Glossary

Plain definitions, including what each term does not mean. Where a term has a fuller page, it is linked.

Accountability gap

The distance between what an AI system reports it did and what can actually be shown to have happened. See the accountability gap.

Action receipt

A portable record of one piece of completed work, including who approved it and whether an independent check confirmed the result. See action receipt.

Agent

Software that uses an AI model to decide what to do next and then does it, rather than only answering. The question that matters is not how capable it is, but what it is allowed to touch.

Approaches

Five broad ways teams try to make AI safe to use at work: pre-deployment testing, observability, post-hoc verification of agent claims, governance frameworks with AI reviewers, and governed execution. Each solves a different problem; serious production stacks combine more than one. See approaches.

Approval binding

Tying an approval to a fingerprint of the exact version approved, so that changing anything afterwards means the approval no longer covers it. Without it, "approved" is a statement about a moment, not about a document.

Attestation

The formal word, used in audit and compliance, for a signed statement that something was checked and found to be so. An action receipt is an attestation in that sense.

Capability vs. authority

What a system is technically able to do, versus what it is permitted to do. A connection that can technically write anywhere but is only permitted to create drafts has capability far beyond its authority; good design narrows the capability until the two nearly match.

Confirmed / Doesn't match / Not confirmed yet

The three possible results of an independent check. See action receipt.

Deterministic

Producing the same output for the same input, every time. The opposite of probabilistic. Determinism is not the same as correctness — deterministic code can be reliably wrong — but it is what makes checking cheap and repeatable.

Fail-closed

When something is missing or unclear, the system refuses rather than proceeds. The common alternative — treating a missing rule or an empty setting as "no restriction" — is how systems end up permitting things nobody approved.

Governed AI execution

See the full definition.

Grounding

Requiring every value a system uses to come from a real, citable source rather than from the model's own guess. A grounded value can still be the wrong business decision, which is why grounding does not replace human judgment.

Hallucination

When an AI system produces confident output that isn't true. In business workflows it rarely looks like nonsense; it looks like a plausible price, a reasonable-sounding product code, or a customer name that is almost right.

Human in the loop

A person involved in an automated process. The phrase means very little on its own: what matters is where the person sits, what they are shown, and whether the system can proceed without them.

Idempotency

Making sure that doing the same operation twice has the same effect as doing it once. Without it, a retried request can quietly become a second quote, a second invoice, or a second payment.

Independent read-back

Going to the system of record afterwards, with a separate identity that cannot write, and checking that what is actually there matches what was approved. The difference between believing a confirmation message and checking.

Provenance

Where something came from, and by what path. For a number in a report: which file, which column, which row. For a decision: who declared it, and when.

Reasoned absence

Recording why something is missing rather than leaving a blank. A blank reads as "nothing to say"; a reasoned absence says "this was not checked, and here's why" — which a reader can act on.

System of record

The place your business treats as authoritative for a given fact: your accounting system for what was invoiced, your CRM for what was agreed. Verification means checking against it, not against the AI's own account.

Tamper-evident

Built so that any change to a sealed record can be detected. Not the same as tamper-proof, and not the same as correct: a sealed record of a wrong claim is still wrong.

Unreceipted

Any AI action with no independent record of whether it actually worked. Most agent activity in production today is unreceipted.