AI & Business Glossary

Information an application retains and makes available for later steps or interactions. Memory may hold current task state or selected information across sessions, depending on the system’s design.

A language model on its own remembers nothing between requests. Whatever it seems to recall has been stored by the application around it and handed back to the model alongside your next message. Short-term memory is the running record of the current task: what the customer said a few messages ago, which steps are done, which file is open. Long-term memory is a small store of notes kept between sessions, such as a preferred tone of voice or a client’s billing contact, that gets pulled back in when it looks relevant.

Memory is easy to confuse with its neighbours. A knowledge base holds the approved documents everyone draws on, while memory holds what was picked up along the way in particular conversations. The context window is the space where both end up for a single request. When a supplier says their assistant remembers things, ask what exactly it keeps and who else can see it, because a quietly stored wrong price is worse than no memory at all.

Example

An assistant retains an approved project preference for future work in that project.

Keep in mind

Define scope, retention, correction and deletion; memory is not the same as retraining the model.

Reusable packages of instructions and supporting resources that help compatible agents perform particular tasks. A skill may include reference material, templates or scripts that are loaded when relevant.

In practice a skill is a folder. Inside is a short written file that says what the skill is for and how to do the job, often with templates, examples or small scripts alongside. The agent starts out knowing only each skill’s name and a brief description. When a request matches, it reads the full instructions and opens the extra files only as it needs them, so dozens of skills can sit ready without crowding the conversation.

For a team, skills are a way to write down how you do something once and have it followed every time: your quote layout, your tone for complaint replies, the checks before an invoice goes out. They sit between a system prompt, which shapes every conversation, and connections such as the Model Context Protocol, which give an agent access to other systems. Treat them like any shared procedure, with an owner who reviews each one and retires it when the process changes.

Example

A proposal-writing skill supplies the approved structure and review checklist.

Keep in mind

A skill guides work; it does not independently grant access or change the model’s trained weights.

An approach to AI in which software can select actions and adjust its next steps towards a goal. The degree of independence varies: some systems suggest actions, while others execute approved tasks using connected tools.

The useful question is who decides the next step. In a fixed workflow, a person designed the path in advance: read the email, look up the account, draft a reply. In an agentic system, the model looks at what it has found so far and picks its next move itself, perhaps searching the CRM, then opening a contract, then deciding a colleague needs to approve a discount. It works in a loop of acting, checking the result and choosing again until the goal is met or it reaches a limit.

Agentic is a sliding scale rather than a yes or no label, and more independence is not automatically better. Every decision the model makes is another place where it can go wrong, and the loops cost more time and money than a single answer. Many business tasks are better served by a predictable workflow with AI handling a step or two. Save the agentic approach for work where the path genuinely changes from case to case.

Example

A service assistant investigates a delayed delivery and prepares the next appropriate action.

Keep in mind

Specify its permissions, stopping conditions and escalation rules before calling a system autonomous.

The process of making AI useful in everyday work through appropriate tools, skills, management support and changed working practices. Adoption concerns sustained use and results, not simply distributing access.

Most adoption problems are not technical. A team can hold licences for months and still use them only for the odd email, because nobody showed them where the tool fits into their actual week, or because the old way still feels safer. Real adoption shows when a task is done differently by default: the meeting summary nobody types up by hand any more, the first draft of a quote that now starts from the assistant.

It tends to work when a few things line up. A manager names the tasks worth changing, people get time to practise on their own work rather than generic demos, and someone collects what goes wrong so the setup can be adjusted. Switching a tool on takes an afternoon. Changing habits takes months, which is why adoption sits so close to change management and AI literacy.

Example

A support team learns when to use answer suggestions and when to escalate a case.

Keep in mind

Measure task outcomes and user feedback alongside usage.

An AI system that uses a model to choose steps and tools while pursuing an assigned goal. Its actions operate within the permissions and controls provided by the surrounding application.

An agent has two parts: a model that decides and ordinary software that acts. The software gives the model a goal, a set of tools it may call, such as searching the CRM, reading a mailbox or creating a ticket, and instructions about when to stop. The model proposes a step, the software carries it out, the result comes back and the model decides what to do next. That loop is what separates an agent from an assistant that simply replies to each message.

The quality of an agent depends as much on that surrounding software as on the model. Good setups give each tool the narrowest access that does the job, log every action and pause for a person before anything hard to undo, like a refund or an email to a customer. When someone offers you an agent, ask what it can touch, what it does when it is unsure and how you would see what it did yesterday.

Example

An agent gathers missing order information before suggesting a resolution to an employee.

Keep in mind

Tool access is not permission to perform every possible action.

An application that helps a person complete tasks through AI, often through conversation. Depending on its configuration, it may answer questions, analyse material, create content or use connected tools.

The assistant is the part you see: a chat window, a side panel in your email or a button in a document editor. Behind it sit a language model and whatever the provider or your own team has connected to it. That might be nothing beyond the conversation itself, or it might include your files, calendar and CRM. Two assistants built on the same model can behave very differently because of what each is allowed to read and do.

The defining trait is that a person stays in the driving seat. You ask, it responds, you decide what to keep. That sets it apart from an agent, which is handed a goal and works through steps on its own, and from a typical website chatbot, which talks to customers rather than helping a colleague with their own work. For many teams it is the natural first step with AI, used for drafting, summarising and finding things in long documents.

Example

An employee asks an assistant to compare two approved supplier proposals.

Keep in mind

The word assistant does not establish which data or actions it can access.

Systematic differences or distortions in an AI system’s behaviour that can create unfair or unsuitable outcomes. They can arise from data, design choices, measurement or the context in which the system is used.

Models learn from examples, so they inherit the gaps and habits of those examples. If past hiring decisions favoured one kind of CV, a model trained on them will tend to favour it too. If most of the text a language model saw reflects one culture, its default assumptions about names, roles or customs will lean that way. Bias can also enter later, through what a team chooses to measure or through using a system with people it was never tested on.

In a small business the risk usually sits in quiet places: a step that filters job applicants, an automated priority score on support tickets, a chatbot that copes well with some writing styles and poorly with others. Nobody intends the unfairness, and it often goes unnoticed until a customer or candidate points it out. Bias is an ethical issue, but it is also a quality problem that costs you good candidates and frustrates customers, and in areas such as recruitment it can carry legal consequences as well.

Example

A speech system performs poorly for customers with particular accents.

Keep in mind

Evaluate relevant groups and situations; a strong overall score can hide uneven performance.

A colleague who helps others apply AI to their work, shares useful practices and brings practical feedback to the implementation team. The role connects central guidance with the needs of individual teams.

Champions work because people learn most from someone who does the same job. A colleague in sales who has found a reliable way to prepare for client calls with an assistant is more convincing than any vendor demo, because the example uses real accounts, real jargon and real time pressure. Good champions are curious rather than evangelical. They share what fails as readily as what works, and they carry the awkward questions back to whoever runs the rollout.

The best candidates are usually respected practitioners, not necessarily the most technical person in the room. Aim for one in each team or department, agree a small amount of protected time for the role and give them a direct line to the people who decide on tools and rules. Without that line the role turns into unofficial helpdesk duty, and the enthusiasm that made someone volunteer fades quickly.

Example

A finance colleague runs short sessions using approved invoice-review examples.

Keep in mind

Give champions time and support; they do not replace accountable process owners.

Systematic checks of whether an AI application meets defined requirements. Evaluations can combine test cases, automated measurements and human judgement to assess accuracy, behaviour, safety and task completion.

An evaluation starts with a set of realistic inputs and a clear idea of what a good result looks like for each. For a quote generator that could be a few dozen past requests with the correct prices, and for an email classifier a sample of messages a person has already sorted. You run the system on the set, compare its output with the expected result and score the differences, either automatically, through human review or with a second model acting as a grader whose judgements are spot-checked.

The point is to replace impressions with evidence. A demo that goes well on a handful of questions tells you little about the hundreds that follow. Evals give you a stable yardstick, so you can tell whether a new prompt or a different model actually helped, and they let you agree with a supplier on what good enough looks like before you sign anything off.

Example

A support assistant is tested on normal questions, missing answers and misleading instructions.

Keep in mind

Repeat relevant evaluations when prompts, models, tools or source content change.

The responsibilities, decisions and controls used to direct and oversee AI throughout its life. Governance connects objectives, risk management, monitoring and accountability across the people who build, buy and use AI.

Governance comes down to a handful of plain questions: which AI tools are in use and for what, who approved them, what could go wrong, how anyone would notice and who can switch something off. In a large company the answers live in committees and registers. In a small one they can fit on a single page, as long as the page is kept current and someone actually reads it.

It covers the whole life of a system, from the moment someone proposes a use, through buying or building it, to daily monitoring and eventually retirement. That sets it apart from an AI policy, which sets the rules for everyday use, and from data governance, which looks after the information the systems depend on. The AI Risk Management Framework from NIST, the US standards institute, offers a useful structure, and the EU AI Act adds legal obligations that depend on how AI is used. Where those apply, the regulator’s guidance or a legal specialist is the place to check.

Example

A business records who owns each AI application and who can approve a significant change.

Keep in mind

Governance needs operating practices as well as written policies.

The knowledge and practical understanding needed to use AI appropriately, recognise limitations and judge its outputs. Useful learning is adapted to people’s roles, the systems they use and the potential consequences of mistakes.

AI literacy is less about how models work inside and more about sound working habits. A literate user knows that a fluent answer can still be wrong, that a model may invent a source, and that pasting a customer file into an unapproved tool is a decision with consequences. They know which tasks the tools handle well, such as drafting, summarising and reformatting, and where extra checking is needed, like figures, legal wording and anything that depends on recent facts.

The right level differs by role. A receptionist using an assistant for replies needs different knowledge from a manager choosing a supplier or an analyst building automations. The EU AI Act also addresses AI literacy for organisations that build or use AI systems. It does not prescribe a particular course or certificate, and because these rules are still being adjusted, the European Commission’s published guidance is the best place to check what is currently expected.

Example

A commercial team practises checking generated claims before using them in proposals.

Keep in mind

EU AI literacy duties should be assessed against current official guidance and the organisation’s circumstances.

An ongoing working relationship in which an organisation and an external specialist develop, operate or improve AI capabilities together. Responsibilities may include delivery, training, monitoring and knowledge transfer, depending on the agreement.

The difference from a one-off project is what happens after launch. AI systems need upkeep: a provider updates a model, a form gains a new field, the source documents change, a team finds a new use. A partnership puts someone in charge of noticing and handling that, usually through regular reviews, a simple channel for small changes and periodic checks that each system still does what it should.

A good arrangement leaves you more capable over time, not more dependent. Ask who owns the automations and the accounts they run on, whether documentation and access stay with you if the relationship ends, and how knowledge moves to your own staff. Agree in writing what is included, how quickly requests are handled and how success will be judged, because those details say more about the relationship than its name.

Example

A company retains specialist support to maintain a growing set of automations.

Keep in mind

Agree who owns what, how results are measured and how knowledge moves to your own team before the work starts.

An organisation’s rules for acceptable AI use, including approved tools, permitted data, review requirements and responsibilities. A useful policy gives people clear instructions for ordinary work and exceptions.

A workable policy is short enough that people actually read it. It usually names the tools staff may use, sorts information into what can and cannot go into them, says which outputs need a human check before they reach a customer and explains how to ask for something new. Concrete examples help more than principles. One sentence saying a contract clause may be polished in the approved assistant but never sent unchecked does more than a paragraph about responsible use.

A policy that only forbids tends to push use out of sight, which is how shadow AI grows. One that offers an approved tool and a quick route for requests keeps the conversation in the open. The policy sits inside wider AI governance and next to your data protection rules, but it cannot make a use lawful on its own. For questions about personal data, contracts or regulated decisions, check with your data protection lead or a legal specialist.

Example

Employees know which assistant may receive customer information and who approves new uses.

Keep in mind

Assign an owner and revise the policy when tools, risks or working practices change.

Improving how work is performed by combining process changes with appropriate AI capabilities. The starting point is the desired outcome and the existing workflow, including delays, errors and human decisions.

The order matters. First you look at the work as it really happens, usually by mapping it with the people who do it: where requests come in, who touches them, where they wait and where they get corrected. Many problems disappear with simple changes at this stage, such as a single intake form instead of several inboxes, or a clear owner for a step everyone assumed someone else handled. Only then do you ask which remaining steps AI could take on, typically reading documents, sorting requests, drafting replies or pulling data together.

Adding AI to a broken process tends to produce faster mess. The useful measure is the outcome for the customer or the team, such as the time from enquiry to answer or the share of cases that need rework, compared with a baseline taken before the change. That keeps the effort honest and shows where AI earned its place and where a simpler process did most of the work.

Example

A team simplifies its quotation handover before adding AI-assisted document preparation.

Keep in mind

Simplify the process first. AI added to a confusing step makes the confusion faster, not smaller.

A structured review of whether an organisation has the data, skills, processes, ownership and technology needed for a proposed AI use. It identifies practical gaps and possible next steps.

A good assessment is tied to a specific use, not to the organisation in the abstract. Asking whether a company is ready for AI gets vague answers. Asking whether it is ready for an assistant that drafts replies to warranty claims gets useful ones: are the warranty terms written down and current, do past cases live somewhere searchable, who would check the drafts, and which systems would it need to read? The work usually combines short interviews, a look at the real data and documents, and a walk through the process with the people who run it.

The output should be a short list of gaps in order of importance, each with a practical next step, rather than a long report of generic advice. Often the most valuable finding is unglamorous, like a price list kept in several versions or a process nobody owns. Be wary of an assessment that declares every idea ready to build, or one that ends in a pitch for the assessor’s own product.

Example

A company checks source-document quality and process ownership before starting a support assistant.

Keep in mind

The assessment method is context-specific; there is no universal score that establishes readiness.

A defined way for software components to exchange requests, data or functionality. A web API might let an authorised application retrieve a customer’s order or create a service ticket.

An API works like a menu that one piece of software offers to others. It lists what can be asked for, such as fetching a customer, creating an invoice or updating an order status, and exactly how each request must be written. Another program sends a request along with a key that proves it is allowed in, and gets back a structured answer it can use straight away, with no person copying data between screens.

For a business, the practical question is whether your tools have good APIs. A CRM, accounting package or booking system with a well-documented API can be connected to almost anything, while one without forces workarounds such as exports, screen scraping or retyping. When you choose software, ask what the API covers, whether it is included in your subscription and whether the system can also notify others when something changes, which is the job of a webhook. AI agents reach your systems the same way, so good APIs matter more now, not less.

Example

A website passes an enquiry directly into a CRM using its API.

Keep in mind

Check permissions, rate limits, error handling and the available operations.

Computer systems that perform tasks such as recognising patterns, interpreting language, generating content or making predictions. AI includes several approaches; generative AI is one part of the field.

AI is an umbrella term, and much of what sits under it is older and quieter than the chat tools that made it famous. Spam filters, fraud alerts on your bank card, voice dictation on your phone and the recommendations in an online shop all rely on it. Most modern AI is machine learning: rather than following rules a programmer wrote, the system learns patterns from many examples and applies them to new cases. Generative AI, which produces text, images or code, is a newer branch built on very large models trained this way.

When a supplier says a product uses AI, the useful follow-up is what kind of job it actually does. Predicting a number, sorting items into categories, finding something in documents and writing a draft are very different tasks, with different strengths, failure modes and costs. Knowing which one you are buying makes it much easier to judge whether it will work on your data and in your process.

Example

A business uses one model to forecast demand and another to draft product descriptions.

Keep in mind

Choose a capability that addresses the problem; the label alone says little about business value.

Using technology to carry out work with less manual intervention. An automation can follow explicit rules, use AI for selected steps or combine both approaches with human decisions.

Most business automation is built from the same few parts. A trigger starts it, such as a new email, a paid invoice or a set time each morning. Then come steps that move or reshape data between systems, conditions that send different cases down different paths, and a final action such as a message to a customer or a task for a colleague. Rule-based steps do exactly the same thing every time, while AI steps handle judgement on messy input, such as reading a free-text email to work out what the sender wants.

The best candidates are tasks that happen often, follow a recognisable pattern and currently involve copying, retyping or chasing people. Low-code tools now let small teams build these flows without a developer, which is both their strength and their risk, because nobody may understand a flow once the person who built it leaves. Keep a simple list of what each automation does, which systems it touches and who owns it.

Example

A submitted form creates a record and assigns it to the right team.

Keep in mind

Design how the process handles missing information, duplicates and failures.

Microsoft’s Azure-based access to OpenAI models for building applications. It is part of Microsoft’s evolving AI platform offering, with deployment, identity and management options controlled through Azure.

Azure OpenAI is aimed at developers and IT teams rather than everyday users. Instead of signing up with OpenAI directly, a company reaches OpenAI’s models through its Microsoft Azure account. Applications connect through an API, and administrators decide which models are deployed, who may call them and how usage is tracked. Microsoft has been folding the service into a broader AI platform, which you may see under the Foundry name and which also offers models from other providers.

For a business, the appeal is usually less about the models than about the wrapping. Organisations that already run on Microsoft can manage access with existing accounts, keep AI costs on the same cloud bill and apply the security and monitoring practices they use elsewhere. It is not the same as ChatGPT or Microsoft Copilot, which are finished applications for staff. Azure OpenAI is a building block, so someone still has to build or buy the application that sits on top.

Example

A company connects an internal document application to an OpenAI model deployed through Azure.

Keep in mind

Verify the chosen deployment’s processing location and terms; using Azure alone establishes neither.

A recorded starting level against which later results can be compared. A useful baseline specifies the metric, time period, sample and conditions under which it was measured.

Without a baseline, any claim of improvement is a guess. People’s memory of how long something used to take is unreliable, and once a new tool is in place nobody can reconstruct the old numbers. So the measurement has to happen before the change, on the real work, over a stretch long enough to even out an unusually quiet or busy week.

It does not need to be elaborate. A couple of weeks of simple tallies in a spreadsheet often do the job: how many enquiries came in, how long each waited for a first reply, how many orders went out with an error. Pick measures a customer would notice, not just the ones that are easy to count, and write down how you measured so the same method can be repeated afterwards. A baseline is also what turns a business case from a hope into something you can check.

Example

An operations team measures handling time and correction rates before changing a workflow.

Keep in mind

Compare equivalent work and conditions; a changing case mix can distort apparent improvement.

A defined reference test or comparison used to assess performance. In AI, a benchmark may measure a capability across models, but its results apply to the tasks and conditions tested.

Public AI benchmarks are fixed sets of questions or tasks, such as maths problems, coding exercises or exam-style questions across many subjects, that new models are run against so their results can be compared. They are good for tracking broad progress and spotting large gaps between models. They have limits too: models can be tuned towards popular tests, test questions sometimes leak into training data, and a high score on exam questions says little about how a model handles your returns policy or your customers’ tone.

The difference from an evaluation is mostly one of purpose. A benchmark compares options against each other on a common test, while an evaluation checks whether one particular application meets your requirements. In practice a business can build both from the same set of its own past cases, which keeps the comparison fair and gives you something to rerun whenever a provider releases a new model.

Example

A team compares two models on the same anonymised support cases.

Keep in mind

A public leaderboard does not establish which model is best for your process.

A structured explanation of why an initiative is worth pursuing, covering the problem, options, expected benefits, costs, risks and delivery requirements. It makes assumptions visible so a decision can be reviewed later.

For a small business, a business case can fit on a couple of pages, but it should still answer the questions a larger one would. What problem does it solve, and how do you know the problem is real? Which options were considered, including doing nothing? What will it cost to build and to run, what should it return, and by when? It should also name what could go wrong and who is responsible for delivery.

Its lasting value is the record it leaves. Writing down assumptions, such as how many enquiries the automation will handle each month or how much checking its drafts will need, lets you compare them with reality later. That is how a team learns which kinds of projects actually pay off for it. The benefit figures are only as good as the baseline behind them, so measure before you estimate.

Example

A company compares improving an existing process with introducing a new automation.

Keep in mind

Include implementation, ongoing operation, human review and change-management costs.

A structured approach to helping people move from current ways of working to new ones. It addresses communication, involvement, skills, support and reinforcement so changes can be adopted and sustained.

Every change asks people to give something up, even if it is only a familiar routine. Change management treats that as real work. People need to understand why the change is happening, want to take part, learn the new way, practise until it feels normal and see it reinforced afterwards so the old habit does not creep back. Skipping any of those stages tends to show up later as quiet workarounds.

With AI and automation the human side is often bigger than it looks. A new tool can change who checks what, which skills are valued and how performance is judged, and those questions deserve straight answers early. Involve the people who do the work in shaping the change, give them a named person to go to when something breaks, and keep listening for some weeks after launch, when the novelty wears off and the real problems surface.

Example

A team redesigns responsibilities and practises the new handover before an automation launches.

Keep in mind

Technical delivery and successful adoption require different kinds of work.

A software interface that communicates through messages or speech. It can use fixed rules, AI models or both, and may connect to tools that retrieve information or perform permitted actions.

Chatbots are much older than the current wave of AI. The classic kind follows a decision tree, where buttons and keywords lead down scripted paths, which is predictable but brittle when someone phrases things unexpectedly. Newer chatbots use a language model to understand free text and write replies, usually grounded in a knowledge base of approved answers so they stay on topic. Many real systems mix the two, with scripts for sensitive flows such as payments and AI for open questions.

A good customer-facing chatbot knows its limits. It answers what it can answer well, hands over to a person with the conversation history intact when it cannot, and never invents a policy to keep the chat going. When you assess one, test it with the awkward messages real customers send: typos, anger, several questions at once and requests it should refuse.

Example

A website assistant answers delivery questions and can start a return request.

Keep in mind

Conversation style alone does not reveal the system’s capabilities or reliability.

OpenAI’s AI application for working with language, information and other supported content. Available capabilities can include drafting, analysis, research and tools, depending on the account and product experience.

ChatGPT is the application, not the model. It is the chat interface, on the web, desktop and phone, that sits on top of OpenAI’s language models and adds features around them, such as reading uploaded files, searching the web and working with images. Its public launch is what brought generative AI into everyday conversation, which is why many people still use its name for AI in general.

In business it is mostly used as a general assistant for drafting, summarising, rewriting, brainstorming and working through documents or spreadsheets. Business versions add central administration, so the company rather than each employee controls accounts and settings. Custom GPTs let a team package instructions and reference files for a recurring task. For anything that needs to connect deeply to your own systems, developers usually build on OpenAI’s API instead of the chat app.

Example

A project lead uses it to turn approved meeting notes into an action list.

Keep in mind

Check account-specific data settings and organisational policy before supplying business information.

Anthropic’s family of AI models and related assistant products. Claude can support tasks such as analysing information, producing content and helping with software, with capabilities determined by the product and configuration.

Claude covers two layers. Underneath are the models themselves, which developers reach through an API and build into their own software. On top sit the products people use directly: the assistant you chat with in a browser or app, versions set up for teams, and Claude Code for software work. The same model can behave quite differently depending on which door you come in through, because each product decides which files it can read, which tools it can use and how much material it can handle at once.

For most small businesses the starting point is the chat assistant, used for drafting, summarising long reports, comparing contracts or talking through a tricky email. It belongs to the same group of general-purpose assistants as ChatGPT and Gemini, and they overlap a great deal. The fair way to choose is to run a handful of your own real tasks through each and compare the results side by side. Agree which account your team uses before anyone pastes in customer data.

Example

A team asks Claude to identify inconsistencies across a set of approved documents.

Keep in mind

Evaluate the particular product and plan, rather than assuming all Claude offerings behave identically.

Anthropic’s coding agent for working with software projects. It can inspect files, propose or apply code changes and use development tools within its configured environment and permissions.

Instead of pasting code into a chat window and copying answers back, a developer gives Claude Code a task in plain language and it works inside the project itself. It reads the relevant files, runs commands such as the test suite, edits code and checks whether the change worked, stopping for approval at the steps it has been told to ask about. It usually runs in a terminal or inside a code editor, close to where developers already work.

The value shows up on routine software work: small fixes, internal tools, scripts that move data between systems, documentation nobody had time to write. It also changes who can build things, because a product manager or analyst can get a working prototype without waiting for the development queue. That is close to vibe coding, where nobody looks closely at the code. Fine for a prototype, risky for anything that holds customer data or takes payments, where a developer who can judge the code still needs to own the result.

Example

A developer asks it to diagnose a failing test and propose a correction.

Keep in mind

Review changes and control command execution, credentials and access to production systems.

The amount of information a model can work with in one request, usually expressed in tokens. Instructions, supplied material, conversation history and output allowances share limits defined by the model and application.

Think of it as the model’s desk. Everything it needs for this answer has to be on the desk at once: your instructions, the files you attached, the conversation so far and room for the reply it is about to write. Apart from what it learned in training, whatever is not on the desk the model simply cannot see. When a long chat outgrows the space, many applications trim or summarise the oldest parts, which is why an assistant can lose track of something you agreed an hour earlier.

This explains a few everyday puzzles. A very long document may be cut short or only partly read, and a fresh conversation often gives sharper answers than one that has wandered through several topics. When an assistant has to work across a large body of material, such as every contract you hold or a whole help centre, the usual approach is retrieval: fetch the few passages relevant to each question rather than trying to fit everything in at once.

Example

A long document and its discussion must fit within the application’s available context.

Keep in mind

A larger window does not guarantee that every relevant detail will be used correctly.

Improving an experience so more of the relevant visitors complete a defined action, such as making an enquiry or purchase. It combines research, clear hypotheses and measurement, often including controlled experiments.

The conversion rate itself is simple arithmetic: of the people who reached a page or step, how many did the thing you hoped for. The optimisation is the discipline around that number. You look at where people drop out, talk to customers or watch recordings of real visits to understand why, form a specific guess about what would help, then change one thing and measure. Controlled tests, often called A/B tests, split visitors between the old and new versions, so you know it was the change, not luck or the season, that moved the number.

Smaller sites often have too little traffic for formal experiments to give a clear answer quickly. That does not make CRO pointless. Much of the gain comes from fixing obvious friction first: a form that asks for too much, a price that is hard to find, a phone number buried in the footer, a page that crawls on a mobile connection. Treat it as part of customer experience rather than a bag of button-colour tricks.

Example

A team tests whether clearer service information improves the rate of qualified enquiries.

Keep in mind

Measure downstream quality as well as the headline conversion rate.

A customised ChatGPT assistant configured with reusable instructions and, where supported, reference material or connected capabilities. The term describes an application configuration, rather than a newly trained foundation model.

Building one is closer to writing a thorough briefing note than to programming. You give it standing instructions, such as the tone to use, the steps to follow and what to decline, and you can attach reference files it should consult, like a price list or a style guide. Where the workspace allows it, it can also be connected to other services to look things up or carry out simple actions. Every conversation then starts from the same briefing, so colleagues do not have to paste a long prompt each time.

A custom GPT is only as good as its instructions and files. Give it an owner who keeps those documents current, and test it with real questions before sharing it with the team. Assume that anyone who can use it could coax it into revealing its instructions and the contents of its files, so leave out anything confidential. Most other assistant products offer a similar idea under their own names.

Example

A team configures a reusable assistant to follow an approved proposal checklist.

Keep in mind

OpenAI documents migration paths to plugins; check the current workspace experience before publishing setup instructions.

A survey measure of how easy or difficult customers find completing a task or resolving an issue. Teams use it to understand friction in a specific interaction or journey.

The question is usually asked straight after a specific interaction, while it is fresh: how easy was it to get this sorted? Customers answer on a scale running from very difficult to very easy, and you track either the share of easy answers or the average over time. Because it is tied to one task, a low score points to a place in the journey rather than a general mood about your brand.

Effort matters because people remember hassle: being transferred, repeating an account number, chasing a reply that never came. Removing that hassle is often a surer route to repeat business than adding extras. CES sits well next to CSAT, which asks how pleased someone was, and it pairs naturally with process mapping, which shows the internal steps behind the friction customers report. Read the free-text comments too, since they usually name the exact step that hurt.

Example

Customers report how easy it was to change a booking.

Keep in mind

Scales and scoring direction vary; define the method before comparing scores.

How customers perceive their interactions with a business across the relationship. It includes how easy, reliable and useful those interactions feel, from finding information to receiving support.

Customer experience is the sum the customer adds up, not the one you add up. It includes the obvious moments, such as the website, the sales call and the support inbox, and the quieter ones that are easy to forget: how the invoice reads, whether the delivery slot is kept, what happens when someone emails out of hours. Each of those is run by a different team or system, but the customer meets them as one company.

For a small business, CX is often where you can outdo bigger rivals, because you can fix things quickly and know customers by name. Good CX work starts with evidence: complaints, survey comments, call recordings, the questions customers keep asking. It is broader than customer service, which is one part of the experience, and broader than user experience, which usually means the design of a product or screen. Repeat business and referrals tend to follow when dealing with you is consistently easy.

Example

A clear promise, accurate delivery update and helpful resolution all shape the same customer’s experience.

Keep in mind

Improving a single touchpoint may not fix a problem elsewhere in the journey.

The sequence of interactions and experiences a customer goes through while pursuing a goal. A journey can span channels, teams and time periods, including activity outside a company’s direct control.

The useful shift is to see the work from the customer’s side. Inside a company, an order passes through sales, finance, the warehouse and support, each with its own systems and targets. The customer sees none of those boundaries, only one long attempt to get something done, including parts you never touch, such as reading reviews, asking a friend or waiting in for a parcel.

Thinking in journeys helps you spot problems no single team owns. A smooth checkout can be undone by a confusing first invoice, and a warm sales call loses its shine when the handover to delivery drops the details. Most businesses have a handful of journeys worth understanding well, such as buying for the first time, getting help and renewing, rather than one master journey. Customer journey mapping is the practical method for drawing one out.

Example

A buyer researches options, requests advice, orders, receives the product and seeks support.

Keep in mind

Define the customer and goal before describing a journey.

Creating a visual account of a customer’s actions, needs, thoughts and difficulties across a journey. A useful map is grounded in research and helps teams identify where improvement is needed.

A typical map runs left to right through the stages of the journey. Under each stage you note what the customer is doing, what they are trying to achieve, what they think and feel, and which channel or person they are dealing with. Many teams add a lower band showing the internal steps and systems behind each moment, which brings it close to a service blueprint and turns the map into a bridge between customer experience and process work.

The workshop is often worth as much as the drawing. Put people from sales, operations and support in one room with real evidence, such as interview notes, complaint themes and a few call recordings, and they quickly see where their handovers hurt the customer. Start with one journey for one type of customer rather than trying to map everything. A map that leads to a few real fixes is worth more than a beautiful poster nobody looks at.

Example

A company maps onboarding to locate repeated requests for the same information.

Keep in mind

Separate evidence from assumptions and assign owners to the improvements identified.

The practices and software used to manage relationships with prospects and customers. CRM systems commonly organise contact details, conversations, opportunities and service activity so teams can coordinate their work.

At its simplest, a CRM is a shared memory of every customer. Each company and contact has a record, and around it sit the emails, calls, meetings, quotes and tickets linked to them. Most systems add a pipeline view, where open deals move through stages, plus tasks and reminders so follow-ups do not depend on someone’s inbox.

In a growing business the CRM often becomes the hub other tools connect to. Website forms feed it, the email tool reads from it, and the invoicing or ERP system exchanges data with it, usually through APIs. That is why you need to decide which system has the final say on each piece of information, such as whether the billing address lives in the CRM or in accounting. When choosing one, start from how your team actually sells and serves customers, and pick the simplest system that fits.

Example

A sales handover includes the customer’s goals and previous conversations in one shared record.

Keep in mind

A CRM’s value depends on data quality and consistent team use.

Keeping customers active with a business over a defined period. Retention can reflect continued subscriptions, repeat purchasing or another clearly specified relationship measure.

Retention is the other side of churn: one counts the customers who stayed, the other the ones who left. What counts as staying depends on the business. For a software firm it is a renewed contract, for a webshop a second order within the season, for a clinic a booked follow-up.

Keeping a customer usually costs less than winning a new one, which is why small improvements in retention add up over the years. Most churn has a cause you can find: a clumsy onboarding, a broken promise, a slow reply to a complaint, or simply nobody checking in. Compare the customers who left with the ones who stayed, and the patterns often show up in your CRM and support history before anyone sends a survey.

Example

A company checks which customers from a starting cohort are still active at quarter-end.

Keep in mind

Define the cohort and activity rule; newly acquired customers should not inflate retention of the original group.

A survey measure of satisfaction with an experience, interaction or product. A common calculation is the percentage of respondents selecting satisfied or very satisfied on a five-point scale.

CSAT is the quick, local question: how satisfied were you with this? It works best when asked close to the moment, about one specific thing, such as a delivery, a support case or a visit. Because it takes seconds to answer, you can ask it often and tie the scores to the agent, product or channel involved, which shows where satisfaction rises and falls.

It answers a different question from NPS, which asks about the overall relationship and whether someone would recommend you, and from CES, which asks about effort. Many teams use CSAT for individual interactions and NPS for the relationship as a whole. Watch who replies: delighted and annoyed customers tend to answer more readily than the quietly indifferent, so a high score from few responses can flatter you. The comment box is often more useful than the number.

Example

After a support interaction, customers rate their satisfaction with the help received.

Keep in mind

State the question, scale, response count and calculation when comparing results.

The decisions, responsibilities and rules that govern how an organisation manages and uses data. It covers ownership, access, quality and handling throughout the data’s life.

In a large organisation, data governance comes with councils and policy documents. In a small business it can be a single page that answers plain questions. Which system holds the master version of customer, product and price data? Who may change it, how long do you keep it, and when may someone export it to a spreadsheet? Those answers, written down and agreed, are the core of it.

It sounds like paperwork, but the gaps show up in daily work. Duplicate customer records, a price that differs between the website and the quote tool, and a former employee who can still open the shared drive are all governance problems. Governance overlaps with data protection law such as GDPR, which sets rules for personal data, but it covers all your data and is about running the business well, not only about compliance. None of this needs special software at first, only clear rules that people actually follow.

Example

A business assigns an owner to customer records and defines who may amend them.

Keep in mind

An AI integration should inherit clear data rules rather than create an uncontrolled copy.

A binding arrangement governing a processor’s handling of personal data for a controller. GDPR Article 28 specifies required terms, including instructions, confidentiality and security responsibilities.

Under GDPR, the controller is the organisation that decides why and how personal data is used, and a processor handles that data on the controller’s behalf. Payroll providers, email platforms and cloud storage services typically act as processors, and an AI supplier often does too. The agreement sets out what the processor may do with the data, how it keeps it secure, when it may bring in other processors, how it helps with people’s requests and breaches, and what happens to the data when the contract ends.

Most established suppliers offer a standard DPA, often accepted in a few clicks. Read it anyway: check where the data is processed, which sub-processors are listed and how quickly they promise to tell you about a breach. A signed DPA is one piece of the picture, not proof that a use is lawful, and when the roles are unclear or the data is sensitive, a privacy specialist is worth calling. In Dutch it is known as a verwerkersovereenkomst.

Example

A business documents an AI supplier’s processor obligations before that processing begins.

Keep in mind

Whether this arrangement is needed depends on the parties’ actual roles.

An assessment of a personal-data processing activity’s risks and safeguards. GDPR Article 35 requires it when processing is likely to create a high risk to people’s rights and freedoms.

A DPIA is a structured piece of work done before the processing starts, not afterwards. It describes what you plan to do with personal data and why, tests whether that is necessary and proportionate, identifies the risks to the people involved and records the measures that reduce them. If a high risk remains that you cannot bring down, the law expects you to consult the data protection authority before going ahead.

Typical triggers include large-scale use of sensitive data such as health records, large-scale monitoring of public spaces, and automated evaluation of people that leads to significant decisions about them, such as credit or job applications. Each national regulator publishes a list of processing that always needs one, so check your own authority’s list. Even when a DPIA is not required, a lighter version of the same thinking is a sensible habit for any new tool that touches customer or staff data. If you have a data protection officer, their advice belongs in the process.

Example

An organisation evaluates privacy risks before introducing large-scale sensitive-data analysis.

Keep in mind

Not every AI use automatically requires a DPIA.

How well data meets the needs of its intended use. Relevant qualities include accuracy, completeness, consistency, timeliness and validity; their importance depends on the task.

Data quality problems are rarely dramatic. They look like a customer entered three times with slightly different spellings, a postcode field full of phone numbers, a product sheet nobody has touched since the range changed, or a status field half the team never fills in. Each one is small, but together they decide whether a report can be trusted, whether an automation sends the right email and whether a customer gets the right answer.

Quality is always relative to a use. A contact list good enough for a newsletter may be too patchy to route support tickets by contract type. The cheapest fixes happen at the point of entry: required fields, drop-down choices instead of free text, checks on forms and one system that owns each type of record. Cleaning up afterwards helps, but the mess returns unless the cause is fixed, and an AI assistant fed bad data will repeat it confidently to every customer who asks.

Example

A support assistant relies on product information that has a named owner and current review date.

Keep in mind

Define measurable quality requirements before treating a dataset as ready for AI.

A branch of machine learning that uses neural networks with multiple layers to learn patterns. These methods support many language, image and speech systems.

A neural network is a long chain of simple calculations whose settings are adjusted during training until its outputs come close to the examples it is shown. Deep simply means many layers stacked on top of each other. Early layers pick up simple patterns, such as edges in a picture, and later layers combine them into more abstract ones, like the outline of a face or the tone of a sentence. Nobody writes these rules by hand: the network finds them in large amounts of data.

Most of the AI tools in the news rest on deep learning, from speech-to-text in a call centre to the large language models behind chat assistants. Businesses rarely need to train such a model themselves, because ready-made models can be used as they are or adapted. For tidy tables of business data, such as orders or invoices, simpler machine learning methods are often just as accurate, cheaper to run and easier to explain.

Example

A visual inspection model learns to recognise recurring product defects from labelled images.

Keep in mind

The appropriate method depends on the task, data and available resources.

A numerical representation of an item, such as text or an image, learned to capture useful relationships. Suitable embeddings allow software to compare related content even when its wording differs.

Picture a huge map on which every sentence, paragraph or image gets a location. An embedding model places items with similar meaning close together, so a question about late deliveries lands near a paragraph about shipping delays even though they share hardly any words. In practice the location is a long list of numbers, and software measures how close two lists are to judge how related the items are.

Embeddings do the quiet work behind semantic search and many retrieval-augmented generation systems. Your documents are split into chunks, each chunk is turned into an embedding and stored, often in a vector database, and every new question is converted the same way and matched against them. You never see embeddings directly, but they decide which passages an assistant reads before it answers. If an assistant keeps citing the wrong part of your help centre, the way documents were split and embedded is one of the first things to check.

Example

A search system connects a question about cancellation with a termination-policy passage.

Keep in mind

Similarity scores depend on the embedding model and do not establish factual correctness.

Software that connects core business functions such as finance, purchasing, inventory and operations through shared processes and data. Specific coverage varies by product and implementation.

Without an ERP, a growing business often runs on separate tools that do not talk to each other: an accounting package, a stock spreadsheet, a purchasing inbox, a planning whiteboard. An ERP brings those into one system with shared data, so a purchase order, a goods receipt and a supplier invoice are linked records rather than three documents someone matches by hand. Most ERPs are sold as modules, and a business switches on the ones it needs.

The difference from a CRM is focus. A CRM looks after the relationship with customers before and after the sale, while an ERP runs the operational and financial core: what you buy, make, hold, ship and invoice. In smaller businesses the line blurs, and some systems do a bit of both. ERP projects have a reputation for running long, mostly because they force decisions about how processes should work, so it pays to agree those processes before configuring anything.

Example

An approved order updates stock planning and financial records within connected business processes.

Keep in mind

Integration must preserve business rules and authorised access to important records.

EU legislation governing certain AI systems and models through obligations that depend on their use, risk and the organisation’s role. Its requirements include prohibited practices, rules for higher-risk systems and transparency duties.

The Act sorts AI by what it is used for, not by the technology. A short list of practices is banned outright, such as social scoring or techniques that manipulate people or exploit their vulnerabilities. A wider group counts as high-risk, such as uses that affect people’s access to jobs, education, credit or essential services, and these carry duties around risk management, data, documentation, human oversight and accuracy. Many everyday uses fall into neither group but can still carry transparency duties, such as telling people they are dealing with a chatbot or labelling certain AI-generated content.

Obligations also depend on your role. Most SMEs will be deployers, using an AI system someone else built, rather than providers who develop it and put it on the market, and deployers carry lighter but real duties. The Act sits alongside GDPR, so personal data in an AI system falls under both. For any use close to the high-risk list, the European Commission’s guidance and a legal specialist are the right next stop.

Example

A company assesses the applicable requirements before introducing AI into recruitment.

Keep in mind

Check the current rules for the specific use; obligations and implementation dates can differ.

The ability to provide understandable information about a system’s behaviour or the factors influencing its outputs. Useful explanations are tailored to the people who must review, use or challenge a result.

Explanations come at different depths. For a simple scoring model you can often show which factors pushed a result up or down, such as order value, contract tier or the number of earlier complaints. With large language models that is much harder, so explainability usually means showing the working from the outside: which documents an answer drew on, which rules applied, what data was used and who approved the result.

The right depth depends on who is asking. A customer refused a refund needs a plain reason they can challenge, a team lead needs to know when to overrule the system, and an auditor needs records of how decisions were made. Explainability is what makes human oversight real, because a reviewer who cannot see why a recommendation was made can only rubber-stamp it. When buying an AI tool, ask the supplier what a user can see behind each output.

Example

A reviewer can inspect the information used to support a service-priority recommendation.

Keep in mind

A model’s fluent explanation is not necessarily a faithful account of its internal computation.

Additional training that adjusts a model using task-specific data. It can help shape behaviour or performance, such as following a classification scheme or producing a consistent output style.

You start from a model that already handles language well and train it a little further on dozens to thousands of examples of exactly what you want, each one an input paired with the ideal output. The model’s internal settings shift so the behaviour becomes its default, without long instructions every time. It is good at teaching a format, a tone or a way of deciding. It is a weak way to teach facts, because prices and policies change and a tuned model can still misremember them.

For most SMEs, fine-tuning is a later step, not a first one. Clear prompts, a few good examples inside the prompt and retrieval from your own documents solve many problems more cheaply and are easier to update. Fine-tuning starts to earn its keep on a narrow, high-volume task, with a large set of carefully checked examples and a way to measure whether the tuned model really does better. It also leaves you with something to maintain when the underlying model is retired.

Example

A team trains an eligible model on carefully reviewed examples of ticket categories.

Keep in mind

It is different from attaching documents or adding information to a prompt.

A model trained broadly enough to support multiple downstream tasks or applications. Organisations can use it directly or adapt it through instructions, retrieval, tools or additional training.

Foundation models are trained up front, at great expense, on very large and varied collections of text, images, code or audio. That broad training is what makes them adaptable: the same model can answer a customer question, translate a contract clause or write code without being built for any of those jobs. The large language models behind ChatGPT, Claude and Gemini are the best-known examples, but there are foundation models for images, speech and other kinds of data too.

The practical consequence is that few businesses need to build AI from scratch, and the real work moves to everything built around the model. Choosing between models is then less about raw intelligence and more about fit: cost per task, speed, data terms and how well it copes with your language and documents. Ask suppliers whether their product lets you swap one model for another later. The term overlaps with large language model, since most LLMs are foundation models, but not every foundation model works with language.

Example

The same base model supports separate applications for drafting, classification and analysis.

Keep in mind

Broad capability does not establish suitability for a particular business task.

A mechanism that lets a model request a defined operation with structured arguments. The surrounding application validates and executes the operation, then can return its result to the model.

A language model on its own can only produce text. Function calling gives it a menu. The developer describes a set of tools, such as finding a customer record, checking stock or creating a ticket, each with a name, a short description and the details it needs. When a question calls for one, the model replies with a structured request instead of prose, the application runs the tool and the model uses the result to write its answer.

This is the step that turns a chatbot into something that can do work, and it is the core mechanism inside AI agents. The model never touches your systems directly. It only asks, and your code decides whether to act, which is where the control lives. A sensible order is to start with read-only tools that fetch information, and to add tools that change records, send messages or move money once confirmation steps and logging are in place.

Example

An assistant requests an order-status lookup using an order identifier.

Keep in mind

Validate authorisation and arguments before execution, especially for actions that change data.

Google’s family of AI models and related assistant experiences. Different Gemini products support different combinations of language, images, audio, tools and workplace integrations.

Gemini belongs to the same group of general-purpose assistants as ChatGPT and Claude, and its main difference is where it lives. It turns up inside Gmail, Docs, Sheets and Meet and on Android phones, so for a team already working in Google’s apps it can be the assistant closest to the email and files people use every day. The models were designed from the start to handle text, images, audio and video together.

Being close to your files cuts both ways. Workplace assistants like this generally work within the sharing permissions people already have, so a drive where too much is shared with everyone becomes a real problem the moment anyone can simply ask about it. Before switching it on across a workspace, tidy those permissions and agree what it should and should not be used for. Developers can also build Gemini models into their own software for tasks such as reading documents, summarising and chat.

Example

A team uses an approved Gemini experience to summarise information for a project.

Keep in mind

Distinguish the consumer assistant, workplace features and developer API when comparing capabilities or data terms.

The EU framework governing personal-data processing, including lawful handling, individual rights and accountability. Its rules apply to relevant AI processing as they do to other processing activities.

It rests on a few principles. You need a lawful basis for each purpose you use personal data for, you collect only what that purpose requires, you keep it accurate, secure and no longer than necessary, and you can show all of this. People keep rights over their data, including seeing it, correcting it, having it deleted and objecting to its use. In the Netherlands you will hear it called the AVG.

GDPR has no exemption for AI. Pasting a customer’s email into an assistant, transcribing a sales call and scoring leads are all processing, so the usual questions apply: what personal data goes in, why, which supplier handles it, how long it is kept and how you would handle a request to see or delete it. Suppliers working on your behalf need a data processing agreement, and processing likely to pose a high risk to people calls for a DPIA. For a specific case, your national data protection authority’s guidance or a privacy specialist is the next stop.

Example

A team limits the personal information included in a document-analysis task.

Keep in mind

An enterprise subscription alone does not establish lawful processing.

AI that produces content such as text, code, images or audio using patterns learned during training and the input it receives. Generated material can be useful without being accurate or suitable for publication.

Under the bonnet, a generative model is a very capable predictor. A text model writes by choosing a likely next word, again and again, based on everything that came before, while image and audio models build their output in a comparable step-by-step way. The model does not look facts up unless the application around it is built to fetch them. Each answer is composed fresh, which is why the same request can come back differently twice.

That makes it a different kind of tool from older AI that sorts or scores things. A spam filter labels an email. A generative model writes the reply. In a small business it earns its keep on first drafts, summaries, rewrites and translations, where a person who knows the facts finishes the job. It is riskiest where nobody checks, such as prices, figures or promises that go straight to a customer.

Example

A team creates a first draft of a product explanation from approved facts.

Keep in mind

Review claims, context and intended use before relying on the output.

A processor designed to perform many calculations in parallel. GPUs support graphics and many AI workloads, including model training and running predictions or generated responses.

A regular processor, the CPU, is built to handle many different jobs quickly, mostly one after another. A GPU takes the opposite approach, with a very large number of simpler cores that all work at the same time. That design was made for video games, where every pixel on the screen has to be worked out many times a second. It turned out to suit AI well, because training and running a model comes down to enormous amounts of multiplication that can be split into small pieces and done side by side.

Most small businesses never touch one directly. When you use a cloud AI service, the GPUs sit in the provider’s data centre and their cost is folded into your subscription or usage bill. The question only becomes practical if you want to run a model on your own servers. Then the memory on the card limits which models fit, and the hardware is expensive to buy, power and keep busy.

Example

An AI service uses GPUs to process many model requests efficiently.

Keep in mind

They are one computing option; AI can also run on other processors and specialised accelerators.

Connecting an AI response to supplied or retrieved evidence, such as documents, database records or search results. Grounding aims to make the response reflect a relevant information source.

A model on its own answers from what it absorbed during training. That knowledge is broad but frozen at a point in time, and it has never seen your price list or yesterday’s orders. Grounding changes the setup. Before the model answers, the application fetches the relevant facts, such as a policy page, a CRM record or a set of search results, and asks the model to answer from those. Retrieval-augmented generation is the most common way to build this.

It is the difference between a plausible answer about return policies in general and a correct answer about yours. Good grounded systems show where an answer came from, say so when the sources are silent and draw on data that is kept current. When you assess a supplier, ask what happens when nothing relevant is found. Grounding narrows the room for hallucination, but a model can still misread or stretch the source it was given.

Example

A response about an order refers to its current status in an approved system.

Keep in mind

Check that the cited evidence actually supports the claim.

Controls that help constrain an AI system’s behaviour and actions. They can include permissions, input and output checks, spending limits, approval steps and monitoring around the model.

Think of the controls you would put around a capable new hire. They get keys to some rooms but not all, payments above a limit need a second signature, and someone reads the logs now and then. Guardrails do the same for an AI system. Some act on what goes in, such as masking personal data or turning away off-topic requests, some limit which systems and actions the model can reach, and some check the output before anyone acts on it.

How many you need depends on what the system can do. An assistant that only answers questions needs fewer than an agent that can send emails, issue refunds or change records. The firmest guardrails live in ordinary software around the model, as permissions and hard limits it cannot talk its way past. When a supplier describes their safeguards, ask which ones are enforced by code and which depend on the model following its prompt.

Example

A payment-related workflow rejects unapproved destinations and requires an authorised employee to approve execution.

Keep in mind

Instructions alone are not a security boundary, and no single control removes every risk.

Generated content that is false, unsupported or inconsistent with the available evidence, even when it sounds plausible. Examples include invented references, incorrect figures and claims absent from the supplied material.

It happens because a language model is built to produce likely-sounding text, not to check facts. When the answer is missing from what it was given, it rarely stops on its own. It fills the gap with something shaped like an answer, in the same assured tone it uses when it is right. And because each answer is composed fresh, the same question can be answered correctly on Monday and wrongly on Tuesday.

For a business, the risk depends on where the output goes. A draft that a colleague reads before sending is a small risk. An answer that reaches a customer directly, or a figure that lands in a quote, is a bigger one. Grounding the model in your own documents, letting it say it does not know and testing it on real questions all lower the rate, but none of them brings it to zero. Plan for the errors that remain rather than hoping a newer model will remove them.

Example

An assistant supplies a product warranty period that does not appear in the policy.

Keep in mind

Retrieval and review can reduce errors; fluent language does not establish truth.

A workflow design that includes human input, review or intervention at defined points. The person may resolve uncertainty, approve an action or correct a result before the process continues.

The person can sit in a few different places. They can approve before anything happens, such as a refund or a message to a customer. They can handle the exceptions, while the system processes the clear cases and passes on the ones it is unsure about. Or they can review a sample afterwards to catch problems early. Where the human goes is a design choice you make step by step, usually based on how costly and how reversible a mistake would be.

The common failure is the rubber stamp. When a system is right almost every time, reviewers start clicking approve without really looking, and the check turns into theatre. The opposite failure is a queue so long that the automation saves nothing. A good setup sends people only the cases that need judgement and feeds their corrections back into the rules or prompts, so the queue shrinks over time.

Example

An employee checks an extracted bank-account change before it reaches the supplier record.

Keep in mind

Give reviewers enough context, time and authority to make a meaningful decision.

Using an already trained model to produce an output from new input. Inference includes classifying a message, recognising an image or generating a response, depending on the model.

If training is the schooling, inference is the working day. Every time a model answers a question, transcribes a call or tags an invoice, that is inference, and it happens again with every request. Training a large model is a rare and expensive project. Inference is the ongoing activity that follows, running whenever someone uses the system.

Most businesses only ever do inference, on a model someone else trained. It is the part you feel: the wait before a chat reply appears and the usage line on your monthly invoice, which language model providers usually count in tokens. Larger models and reasoning models do more work per answer, so they tend to be slower and pricier. Matching the model to the task, a small fast one for sorting emails and a bigger one for tricky drafting, is often the simplest saving available.

Example

An incoming service request is classified into an agreed category.

Keep in mind

Operating cost depends on the model, workload and infrastructure, not just the number of requests.

Using software, including AI where appropriate, to classify documents, extract information and feed validated results into a business process. It can combine text recognition, interpretation, checks and human review.

The interesting part comes after the reading. Older document capture relied on fixed templates, one per supplier layout, and broke whenever a layout changed. Newer systems use AI models that cope with layouts they have not seen before, then check what they found: does the purchase order exist, do the line totals add up, does the VAT number match the supplier on file. Those checks are what turn extracted text into data a process can trust.

Typical candidates in a small business are invoices, purchase orders, delivery notes, claim forms and signed contracts. Ask a supplier how much of a normal week would pass through with no human touch, and test with your own documents, including the crumpled scans, the handwritten notes and the supplier who sends everything as one long PDF. A supplier’s demo documents are almost always cleaner than your inbox.

Example

A supplier invoice is classified, key fields are extracted and exceptions are sent for review.

Keep in mind

Evaluate the full document-to-decision process, not just text-recognition accuracy.

A selected measure used to judge progress towards an important objective. A useful KPI has a clear definition, owner, data source and review frequency.

Every KPI is a metric, but most metrics are not KPIs. A number earns the label when it is tied to an objective someone cares about and a decision depends on it. It also helps to mix two kinds. Lagging indicators such as revenue or churn tell you what already happened, while leading indicators such as quotes sent within a day or demos booked give you an earlier signal you can still act on.

The classic trap is that people optimise for the number rather than the goal behind it. Put a target on average call length and some agents will rush customers off the phone. Before any automation or AI project, agree which KPI should move and record where it stands today. Without that baseline, nobody can say afterwards whether the project paid off.

Example

A support team tracks resolution quality and time to a useful first response.

Keep in mind

Choose a small set that reflects outcomes and makes unintended trade-offs visible.

An organised collection of information that people or software can use to answer questions and complete work. In an AI application, it may supply reference material through search or retrieval.

In practice it can be a help centre, an internal wiki, a folder of policies or a set of product sheets. For an AI application, the content is usually split into short passages and indexed so the system can search by meaning rather than exact keywords. When a question comes in, the most relevant passages are found and handed to the model to answer from. This is the retrieval step in retrieval-augmented generation.

The weak point is rarely the technology. In many businesses knowledge is spread across shared drives, old email threads and a few people’s heads, with two versions of the same policy that quietly disagree. An assistant fed that material will repeat the confusion with great confidence. A useful test before any AI project: if a new colleague could not answer a customer correctly from the knowledge base alone, an assistant will not either.

Example

A support assistant consults maintained policy pages and approved product documentation.

Keep in mind

Assign content owners, review dates and access controls.

A machine-learning model trained on large amounts of language data to process and generate text. LLMs are components within applications and can be combined with retrieval, tools and other capabilities.

An LLM learns by reading an enormous amount of text and practising one task: predicting the next small chunk of a sentence, called a token. Done at sufficient scale, that simple exercise produces a model that can summarise, translate, classify and draft. It is then trained further on examples of good answers and on human feedback, so it follows instructions instead of just continuing your text. The word large refers both to the training data and to the number of internal settings, or parameters, that the model adjusts as it learns.

Two limits are worth knowing. A model’s knowledge stops at the point its training data was collected, and it can only consider a certain amount of text at once, called its context window. Everything else, from searching your files to remembering last week’s conversation, is added by the application. That is why the same model can feel sharp in one product and clumsy in another.

Example

An application uses an LLM to turn a support case into a draft response.

Keep in mind

The model and the application around it have different responsibilities and limits.

Approaches to building software with visual tools and reusable components. No-code aims to minimise coding, while low-code also allows code where additional flexibility or integration is needed.

Instead of writing code, you assemble an app or automation from ready-made blocks: a form, a table, a screen, or a rule that adds a row and emails the manager whenever a form arrives. Connectors to common tools handle the plumbing, and the platform takes care of hosting, logins and updates. Low-code tools add an escape hatch, so a developer can write a few lines for the one step no block covers, such as an unusual calculation or a system without a ready connector.

The big gain is that the people closest to a process can build and adjust it without waiting in an IT queue. It suits internal tools, approvals and the glue between your existing apps. It is a weaker fit for very high volumes, intricate logic or a customer-facing product where speed and polish matter. Leaving a platform usually means rebuilding, because what you made rarely exports cleanly to anything else.

Example

An operations team assembles a request form and approval flow using a visual builder.

Keep in mind

Data access, maintenance and ownership still need deliberate decisions.

Methods that learn patterns from data to make predictions or perform tasks. A model’s behaviour is learned through training rather than specified entirely as hand-written rules.

The difference shows up in how you would build something like a fraud check. With rules, someone writes them down, for example flag any large order shipped to a new address. With machine learning, you show the system many past orders marked as fraudulent or fine, and it works out for itself which combinations of signals matter. Other methods find structure without any labels, such as grouping customers who behave alike.

You already meet it in spam filters, payment fraud checks, product recommendations and the lead scores in many CRMs, usually built in by the supplier. Deep learning is one branch of machine learning, and large language models are one product of that branch. Building your own model needs plenty of clean history with the outcome recorded. When that history is thin, a well-chosen rule or a simple spreadsheet forecast can do the job just as well.

Example

A forecasting model learns relationships between historical orders and future demand.

Keep in mind

Performance on past data does not guarantee useful results on changing real-world data.

Microsoft’s branding for several AI experiences. In workplace use, it often refers to assistance within Microsoft 365 and related tools; access to organisational information and actions depends on the product, licensing and configuration.

In everyday work it shows up as an assistant inside apps your team already opens: drafting in Word, summarising a long thread in Outlook, recapping a Teams meeting or explaining a spreadsheet in Excel. The workplace versions can draw on your organisation’s emails, files and calendars, limited to what each user already has permission to see. That cuts both ways. If old folders are shared too widely, the assistant can surface documents people never knew they could open.

For a business already working in Microsoft’s office apps, the appeal is that nobody has to learn a new tool or copy data elsewhere. Before rolling it out, tidy up file sharing and decide which teams get access first. Then judge it on a handful of concrete tasks people do every week, since a general sense of helpfulness is hard to justify on a licence bill.

Example

An employee uses a configured Copilot experience to prepare for a meeting.

Keep in mind

Distinguish workplace Copilot from GitHub Copilot and other products sharing the name.

An open protocol for connecting AI applications with external tools, resources and contextual information through a shared interface. It helps integrations use a consistent structure across compatible clients and servers.

The usual comparison is a universal plug. Before a shared standard, every AI app needed its own custom connection to every tool, so linking a few assistants to a few systems meant building and maintaining each pairing separately. With MCP, a service is wrapped once in an MCP server, and any compatible AI application, the client, can use it. A server can offer actions the AI may take, data it may read and ready-made prompts.

For a business, this mostly matters when you compare AI tools: can your assistant reach your CRM, ticketing system or file storage without bespoke development? Treat each MCP server like any app you install with access to company data. Find out who built and maintains it, and start with read-only access before you let it create or change records.

Example

An assistant connects to an approved document-search service through an MCP server.

Keep in mind

The protocol does not automatically grant permission or establish that a connected service is trustworthy.

The process of adjusting a machine-learning model using data and a learning objective. Training shapes the model’s parameters; using the trained model on new input is inference.

A model starts out with its internal settings, the parameters, set more or less at random. It makes a prediction on a training example, the error is measured against the right answer, and every parameter is nudged slightly in the direction that would have reduced that error. Repeat that over a huge number of examples and the settings gradually settle into something useful. For the largest models this takes vast amounts of data and computing power, which is why few organisations train them from scratch.

A small business almost never needs to. The realistic options are using a ready-trained model with good prompts and your own documents, fine-tuning one on a modest set of examples, or training a small model for a narrow job like demand forecasting. When a supplier says they train on your data, ask which of those they mean. Training changes the model itself, while retrieval only reads your documents, and that difference matters for privacy and for removing data later.

Example

A forecasting model is trained using historical demand data.

Keep in mind

Keep training and evaluation data appropriately separated to avoid misleading performance estimates.

A system in which multiple AI agents perform coordinated roles or subtasks. They may divide work, exchange information or check results under a shared orchestration design.

A common pattern is a lead agent that breaks a job into pieces, hands each piece to a specialist and assembles what comes back. Other designs run as a relay, each agent passing its work to the next, or pair a worker with a reviewer that critiques the output before it moves on. Each agent can get its own instructions, tools and permissions. That separation is useful in itself, because the agent that reads the open web need not be the one allowed to send emails.

It is easy to confuse this with a single agent that uses several tools, which is often all a task needs. With more agents, every handover is another place where context can be lost or misread, and small errors can compound along the chain. Whoever runs it needs a clear record of what each agent received and returned, otherwise tracing a bad result back to its cause turns into guesswork.

Example

Separate agents research a topic and check whether the draft cites the gathered evidence.

Keep in mind

Use additional agents only when their contribution justifies the extra coordination, cost and testing.

AI that works with more than one type of information, such as text, images, audio or video. Supported input and output combinations differ between models and applications.

Earlier systems handled each type of information with a separate model: one turned speech into text, another read the text, a third labelled images. A multimodal model takes several of these in at once, so it can relate them to each other, such as reading a chart together with the paragraph that explains it. Understanding a format and producing it are separate abilities. A model that reads images well may not create them at all.

In a small business this shows up in quieter places than the demos suggest: reading scanned forms and photographed receipts, answering a question about a screenshot in the support inbox, transcribing calls or writing alt text for product photos. It also lowers the bar for intelligent document processing, since one model can often read a messy page without a separate OCR step. Try it on your real material, blurry phone photos and handwriting included, before you rely on it.

Example

A model compares a damage photograph with the accompanying written service request.

Keep in mind

Being multimodal does not mean a model supports every media format or performs every task well.

A workflow automation platform that connects services and combines processing steps, including optional AI components. Workflows can run in supported hosted or self-managed environments, depending on the chosen setup.

You build a workflow on a canvas by linking blocks, called nodes. A trigger starts each run, such as a submitted form, an incoming email, a webhook or a schedule. Every node after that does one job: fetch a record, reshape the data, ask an AI model to classify a message, branch on a condition or update another system. While testing you can open any node and see exactly what went in and what came out, which makes faults much easier to find.

It sits between no-code tools and custom development. Ready-made connectors cover many common apps, a general request node reaches any system with a web API, and a code node lets a developer add logic where the blocks run out. That flexibility comes with a steeper learning curve than the simplest automation tools. Running it on your own servers keeps the data within infrastructure you control, but updates, backups and security then become your job.

Example

A workflow receives an enquiry, enriches its details and creates a CRM task.

Keep in mind

Confirm current licensing and hosting responsibilities before choosing a deployment.

The area of computing concerned with processing human language. It includes tasks such as extracting information, translating, classifying text and identifying themes, using a range of statistical and AI methods.

The field is much older than the current wave of AI. Early systems used hand-written grammar rules and keyword lists, which worked until a customer phrased something in a way nobody had anticipated. Statistical methods followed, learning patterns from large amounts of text, and today most of the work is done with deep learning, including large language models. So NLP is the discipline, and LLMs are one powerful family of tools within it.

Its business value is that it lets you read, at scale, what customers write in their own words: emails, reviews, chat logs, survey comments. That is the raw material for voice of the customer work and for routing messages to the right team. Not every language task needs the biggest model, though. For a stable job like tagging support tickets, a smaller purpose-built classifier can be quicker, cheaper and more consistent.

Example

A team groups incoming complaints by recurring issue.

Keep in mind

Check performance on the languages, terminology and writing styles used by your customers.

A metric based on a zero-to-ten recommendation question. It subtracts the percentage of detractors, scoring zero to six, from promoters, scoring nine or ten; scores of seven or eight are passives.

Because it subtracts one group from another, NPS is not an average, and it behaves in ways that surprise people. Passives count towards the total number of responses but neither add nor subtract, so a customer base made up entirely of passives scores zero, however content they are. Turning a detractor into a passive lifts the score just as much as turning a passive into a promoter.

It works best as a trend within your own customer base, measured the same way each time, rather than as a comparison with other companies whose methods and customers differ. With only a small number of responses the score can jump around from one round to the next, so look at a longer stretch before reacting. Where a customer satisfaction score usually asks about a single interaction, NPS speaks to the relationship as a whole. The habit that pays off most is closing the loop: someone gets back to unhappy customers quickly and fixes what they can.

Example

A business tracks recommendation sentiment among a defined customer group over time.

Keep in mind

The score ranges from minus 100 to plus 100; investigate the reasons behind it.

An AI model made available under terms and with materials that support use, study, modification and sharing. Publishing model weights alone does not establish that a model meets an open-source definition.

A model is more than a download. There are the weights, the vast set of learned numbers that make it work, the code used to train and run it, and information about the data it learned from. When only the weights are published, often under a licence with conditions attached, people call it an open-weight model. The open-source definition goes further and also asks for the full code and enough detail about the training data that a skilled person could build a substantially equivalent model.

The appeal for a business is control. You can run an open model on your own servers or with a hosting provider of your choice, decide where the data goes, adapt it to your tasks and avoid depending on one supplier’s pricing or product decisions. In return you take on the hosting, updates, security and testing that a commercial service would otherwise handle. Having a provider host an open model for you is a middle route between the two.

Example

A technical team reviews a model’s licence and released materials before planning its own deployment.

Keep in mind

Check use restrictions, documentation and support requirements; open access is not the same as unrestricted use.

Technology that converts text visible in an image or scanned document into machine-readable text. It makes previously image-based information available for search, copying or further processing.

The software first tidies the image, straightening a skewed scan and sharpening the contrast, then finds the regions that contain text and recognises the characters within them. Modern engines use machine learning to read whole words and lines rather than matching letters one at a time. Alongside the text they usually record where each word sits on the page, which later steps use to tell a total from a reference number. Results drop with poor input, and low-resolution phone photos, creases, stamps across the text and handwriting all cause errors.

You probably use it already, in scanner software, PDF tools, expense apps that read receipts and accounting packages that capture bills. A searchable PDF is simply a scan with an invisible layer of recognised text behind the image. Many PDFs need no OCR at all, because they were created digitally and already contain real text. Multimodal AI models can now read text from images directly, but classic OCR is usually faster and cheaper for large volumes of clean scans.

Example

A scanned delivery note becomes searchable text.

Keep in mind

Recognising characters is different from validating what a document means or extracting the correct business fields.

Coordinating the sequence and dependencies of tools, services and human steps in a process. Orchestration determines how work moves between components and what happens after a result, failure or approval.

A conductor plays no instrument but decides who comes in when, and orchestration does the same for a process. The orchestrator keeps track of where each case stands, what has already happened and what it is waiting for, such as a reply from another system, a colleague’s approval or a deadline. That is what separates it from simple automation, which fires one action when one thing happens. Orchestration handles processes with branches, parallel steps and cases that stay open for days.

In AI systems the word also covers deciding which model, agent or tool handles each part of a task and passing the right context between them. That can be fixed in advance as a defined workflow, or decided on the fly by a model choosing the next step. Fixed designs are more predictable and easier to test, while model-led ones adapt better to unusual cases. Either way, you should be able to see at a glance where every case is right now.

Example

A document workflow waits for validation before updating the accounting system.

Keep in mind

Define timeouts, retries and recovery so partial failures do not silently lose work.

A limited real-world trial used to learn whether a proposed solution works for its users and operating conditions. A pilot should have a defined scope, success criteria and a decision about what follows.

A pilot puts a new way of working in front of the people who will actually use it, with real customers, real volumes and the usual interruptions. It sits between a proof of concept, which asks whether the idea can work at all, and a full rollout, which asks everyone to change how they work. Keeping it small means mistakes stay cheap and easy to see.

The part most pilots skip is the ending. Agree before you start what you will measure against your baseline, how long the trial runs, and which results mean stop, adjust or expand. Without that, a pilot can drift on for months as a half-solution nobody quite owns. A good pilot also tells you what a demo never will: how often staff work around the tool, how customers react, and how much extra effort it quietly adds to the team’s week.

Example

One team trials a support assistant on a restricted set of enquiries.

Keep in mind

Include operational ownership and support, not only a demonstration of technical capability.

Documenting the steps, decisions, responsibilities and handovers involved in completing work. A process map helps teams understand how work moves and where delays, duplication or unclear ownership occur.

Most process maps are boxes and arrows laid out in swimlanes, with one horizontal band per role or team, so you can see who holds the work at each moment. Each time an arrow crosses from one lane into another there is a handover, and handovers are where work tends to wait, get lost or get done twice. Decision points branch the map into the different routes a case can take.

You don’t need special software to start. Sticky notes on a wall and an honest conversation with the team will teach you more than a polished diagram drawn from a manager’s memory. Map the process as it runs today before designing how it should run, and mark where information is retyped, where someone waits for an answer and where nobody is sure who decides. That current-state map is also what you want in hand before automating anything, since software repeats the steps you give it, unnecessary ones included.

Example

A business maps what happens between a signed proposal and the start of delivery.

Keep in mind

Validate the map with people doing the work, including how they handle exceptions.

Using event data from business systems to investigate how processes actually run. It can reveal common routes, variations, waiting times and differences between intended and observed behaviour.

Every time someone creates, approves or ships something in a system such as an ERP or CRM, the system usually records what happened, to which case and when. Process mining gathers those records into an event log and rebuilds the journey of each order, invoice or ticket from them. Lay thousands of journeys side by side and you get a map drawn by the data rather than by memory, detours included.

The difference from process mapping is the source. A process map captures what people say happens, process mining shows what the systems recorded, and the most useful insights often come from comparing the two. It pays off mainly where high volumes run through one or two core systems, such as order handling or accounts payable. Some ERP and automation platforms include process-mining features, so check what you already have before buying a dedicated tool.

Example

Order-system timestamps reveal where approved orders repeatedly wait for the next step.

Keep in mind

Results depend on reliable identifiers, timestamps and coverage of the relevant activities.

Input that asks or guides an AI system to perform a task. A useful prompt can specify the objective, relevant context, constraints and the form of the desired response.

In practice a prompt can be a single line or a long brief with pasted emails, an attached spreadsheet and a few examples of what good looks like. Unless it is connected to your systems, the model knows only what it picked up in training and what you put in front of it, which is why a vague request gets a generic answer. Think of briefing a sharp temp on their first morning, quick and capable but new to your customers and your tone.

Good prompts read like good briefs. Say what the result is for and who will read it, include the material it should rely on, and name what to avoid. Asking for a specific format, such as a table or a short email, saves a round of rework. When an answer misses, add what the request lacked rather than rephrasing it. Prompt engineering is the more systematic version of this skill, used when one prompt will run many times inside a product or workflow.

Example

Ask for a comparison of three approved proposals using the same evaluation criteria.

Keep in mind

Provide enough relevant context and make uncertainty acceptable.

Designing and testing instructions and context to improve how an AI system performs a task. It can involve examples, clearer constraints and consistent output requirements.

For a one-off question, a good prompt is enough. Prompt engineering earns its keep when the same instructions will run hundreds or thousands of times, for example inside a tool that writes a CRM note after every sales call. The work then looks more like product testing than writing: you collect a set of real inputs, run each version of the prompt against them, and compare the outputs with what a careful colleague would have produced.

The typical moves are unglamorous. Spell out the task and the awkward cases, show a few worked examples, fix the output format, and tell the model what to do when information is missing. A change stays only if it improves results across that whole set of examples, which is why prompt engineering leans so heavily on AI evaluations. When careful prompting still falls short, the next options are usually better context through retrieval and, less often, fine-tuning a model.

Example

A team tests alternative instructions for categorising customer requests.

Keep in mind

A well-written prompt improves guidance but does not guarantee correct or safe behaviour.

An attempt to make an AI application follow unintended instructions through its inputs, including retrieved documents or tool results. The attack exploits confusion between material to analyse and instructions to obey.

A language model reads everything as text and cannot reliably tell the instructions it was given from instructions hidden in the material it works on. A line buried in an email, web page or PDF that tells the assistant to ignore its earlier instructions and forward the conversation may be followed as an instruction rather than read as content. Users can try this in a chat, but the indirect route, through content the assistant reads on its own, is harder to spot because the planted text can be hidden from human readers.

The risk grows with what the assistant can do. A chatbot that only answers questions might embarrass you, while an agent that can send emails, change records or make payments can do real damage. No known technique makes a model fully immune, so the defences sit around it: narrow permissions, human approval before consequential actions, and keeping sensitive data away from tools that read outside content. Ask suppliers how their product handles documents containing instructions.

Example

A retrieved document tries to redirect an assistant into revealing unrelated information.

Keep in mind

Treat external content as untrusted and enforce action permissions outside the model.

A focused experiment that tests whether a specific technical idea or assumption is feasible. It creates evidence for a decision before committing to a complete operational solution.

A proof of concept answers one narrow question, usually whether something can be done at all with your data and your systems, and deliberately ignores almost everything else. It is often built quickly by one or two technical people and thrown away afterwards. That is fine, because its value is the answer, not the software.

It is easily confused with a pilot. The PoC tests the technology on a sample in controlled conditions, while the pilot tests the whole solution with real users in daily work. A useful PoC is defined before anyone starts: which sample it runs on, what result counts as success, and what you will decide if it fails. Agreeing that last point in advance stops a disappointing result from turning into a vague request for more time.

Example

A team checks whether key fields can be extracted from a representative document sample.

Keep in mind

Feasibility does not establish production reliability, user acceptance or financial value.

Processing personal data so it cannot be attributed to a person without separately held additional information. Pseudonymised data remains personal data under the GDPR.

In practice you replace direct identifiers, such as a name, email address or customer number, with a code, and keep the key that links codes back to people separately, under tighter access. Analysts can then work with the records without seeing who is who, while the business can still re-link a record when it has a legitimate reason. That reversibility is the key difference from anonymisation, where the link to a person is meant to be gone for good and the data generally falls outside the GDPR.

It is a sensible safeguard when sharing data with an outside analyst or testing a new AI tool on real customer records. The GDPR names it as one of the measures that can reduce risk, but it does not lift your other obligations. The key file needs protecting as carefully as the original data, because whoever holds it can undo the whole exercise. For sensitive or large-scale processing, the regulator’s guidance and a privacy specialist are the right next stop.

Example

A research dataset uses coded identifiers with the identifying key held separately.

Keep in mind

Removing names alone may leave other ways to identify someone.

A model designed to use additional internal computation when working through a task. Depending on the model and task, this can help with planning, complex analysis or checking alternative approaches.

A standard language model starts writing its answer straight away. A reasoning model first works through the problem in a series of intermediate steps, trying an approach, noticing errors and revising, and only then writes the final answer. That thinking is often hidden or shown as a short summary, and many products let you choose how much effort the model spends. The extra steps are made of tokens too, so they cost time and money even when you never see them.

The extra thinking pays off on tasks with several moving parts: reconciling figures, checking a contract against a policy, planning a sequence of steps or tracing an error in a spreadsheet formula. For a quick customer reply, a summary or a simple classification, a faster standard model usually does just as well. A sensible setup often uses both, sending only the harder requests to the reasoning model. Careful reasoning still starts from the information provided, so a wrong input simply produces a more carefully argued wrong answer.

Example

A system evaluates several constraints before suggesting a delivery schedule.

Keep in mind

Compare quality, response time and cost on your workload; more reasoning is not always better.

An approach that retrieves relevant information and supplies it to a generative model when answering a request. It can connect an assistant to maintained documents without changing the model’s trained parameters.

RAG works in two steps. When a question comes in, a search component looks through your documents, help articles or past tickets and picks out the passages most likely to help. Those passages are added to the prompt alongside the question, and the model is asked to answer from them, ideally naming which source it used. It is much like handing a new colleague the right pages of the handbook before they reply to a customer.

The appeal for a business is upkeep. Update a policy document and, once the search index has picked up the change, the assistant’s answers follow without any retraining. Much of the quality lives in the retrieval step: how documents are split into passages, whether outdated versions are removed, and whether the search understands how customers actually phrase things. When a RAG assistant gets something wrong, look at what it retrieved before blaming the model.

Example

An assistant retrieves the applicable service policy before drafting a response.

Keep in mind

Test retrieval quality and source access; relevant-looking passages can still lead to incorrect answers.

A measure of net benefit relative to investment cost. For a defined period, a simple calculation is: (benefits minus costs) divided by costs, multiplied by 100.

The formula is simple, the inputs are not. Costs should cover everything the change requires: licences, setup and integration, the hours your team spends learning and testing, and the ongoing effort to keep it running. Benefits are what you can measure against a baseline, such as fewer errors, invoices paid sooner, more enquiries converted or overtime avoided. Leave out costs, or count benefits you hope for rather than ones you can observe, and the percentage looks impressive while meaning very little.

Two habits make ROI more honest for AI and automation projects. First, estimate a range rather than a single number, including a cautious case you would still be happy with. Second, agree in advance how the benefit will be measured and who will check it once the change has bedded in, so the business case and the result can be compared. ROI is also only one lens, and payback period, risk reduced and customer experience improved can matter just as much.

Example

Illustration: €15,000 of measured benefit against €10,000 of total cost gives 50% ROI.

Keep in mind

State the period and assumptions; saved time is not automatically a cash saving.

Software automation that carries out defined actions in digital systems, often through their user interfaces. It is commonly used for repetitive tasks such as transferring records or completing standard forms.

An RPA bot works a screen the way a person would. It logs in, clicks through menus, copies a value from one window and types it into another, following recorded or scripted steps exactly, at any hour and without typing errors. That makes it a practical bridge to older systems with no API, where the only way in is through the same screens your staff use.

Its dependence on the screen is also its weak spot. A moved button, a renamed field or an unexpected pop-up can stop a bot, so someone has to monitor it and repair it when an application changes. Where a system offers an API or a ready-made integration, connecting through that is usually more dependable. RPA suits stable, rule-based tasks on systems you cannot easily change, and deserves a second thought when the process itself still needs fixing.

Example

A software robot enters validated records into an older application.

Keep in mind

Modern RPA platforms can incorporate AI and document processing; capabilities are not limited to screen clicking.

Searching for content using meaning and context rather than relying only on exact matching words. Many implementations compare embeddings, sometimes combined with keyword search and filters.

Keyword search looks for the words you typed, so it struggles when a customer writes “refund” and your help article says “reimbursement”. Semantic search converts both the query and every document into embeddings, lists of numbers that place texts with similar meaning close together, and returns the nearest matches. It copes better with full questions, loose phrasing and, with the right model, other languages, because it matches the gist rather than the wording.

It has its own blind spots. Exact strings such as product codes, invoice numbers and surnames are often better served by plain keyword matching, which is why many systems combine the two, an approach usually called hybrid search, and add filters for date, product or department. Semantic search is also the engine inside many RAG assistants, so the quality of an assistant’s answers often depends on whether this step finds the right passage in the first place.

Example

A search for ending a subscription finds a page titled cancellation policy.

Keep in mind

Test real customer phrasing, specialised vocabulary and access restrictions.

AI use that takes place outside an organisation’s approved tools, oversight or normal controls. It can include unreviewed applications, personal accounts or unofficial integrations used for work.

Shadow AI rarely starts with bad intent. Someone finds that a free chatbot drafts their emails much faster, a browser extension summarises meetings, or a colleague connects an AI tool to the shared drive to save an afternoon. Each choice makes sense on its own. Together they can mean customer data, contracts or pricing flowing into services nobody has checked, under terms nobody has read, from accounts the business cannot see or close when someone leaves.

Banning everything tends to push the use further out of sight rather than stopping it. A better first step is to ask people what they already use and why, which shows you where the real demand is without anyone feeling caught out. From there, a short AI policy, sensible settings on the tools you approve and some plain training on which data can go where will go a long way. Expect to revisit the list regularly, since new tools appear faster than any approval process can keep up with.

Example

A team uses an unapproved summarisation service for internal customer discussions.

Keep in mind

Provide practical approved options and a clear route for requesting new tools.

The amount of engineering, integration and ongoing management a solution requires. Useful choices range from configuring an existing tool to connecting a workflow or building a dedicated application.

Picture the options from lightest to heaviest. At the light end you configure a tool you already pay for, such as a chat assistant with your own instructions and documents. In the middle you connect systems, so a workflow moves data from your inbox to your CRM with an AI step along the way. At the heavy end you build something of your own, with custom code, hosting, security reviews and someone responsible for keeping it running.

Each move towards the heavy end adds capability, and also cost, risk and reliance on specialist skills. Much of that cost arrives after launch, in maintenance, updates and support, which is why the lightest option that genuinely solves the problem is a sensible default. When a supplier proposes a custom build, ask what a simpler version would miss, who will maintain it in a few years and what happens when the person who built it moves on.

Example

A team compares a reusable assistant with a custom portal for the same task.

Keep in mind

More complex is not more valuable. Choose the lightest option that solves the problem and that someone in your team can look after.

A model response arranged according to a specified structure, such as fields in a JSON schema. Supported features can constrain the format so other software can process it more consistently.

Left to itself, a language model answers in flowing text, which suits a person and frustrates software. With structured output you describe the shape you need in advance, for instance a field for the customer’s name, a date in one agreed format and a category picked from a set list, and the model fills in that shape. Some platforms enforce the shape strictly while the answer is being generated, whereas others rely on instructions plus a check afterwards.

This is what lets AI sit inside a workflow instead of beside it. Once an email or a scanned form becomes predictable fields, the next steps can be ordinary automation: create the CRM record, route the ticket, start an approval. Simple checks on each field, such as whether a date is plausible or the line items add up to the total, catch many problems before they spread into other systems.

Example

An extraction step returns supplier name, invoice date and total in defined fields.

Keep in mind

Valid structure does not establish correct values; handle refusals and incomplete responses too.

Application-level instructions that establish how an AI assistant should behave, such as its role, boundaries and response style. Products and APIs differ in their instruction roles and precedence.

The system prompt is the standing brief behind an assistant, written by whoever sets it up and applied to every conversation, usually out of the user’s sight. It typically covers who the assistant serves, what tone to use, which topics to avoid, which sources to trust and what to do when it is unsure. When you set up a custom assistant in one of the mainstream chat tools, the instructions field you fill in plays much the same role.

A good one reads like a clear job description, with a few concrete examples of good and bad answers, and gets revised as real conversations reveal its gaps. Its limits matter as much as its content. Models follow these instructions most of the time, not always, and a determined user or a malicious document can sometimes talk an assistant out of them. Hard rules about data access and actions therefore belong in the software around the model, not only in the prompt.

Example

An assistant is instructed to identify missing evidence before making a recommendation.

Keep in mind

Keep confidential secrets out of prompts; instructions are not secure storage.

A unit used to represent information for a model. For text, a token may be a word, part of a word or a character; other inputs may use different representations.

Before a model reads anything, the text is cut into pieces drawn from a fixed vocabulary. Common short words are often a single token, while longer or rarer words and names are split into several pieces. The model reads tokens, predicts the next one, and writes its answer one token at a time.

Tokens matter to a business for two practical reasons. Most AI APIs charge by the token, counting what you send and what comes back, usually at different rates, so long documents and long answers cost more. Every model also has a context window, a maximum number of tokens it can handle at once, which limits how much material you can give it in one go. Reasoning models add a third consideration, because the tokens they spend thinking are usually billed even when you never see them.

Example

A long report consumes more input tokens than a short question.

Keep in mind

Token counts vary by model and language; words and tokens are not interchangeable.

A neural-network architecture that uses attention mechanisms to model relationships between parts of its input. Transformer-based models underpin many modern language and multimodal systems.

Earlier language models mostly read text in order, one word after another, and tended to lose track of what came much earlier in a long passage. Attention lets a transformer look at a whole passage together and work out, for each word, which other words matter most for understanding it. Because that work can be spread across many processors at once, transformers could be trained on far more text than earlier designs, which opened the way for large language models.

The T in GPT stands for transformer, and the same family of designs now sits behind many chatbots, translation tools and transcription services. You will rarely need to choose an architecture yourself. The word is mostly useful for following technical conversations and product documentation, where it tells you the broad family a model belongs to.

Example

A language model uses contextual relationships to interpret an ambiguous sentence.

Keep in mind

Architectures vary; the term alone does not specify a model’s quality, limits or capabilities.

Information that does not fit neatly into a predefined table of fields. Examples include documents, messages, recordings and images, although these items can also carry structured metadata.

Think of the difference between a row in your invoicing system and the email thread that led to it. The row has fixed columns that any report can add up. The thread holds the context, the promises made and the reason the customer nearly left, but no software can count any of that without first reading and interpreting it. In most businesses much of what people know lives in this second kind of place: inboxes, call notes, PDFs, chat messages and shared folders.

For a long time this material was hard to use at scale, because getting meaning out of it needed a person to read it. Language models, OCR and speech-to-text have made it practical to tag large volumes of emails by topic, pull key terms from contracts or search call recordings for recurring complaints. The usual pattern is to turn unstructured inputs into structured records your systems can act on, while keeping the original attached so anyone can check the source.

Example

Customer complaints arrive as free-text messages and attached photographs.

Keep in mind

Extracting usable fields requires interpretation and validation, not just moving files between systems.

A database or database capability for storing and searching numerical vectors. It supports similarity search, often alongside metadata filters, to find information represented by related embeddings.

An embedding turns a passage of text, an image or a product description into a long list of numbers that captures its meaning. A vector database stores huge numbers of these lists and can quickly find the ones closest to a new query, much as a map app finds the cafés nearest to where you stand, only across hundreds or thousands of dimensions instead of two. Each stored vector usually carries a link back to the original content plus labels such as source, date or department, so results can be filtered and traced.

You rarely buy one on its own. It usually sits inside a larger setup, most often as the memory behind a RAG assistant or a semantic search feature, and many ordinary databases and search tools now offer vector search as a built-in capability. For a small business the real question is less which vector database to pick and more whether the platform you already use does enough.

Example

A document-search application retrieves passages close to a question’s embedding.

Keep in mind

It is one retrieval option; permissions, source quality and search design still matter.

An informal term for creating software by describing an intended result to an AI coding tool and iterating on what it produces. Usage varies, especially in how closely the person inspects the generated code.

The phrase describes a loose, conversational way of building. You say what you want, run whatever the AI produces, and when something looks wrong you describe the problem and let it try again. In its purest form the person barely reads the code and judges the result only by whether it seems to work. Others use the term more loosely for any heavily AI-assisted programming, so it is worth asking what someone means by it.

For non-technical teams it is a useful way to sketch an idea. A quoting tool, a dashboard mock-up or a simple internal form can go from description to something clickable in an afternoon, which makes conversations with developers and suppliers far more concrete. The risks sit in what you cannot see: how it stores data, who can reach it, and how it behaves with odd inputs or more than a handful of users.

Example

A founder experiments with an internal calculator by requesting changes in ordinary language.

Keep in mind

Useful prototypes still need appropriate engineering, verification and maintenance before operational use.

A systematic approach to understanding customers’ needs, expectations and experiences through feedback and other signals. It combines listening with analysis and action across relevant parts of the business.

A VoC programme brings together what customers tell you directly, through surveys, reviews, interviews and complaints, with what they show you indirectly, through support tickets, sales call notes, cancellation reasons and how they actually use your service. The point is to see these as one picture rather than as scattered inboxes owned by different teams. Done properly it runs as a loop: collect, find the patterns, act on them, and tell customers what changed.

In a small business the raw material usually exists already, and the gap is in reading it regularly and deciding something. A short monthly review of the top complaint themes with the people who can fix them often achieves more than a large survey nobody acts on. Scores such as CSAT tell you how customers feel, the comments next to them tell you why, and you need both. AI can tag free-text feedback by theme and sentiment at volume, though someone should still read a sample to check the tags make sense.

Example

A team combines interview findings with complaint themes to prioritise service improvements.

Keep in mind

Feedback represents the people and channels captured; check who is missing.

A way for one system to send another system a notification when a specified event occurs. The receiving application can use that message to begin a workflow or update its state.

Without webhooks, one system has to keep asking another whether anything has changed, a habit called polling that is slow and wasteful. A webhook turns that around. You give the sending system a web address that belongs to the receiving system, and the moment the event happens it posts a short message there with the details, such as which order, which customer and what changed. It is the difference between checking the letterbox every hour and having the postman ring the bell.

Webhooks sit behind much of everyday automation: a web form that creates a CRM record, a new booking that posts in a team chat, a signed contract that kicks off onboarding. Most automation platforms can start a workflow from an incoming webhook, which makes them a common way to link tools from different vendors. When an automation suddenly stops firing, a changed or expired webhook address is one of the first things to check.

Example

A payment service notifies the order system when a payment is confirmed.

Keep in mind

Authenticate messages and handle duplicates or retries before taking consequential actions.

A defined sequence of tasks, decisions and handovers used to complete work. Workflows can be manual, automated or mixed, and may contain AI steps within an otherwise predictable process.

Every workflow has a trigger that starts it, a series of steps, decisions that send work down different paths, and an end point where someone can say it is done. Automation tools show the same idea as a chain of connected blocks: when this happens, check that, then do these things in order. Each step has an owner, which can be a person, a piece of software or an AI model, and the handovers between them are where delays and errors usually creep in.

Adding AI tends to work best when the AI handles a contained step, such as reading an incoming email or drafting a reply, while the steps around it stay predictable and easy to check. This predictable path is the main difference from an AI agent, which decides its own next steps rather than following a route you designed. Even workflows that never get automated are worth writing down, because a written path makes training new starters and covering holidays far easier.

Example

An enquiry moves through capture, qualification, assignment and follow-up.

Keep in mind

Describe the business flow before choosing the software that will execute it.