The words, without the fog

Software that produces useful output from messy input without being given explicit rules for every case. In practice, when someone says AI today they almost always mean a large language model.

Why it matters. The word covers everything from a spam filter to a chatbot, so it is nearly useless in a procurement conversation. Ask which specific technique is being proposed and what it does when it is wrong.

See also Large language model, Deterministic logic

A model trained on very large amounts of text that predicts what text should come next. Claude, GPT and Gemini are all large language models.

Why it matters. It predicts plausible text. It does not look anything up unless you connect it to something. That single distinction explains most AI project failures.

See also Hallucination, Retrieval augmented generation

The instruction you give a model. In a production system the prompt is usually assembled by code rather than typed by a person.

Why it matters. Prompts are code. They need version control, review and tests. Treating them as copy that anyone can tweak is how a working system quietly stops working.

See also Context window, Evaluation

The unit a model reads and writes. Roughly three quarters of a word in English. Providers bill per token, in and out.

Why it matters. Token count is your unit cost. A feature that looks cheap in a demo can be expensive at volume, and the difference is usually how much context gets sent on every single request.

See also Context window, Inference

How much text a model can consider at once, prompt and response together. Measured in tokens.

Why it matters. Everything the model knows about your request has to fit in here. It is why systems get built to fetch the right ten paragraphs rather than send the whole handbook every time.

See also Token, Retrieval augmented generation

One run of a model. Sending a request and getting an answer back.

Why it matters. Training is a one-off cost that someone else usually paid. Inference is the recurring one you pay per use, forever. Model choice is largely an inference cost decision.

See also Token

Ordinary code. Same input, same output, every time, and you can read the rule that produced it.

Why it matters. A great deal of what gets proposed as AI is better done as a rule. Deterministic logic is cheaper, faster, testable and auditable. Reach for AI where the input is genuinely unpredictable, not because it is on the roadmap.

When deterministic logic beats AI

See also Artificial intelligence

A model given tools and a goal, allowed to decide which tools to use and in what order, looping until it thinks it is finished.

Why it matters. The step change is that it acts rather than answers. That is also the risk: an agent with write access can do real damage quickly, so the interesting questions are what it can reach and who approves what.

What the MCP release changes for enterprise agents

See also Model Context Protocol, Human in the loop, Orchestration

An open standard for connecting AI models to tools and data sources. One protocol instead of a bespoke integration per model and per system.

Why it matters. Before it, every AI integration was custom, so switching model provider meant rebuilding the plumbing. MCP is what makes vendor neutrality practical rather than aspirational.

Stateless, secure and extendable

See also AI agent, API

Fetch the relevant documents first, put them in the prompt, then ask the model to answer using them.

Why it matters. It is the cheapest way to make a model answer from your data rather than its training. Try this before anyone proposes fine-tuning or a custom model.

See also Fine-tuning, Vector database, Hallucination

A store that finds text by meaning rather than by keyword, by comparing numerical representations called embeddings.

Why it matters. It is the retrieval half of retrieval augmented generation. Worth knowing it is a component you can swap, not a platform you marry.

See also Retrieval augmented generation, Embedding

A list of numbers representing a piece of text, arranged so that similar meanings sit near each other.

Why it matters. It is what lets a search find the right answer when the user typed none of the same words the document uses.

See also Vector database

Continuing to train an existing model on your own examples so it adopts a particular style, format or task.

Why it matters. Expensive, slow to change, and it does not teach the model facts reliably. It is the right answer far less often than it is proposed. Try retrieval first.

See also Retrieval augmented generation, Inference

The code that decides what happens in what order: which model, which tool, what to do on failure, when to ask a person.

Why it matters. This is where most of the actual engineering lives. The model is a component. The orchestration is the system, and it is what you own.

See also AI agent, Human in the loop

A defined way for one system to ask another for something. The contract between two pieces of software.

Why it matters. If a platform has no usable API, nothing else you want to do is possible. It is the first thing to check, before features.

See also Model Context Protocol, Microservices

A model producing confident text that is wrong. Not a bug that gets fixed, a property of how the technology works.

Why it matters. Any design that assumes correct output is already broken. Systems that survive contact with production ground answers in retrieved sources and put a person in front of anything expensive to get wrong.

How we handle it

See also Retrieval augmented generation, Human in the loop, Evaluation

A written set of test cases and the criteria for a correct answer, run against the system so you can tell whether a change made it better or worse.

Why it matters. This is the single most skipped step and the most common reason a pilot never ships. Without it nobody can prove the new version is an improvement, so the team keeps swapping models and hoping.

Your pilot is not failing on the model

See also Hallucination, Prompt

Checks around a model that block or alter unacceptable input and output, separate from the model itself.

Why it matters. Instructing a model to behave is a request. A guardrail is a control. Compliance conversations need the second one.

See also Human in the loop, AI governance

A person reviews or approves before an action takes effect, rather than after.

Why it matters. It is both the practical answer to hallucination and, for high-risk uses, a legal requirement under the EU AI Act. Design where it sits early; retrofitting approval into an automated flow is painful.

Check your AI Act risk tier

See also High-risk AI, Guardrails

AI tools used inside the organisation that nobody approved, procured or knows about.

Why it matters. It is where your data actually leaks. One person installing a tool and pointing it at a repository is enough, and no policy document prevents it on its own.

One person on your IT team used Grok

See also Data residency, AI governance

The European Union's regulation on artificial intelligence. It sorts AI uses into risk tiers and attaches obligations to each, with fines up to 35 million euros or 7 percent of worldwide turnover.

Why it matters. It applies based on where the system is used, not where your company is. The prohibitions and the transparency duties are already in force. The high-risk obligations follow.

Check your risk tier in 5 minutes

See also High-risk AI, Provider vs deployer, AI governance

Under the EU AI Act, AI used in areas like hiring, credit, education, essential services or biometrics. It carries the heaviest obligations: risk management, documentation, human oversight and a conformity assessment before go-live.

Why it matters. It catches more organisations than expected. If AI touches a decision that materially affects a person, assume this tier until you have checked.

Find out which tier you are in

See also EU AI Act, Human in the loop

Under the EU AI Act, a provider builds and puts an AI system on the market. A deployer uses one. The obligations differ sharply.

Why it matters. Most organisations assume they are only deployers and are wrong, because building on top of a model can make you a provider. You also cannot rely on a vendor having done its part: buying it is not a defence.

Work out which you are

See also EU AI Act, High-risk AI

Knowing which AI systems you have, what each is used for, who owns it, and being able to show that on request.

Why it matters. It sounds like paperwork and is mostly an inventory problem. Organisations that cannot list their AI systems cannot comply with anything, and that is the usual starting position.

See also EU AI Act, Guardrails

Which country or region your data is physically stored and processed in.

Why it matters. It constrains model and hosting choice, so it is a design input rather than a deployment detail. Discovering it late usually means rebuilding.

See also Shadow AI, AI governance

Microservices, API-first, Cloud-native, Headless. A set of principles for building systems out of replaceable parts rather than one large platform.

Why it matters. The point is being able to change one thing without a replatform. That is what a monolith takes away and what it costs you every time the market moves.

Score your composable maturity(opens in a new tab)

See also Composable architecture, Headless, Microservices

Assembling a system from best-fit components connected by APIs, so any one can be replaced without rebuilding the rest.

Why it matters. It buys optionality, and it costs integration work. Done without senior architecture it produces a distributed monolith, which is worse than what it replaced.

When to go composable and when to stay put

See also MACH, Microservices

A system that manages content or commerce and serves it over an API, with no opinion about how it is displayed.

Why it matters. It is what lets one source of content feed a website, an app, a kiosk and now an AI agent. It also means the front end is now your responsibility.

See also MACH, Digital experience platform

Building an application as a set of small independent services that talk over APIs, rather than one deployable unit.

Why it matters. Independent deployment is the benefit. Operational complexity is the price. Both are real, and small teams routinely underestimate the second.

See also MACH, API

A suite that bundles content management, personalisation, analytics and campaign tools into one product.

Why it matters. Convenient at the start, expensive to leave. The composable question is whether you want one vendor's version of every capability or the best available version of each.

CMS, DXP and MACH compared

See also Headless, Composable architecture

Making your content easy for AI assistants to find, understand and quote, rather than optimising for a ranked list of blue links.

Why it matters. People increasingly ask an assistant instead of searching. If the assistant cannot read your site, you are not in the answer, and no amount of traditional ranking fixes that.

AI is answering your customers

See also Agent experience, Structured data

How well a site works for an AI agent trying to accomplish something on it, as opposed to how it looks to a person.

Why it matters. Agents are becoming a real share of traffic, and they fail on things a person tolerates: content behind script, no machine-readable pricing, forms that need a mouse.

See how the web treats AI agents(opens in a new tab)

See also Answer engine optimisation, Structured data

Machine-readable labelling embedded in a page that states what things are: this is a price, this is an author, this is a definition.

Why it matters. It removes guesswork for search engines and AI assistants. Cheap to add, and it is the difference between being quoted accurately and being paraphrased wrongly.

See also Answer engine optimisation, Agent experience