Reference
Lexicon
Working definitions for terms that get used to mean several things in the same meeting. Each entry says what we mean by it and what the word usually smuggles in when nobody defines it.
- Agentic
- A system that takes multiple steps toward a goal without a human approving each step, and can call tools that change state.
- ContestedUsed for anything from a two-step prompt chain to a system with write access to production. The question that settles it: what can it change without asking, and who is accountable when it does.
- Pilot
- A time-boxed deployment whose purpose is to produce evidence for a specific decision that a named person will make.
- ContestedUsually means an indefinite deployment with no decision attached. If nobody can say which decision the pilot informs and who makes it, it is not a pilot. See Entry 001.
- Adoption
- The share of eligible work in a defined process that now passes through the tool.
- ContestedMost often reported as accounts activated or weekly active users, which measures curiosity. It also implies a zero baseline that does not exist. See Entry 003.
- Return
- Measured effect over total cost, with the total cost itemised in the same table.
- ContestedRoutinely quoted against a licence-only denominator, which produces a number that cannot be compared with anyone else's, including your own last quarter. See Entry 002.
- Human in the loop
- A person who reviews output before it takes effect, has the authority and the time to reject it, and is accountable for the result.
- ContestedFrequently a person who receives a notification. If reviewing properly takes longer than the task saved, the loop exists on the diagram and not in practice. See Entry 007.
- Guardrails
- Controls that prevent a defined class of output or action, with a stated failure mode when they trigger.
- ContestedOften used for a system prompt asking the model politely. A guardrail you cannot test is a preference.
- Grounding
- Constraining output to retrieved source material, with the source shown alongside the answer.
- ContestedSometimes used for retrieval that happens but is never surfaced. Without visible provenance the user still has to verify from scratch, which is most of the cost.
- Production
- Running against real volumes, for real users, funded from an operating budget, with someone on call.
- ContestedApplied to anything with a URL. The operating-budget clause is the one that separates the two, and it is the one that predicts survival.
- Evaluation
- A repeatable measurement against a fixed set of cases, run before and after every change, with a threshold agreed in advance.
- ContestedCommonly a demo to stakeholders. If the pass mark was decided after seeing the result, no evaluation occurred.
- Centre of excellence
- A small team that owns shared infrastructure, evaluation standards and access patterns so delivery teams do not each solve them.
- ContestedBecomes an approval queue with a nicer name when it owns permission instead of infrastructure. The test is whether it makes delivery teams faster or slower.