# AI Governance Assessment Version 0.1 · CC BY-NC 4.0 — attribute to Hanif Karimi, contextualagentic.com Eighteen questions to ask before an agent is given authority over anything that matters. Each has a scoring note, and the scoring notes are deliberately hard to pass by intending to. > Drafted for the companion material of The Contextual Agentic Enterprise. Not a compliance instrument and not mapped to any regulation. Use it to find the argument you cannot yet make. ## Scoring - **0** — No. Or: we intend to. - **1** — Partly — it exists for some paths, or on paper only. - **2** — Yes, and I can show you the artefact right now. ## Purpose ### GA-01 — Can you state what this agent is for in one sentence without using the word "AI"? **What a 2 looks like** — A written purpose that names the outcome and the system of record it touches. ### GA-02 — Is there a decision it must never make, written down? **What a 2 looks like** — An exclusion list, and a control that enforces at least one item on it. ### GA-03 — Who is accountable for its actions by name, not by team? **What a 2 looks like** — A named owner who knows they are the owner. ## Authority ### GA-04 — What is the largest irreversible thing it can do unaided? **What a 2 looks like** — A specific answer with a number, derived from the grant rather than guessed. ### GA-05 — Who approved that, and when? **What a 2 looks like** — An approval record with a date and a scope. ### GA-06 — How would you remove its authority in the next five minutes? **What a 2 looks like** — A procedure someone has actually executed in a drill. ## Context ### GA-07 — Where does it get its facts, and how do you know they are current? **What a 2 looks like** — A named source with a freshness guarantee and a staleness alert. ### GA-08 — Can content it retrieves change what it does? **What a 2 looks like** — A prompt-injection test suite in CI with failing cases recorded. ### GA-09 — Does it see anything the person it acts for could not see themselves? **What a 2 looks like** — Entitlement filtering at retrieval, with a test for two principals. ## Enforcement ### GA-10 — What stops it doing the thing you most fear, and is that thing code? **What a 2 looks like** — A named control outside the model, with the test that exercises it. ### GA-11 — What happens when that control is unavailable? **What a 2 looks like** — A fault-injection result showing a refusal. ### GA-12 — Has anyone tried to talk it past its controls? **What a 2 looks like** — A red-team exercise with findings and fixes. ## Evidence ### GA-13 — If it did something wrong last Tuesday, what could you show an auditor? **What a 2 looks like** — A record retrieved in front of the asker, in under a minute. ### GA-14 — Could that record have been edited? **What a 2 looks like** — Retention lock or a verified hash chain, and a verification run. ### GA-15 — Are refusals recorded as carefully as successes? **What a 2 looks like** — A refusal count you can quote, and someone who watches it. ## Oversight ### GA-16 — How many decisions a day does a human have to review, and can they? **What a 2 looks like** — A measured review load compared against measured human capacity. ### GA-17 — When did a reviewer last say no? **What a 2 looks like** — A specific recent example. "Never" is an answer worth investigating. ## Economics ### GA-18 — What is the denominator in your cost-per-task figure? **What a 2 looks like** — A definition that includes failures, reviews and rework, agreed with finance.