A TO Z
Twenty-eight terms, and what each one costs you.
A
- AgentAlso: autonomous agent
- A system that decides what to do next rather than following a fixed script: it reads a situation, picks an action from a set you defined, and asks a person when it is unsure.The decision it affects: an agent needs approval points designed in before it is built. A script does not.Custom AI agents
- APIApplication programming interface
- The documented way one piece of software lets another read or change its data.Ask this before scoping: if a tool you depend on has no API, integrating it costs more and breaks more often, and that belongs in the estimate rather than in month three.AI integration and APIs
- Automation
- Work that used to wait for a person now happening without one.Only worth measuring against the number it was meant to move: replies sent, hours returned, leads not lost. “Efficiency” is not a measurement.AI workflow automation
B
- Benchmark
- A published score for a model on a fixed public test set.Useful for narrowing a shortlist, close to useless for predicting performance on your data, which is why verification runs against your own target, not a leaderboard.
C
- Confidence threshold
- The score below which the system stops and asks a person instead of acting.The single most important number in an automation you actually trust. Set it too high and nothing runs; too low and it invents. It should be yours to change without a developer.
- Computer vision
- Reading images or video as data: counting, classifying, detecting a defect, matching a plate.Worth it where a person is currently looking at a screen in order to produce one number or one yes/no.Computer vision
- Context window
- How much text a model can consider in a single answer.It caps how much of your document or thread the system can take into account at once. It is a limit on what it can know in one pass, not on what it can remember over time.
E
- Embedding
- A numeric representation of text, used to find things by meaning rather than exact words.What makes “find our past quotes for jobs like this one” possible when nobody tagged them.
- Evaluation setAlso: eval
- Real examples with the right answer attached, used to check a system before it goes live.If a vendor cannot tell you what they will evaluate against, they cannot tell you when the work is finished. Ask for it in writing at scope.
F
- Fine-tuning
- Further training of a model on your own examples.Rarely the first answer. Giving the model the right context at the moment it answers is usually cheaper, faster to change and far easier to audit.
G
- Guardrail
- A hard rule the system cannot talk its way past: topics it will not discuss, actions it cannot take, data it cannot read.Distinct from a prompt instruction, which is a request. Anything that must never happen belongs here.
H
- Hallucination
- A confident, fluent, wrong answer.Reduced by grounding answers in your documents and by holding output for review. Never eliminated, which is why destructive and financial actions stay behind a human step.
- Human in the loop
- A named person who approves, corrects or overrides output before it has effect.The difference between an automation you can defend to a customer and one you cannot.How we work
I
- Inference cost
- What it costs each time the system answers.It scales with volume, so a per-answer cost that looks trivial in a demo is a real line item at ten thousand enquiries a month. Ask for it per unit of work, not per month.
- Integration
- The connection between the new system and the tools you already run on.Usually the largest hidden cost in an AI project, and the part most demos skip entirely.AI integration and APIs
L
- Latency
- How long an answer takes.Matters most on voice, where anything much over a second reads as a dropped line. Barely matters at all on overnight batch work, so do not pay for speed you cannot hear.Voice AI agents
- LLMLarge language model
- The component that reads and writes text.One part of a system, not the system. The value is in the orchestration, the data it can reach and the approval points around it.
M
- Model provider
- The company hosting the model your system calls.Worth knowing which one, which region it runs in and what their retention policy says, because that answer is now part of your own data policy.Security and AI governance
O
- Orchestration
- The layer deciding which step runs in which order, what happens on failure, and where a person is asked.Boring, invisible, and the reason a system still works in month six.
P
- Prompt
- The instructions given to a model at the moment it answers.A request, not a constraint. If the answer to “what stops it doing X?” is “the prompt says not to”, that is not an answer.
- PIIPersonally identifiable information
- Data that identifies a person.Determines what may leave your systems, which provider may see it and how long anything may be kept. Settled before the build, not during it.Privacy notice
R
- RAGRetrieval-augmented generation
- Finding the right passage in your own documents and giving it to the model before it answers.The practical answer to “how does it know our policies?”, and cheaper than fine-tuning for almost every case that sounds like one.
- Retention
- How long data is kept, by you and by every provider in the chain.A number you should be able to state in one sentence per system. If nobody can, that is the finding.
S
- System prompt
- The standing instructions a system carries into every conversation: role, tone, what it may promise.Changing it changes behaviour everywhere at once, so it belongs under version control and in the handover pack.
T
- Token
- The unit models read and write in, and the unit you are billed in: roughly three quarters of a word in English.Useful only for estimating cost. Nobody should be quoting you a system in tokens.
- Tool useAlso: function calling
- A model calling a real function: checking a diary, creating a booking, sending an email.The point where an assistant becomes a system, and where approval points stop being optional.Custom AI agents
V
- Voice agent
- A system that answers or places calls: speech in, decision, speech out, and a written record afterwards.Judged on latency, interruption handling and what it does when it cannot help, not on how human it sounds.Voice AI agents
W
- Workflow
- The actual sequence of steps your business runs, including the undocumented ones.Automating a workflow nobody has written down is the most common way an AI project fails. That is what the audit is for.Book the audit