Agent
Software that can decide on a next step and take it. The word says nothing about how much it is allowed to do, which is why we prefer to talk about trust levels.
AI employee
An agent with a job description, an identity, memory of your organisation and a granted boundary. Owns a piece of work rather than answering questions about it. More here.
Approval queue
Where supervised work waits for a person. A queue that shows only a recommendation trains people to click yes; one that shows the reasoning lets them review.
Audit trail
A sequence of events showing what the software did. Necessary, but not the same as a decision record, which says what was concluded and why.
Autonomy
How much an AI employee may do without asking. Here it is granted by a person, stated in business terms, and revocable in one click, never a property of the model.
Ceiling
A value limit inside a granted boundary: approve up to this, escalate above it. Enforced outside the model, so nothing written in a document can raise it.
Context window
How much text a model can consider at once. Often confused with memory; it is working space for one task and is gone afterwards.
Decision record
What was decided, on which documents, by which AI employee at which trust level, and who approved it. Produced as a by-product of the work. More here.
Demotion
Moving an AI employee down a trust level. Should take one click and no engineer. If it does not, autonomy was not really granted, it was assumed.
DPA
Data processing agreement. Sets out that you are the controller, we are the processor, and what we may do with personal data on your instruction. Ours is here.
Escalation
Handing an item to a named person because it falls outside the boundary. Should trigger on the shape of the work, not on a confidence score nobody can defend.
Exception
The item that does not match the pattern. In document work the exceptions are the job, so a pilot on clean samples proves very little.
Grounding
Tying an answer to a specific source you can open. An ungrounded answer may be right, but you cannot check it, so it cannot be used as evidence.
Guardrail
Any constraint on behaviour. Worth asking whether a given guardrail is a prompt asking the model to be careful, or a rule enforced where the model cannot reach.
Hallucination
A fluent answer that is not true. Grounding and floors reduce the consequence; nothing eliminates the behaviour, and anyone saying otherwise is overselling.
Human in the loop
A person approves before an action lands. Only meaningful if they are shown enough to actually judge it, and if the approval rate is sampled by someone else.
Job description
The written scope of an AI employee: which work, which systems, which limits, who it escalates to. Versioned, because a decision has to be read against the version in force at the time.
Knowledge graph
The structure holding what has been learned about your organisation, its entities, relationships and history, rather than a pile of document chunks. One per organisation. More here.
Memory
What is retained between tasks: who a counterparty is, how this client's contracts usually read, what happened last time. Not the same as search over a folder.
Model provider
The service running inference. Ours are named on the trust and compliance page, with what reaches each one and the fact that none may train on it.
Orchestration
Coordinating several steps, tools or agents to complete one piece of work. An implementation detail. It tells you nothing about whether the result is governed.
Owner
The named person accountable for an AI employee. Every one needs a current owner and a fallback; an unowned agent is where governance quietly stops.
Permission scope
Which systems an AI employee may reach and at what depth, enforced per request against the source system rather than held as one admin token.
Prompt injection
Text inside a document that tries to instruct the system reading it. Not a solved problem anywhere; the mitigation is floors the model cannot cross. Our position.
Provenance
Where a fact came from and when, kept with the fact. Without it a memory becomes an assertion nobody can check or correct.
RAG
Retrieval-augmented generation: find relevant text, then answer over it. Useful and widely available. It is a retrieval technique, not memory of an organisation.
Residency
Where data is processed and stored. We run in one region, the United Kingdom, and there is no residency selection to offer.
Subprocessor
A third party that processes your data on our behalf. There are three, each named on the trust and compliance page with what it does and what reaches it.
Supervised
The middle trust level: the AI employee acts, but each action is approved first. Most work should stay here longer than people expect.
Tenancy
How one customer's data is separated from another's. Ours: a dedicated knowledge graph per organisation on a shared database, with tenant binding tested in CI.
Tool use
A model calling out to software to look something up or change it. The moment tool use includes writes, governance stops being optional.
Trust level
Learning, supervised or trusted. Raised by a person on evidence, lowered in one click, and never changeable by anything the system reads. In full.