34 terms you will meet in any AI conversation, written for people who have a business to run rather than a model to train. Each one says what it means and, more usefully, why it should matter to you.
The trained system that does the actual work. You send it text, it sends text back. Different models have different strengths, costs and speeds.
Why it matters: You are not buying 'AI' — you are choosing a specific model for a specific job, and that choice is measurable.
The kind of AI behind ChatGPT and similar tools. It predicts likely text, which makes it excellent at reading, summarising and drafting.
Why it matters: It is a very capable reader and writer. It is not a calculator or a database, and treating it as one is where projects go wrong.
The instruction you give the model. In a real system this is written by engineers and reused thousands of times, not typed fresh each time.
Why it matters: Most of the quality difference between a good and bad AI system is in prompts nobody outside the team ever sees.
How AI usage is counted and billed — roughly three-quarters of a word. A page of text is about 500 tokens.
Why it matters: Your bill is per token, in and out. Sending whole documents when a paragraph would do is the most common way to overspend.
How much text a model can consider at once. Beyond that limit, earlier material drops out of view.
Why it matters: It is why a system cannot simply 'read all our files'. Getting the right subset in front of the model is most of the engineering.
One run of the model — you ask, it answers. Distinct from training, which is how the model was built in the first place.
Why it matters: This is what you pay for per use. Training a model from scratch is almost never what a business actually needs.
When the model states something false with complete confidence. It is not lying — it has no notion of certainty.
Why it matters: This is the single biggest risk in any AI project, and the reason anything that must be right needs checking designed into the process.
The constraints stopping a system doing something unwanted — refusing off-topic requests, enforcing an output format, escalating when unsure.
Why it matters: The difference between a demo and something you would let near a customer or a regulator.
When text the system reads contains hidden instructions that try to hijack it — for example a document saying 'ignore your instructions'.
Why it matters: If your system reads documents from outside your business, this is a live security concern, not a theoretical one.
Designing the system so a person reviews certain outputs before anything happens, usually where the model is unsure.
Why it matters: Usually the right answer for regulated work. Full automation of a judgement call is rarely worth the risk it adds.
A number saying how sure the system is. Useful when calibrated properly, misleading when it is not.
Why it matters: Ask how the score was calibrated. An uncalibrated confidence number is worse than none, because people trust it.
Retrieval-Augmented Generation. Rather than hoping the model already knows, the system looks up the relevant documents first and gives them to the model to answer from.
Why it matters: This is how you get answers grounded in your own information, with a citation back to the source rather than a guess.
A numerical fingerprint of a piece of text that captures its meaning, so a computer can compare meanings rather than matching words.
Why it matters: It is what lets search find 'invoice dispute' when the document says 'billing disagreement'.
Searching by meaning using those fingerprints, instead of by keyword.
Why it matters: Why a good AI search finds the right record when nobody tagged it and the wording is completely different.
Forcing the model to reply in a fixed shape — specific fields, specific types — rather than free prose.
Why it matters: The difference between something you can feed into your existing systems and something a person has to retype.
A system that takes several steps on its own — look this up, then check that, then draft the answer — rather than answering in one shot.
Why it matters: Powerful for multi-step work, and considerably harder to keep predictable. Worth being sceptical of anyone selling agents as a default.
Further training a model on your own examples so it adapts to your style or task.
Why it matters: Often reached for too early. Better prompts and retrieval usually get you further, faster and cheaper — and are easier to change.
A model that handles more than text — images, audio, or documents as pictures rather than as characters.
Why it matters: What makes it possible to work with scanned paperwork and photographs, not just clean digital text.
A setting controlling how varied the output is. Low means consistent and predictable; high means creative and less repeatable.
Why it matters: Business systems that extract facts should run low. Anything claiming creativity in a compliance workflow is a warning sign.
Optical Character Recognition — turning a picture of text, like a scan or photo, into text a computer can read.
Why it matters: The first step for anything on paper. Quality here sets the ceiling for everything downstream.
Sorting each document into a category — invoice, referral, contract, correspondence.
Why it matters: Usually the highest-value automation in a document-heavy business, and among the most reliable.
Pulling specific fields out — dates, amounts, names, reference numbers — into a structured record.
Why it matters: This is the part that actually removes retyping, which is where the hours usually go.
Personally Identifiable Information — anything identifying a living person.
Why it matters: Determines your legal obligations. Handle it wrongly and the problem is regulatory, not technical.
Replacing personal details with consistent stand-ins, so records stay usable and linkable without identifying anyone.
Why it matters: How sensitive documents can be processed and shared without exposing the people in them.
Systematically testing whether the AI is actually right, against examples where the correct answer is known.
Why it matters: If a supplier cannot tell you their accuracy number and how they got it, they do not have one.
A fixed set of tasks used to compare models on equal terms.
Why it matters: Public benchmarks are a starting point. What matters is how models perform on *your* documents, which is rarely the same.
Examples deliberately kept back from development, used to test honestly at the end.
Why it matters: Testing on data the system was built against tells you nothing. This is how you avoid fooling yourself.
How long you wait for an answer.
Why it matters: Fine to ignore for overnight batch work. Critical if someone is sat waiting mid-conversation with a customer.
How much the system gets through in a given time — pages an hour, cases a day.
Why it matters: Usually the number your business case rests on, and the one worth agreeing before you start.
The connection one piece of software uses to talk to another. AI models are used through one.
Why it matters: Means the AI can plug into systems you already run, rather than being another window your team has to check.
Running the AI on infrastructure you control, rather than sending data to a provider.
Why it matters: Occasionally required by regulation or contract. Usually more expensive and slower to improve than the alternative — worth checking whether you truly need it.
Which country your data is physically processed and stored in.
Why it matters: Often the first question your legal team asks. Most major providers can keep processing inside the UK or EU if you configure it deliberately.
A cap on how many requests you can make in a period.
Why it matters: Why a system that flies in testing can stall under real load. Worth designing for rather than discovering.
Tracking changes to instructions the way you track changes to code.
Why it matters: Without it, nobody can explain why the system behaved differently last month. Basic hygiene that is frequently skipped.
Nothing matches that. Try a plainer word — or ask us directly.
That is fine, and it is not your job to learn it. Describe the problem in your own words and we will translate.
Ask in plain English