How Do Enterprises Govern AI Agents? What the Index Asks, Dimension by Dimension
Governing AI agents comes down to rules written where the work happens. The Index checks eight of them. Here is each one, the control, and the question to ask.
Governing AI agents comes down to eight rules written where the work happens: a human yes, verified facts, minimum access, test then watch, stop and ask, written criteria, undo, and ownership. The AI Readiness Index asks 18 questions across them and returns a score and one of four bands.
Governance questions tend to arrive as a policy document. A policy says what should happen. The Index asks whether the rule exists somewhere a person could point to, and who would notice if it stopped being followed. The Index works from the other end. Every question checks a management habit that produces reliable AI work, and full credit on a question means you could show where the habit is written down. Habit alone does not count. The parts below take the eight dimensions one at a time. Each gives the rule, the control behind it and the question to put to the team that runs the AI agent.
The Index scores 18 questions across eight weighted dimensions
The Index asks 18 plain-language questions in about four minutes. Each answer is No, Partly or Yes, worth zero, partial or full credit, and full credit means the rule is written down. The eight dimensions carry different weights out of 100: The Yes Gate 20, Truth in Every Draft 15, Scoped Access 15, Graduated Trust 12, Stop and Ask 10, Written Criteria 10, Ownership and Skills 10, and Reversibility 8.
The total places a team in one of four bands. Ad hoc is below 40: AI is in use, and fewer than half of the rules are written down. Watched is 40 to 64: humans review the work, and many of the rules are not written down yet. Gated is 65 to 84: some of the rules are written down. Governed is 85 and above: most of the rules are written down. The AI Readiness Index is free, takes no signup, and ends with the top three fixes ranked by impact.
The Yes Gate governs what reaches the outside world
The rule: nothing sends without a yes. The control is a named human who clicks yes before any AI-assisted action that reaches the outside world, meaning send, publish, pay or delete, plus a written draft-versus-send rule in every tool's instructions. The Index's own fix for a gap here is often one line: draft, never send.
Ask the team: for each AI tool in use, can anyone say which actions it takes alone and which wait for a person? The Index asks exactly that. It carries the most weight of the eight.
Truth in Every Draft governs claims and numbers
The rule: never invent a credential, a number or a relationship. The control has two parts. The AI tool is instructed to ground its output only in verified facts you gave it, and a person verifies claims and numbers before anything goes to a client, a donor or the public.
Ask the team: before something with a number in it goes out, which person checks it, and where is that written?
Scoped Access governs what each tool can reach
The rule: minimum access, revocable in one click. The control is a scoped, revocable credential for each tool instead of anyone's real password, access trimmed to the one job, and a known path to shut a tool off.
Ask the team: if a tool misbehaved today, could you revoke its access in under five minutes without breaking anything else? The Index's fix for a no is a one-page list of each AI tool, the login it uses and who can turn it off. The page on what an AI agent can touch shows the settings behind this one.
Graduated Trust governs how a workflow earns a schedule
The rule: test case first, watched before scheduled. The control is a throwaway test case before a new workflow touches real data, then runs reviewed by a person before anything goes on a schedule.
Ask the team: which of your AI workflows ran watched before they ran unattended, and who watched?
Stop and Ask governs what happens when something is unclear
The rule: stop and ask beats a confident guess. The control is an instruction, to the tools and to the people using them, to stop when something is unclear, and a written who-to-ask path when output looks wrong, with a named person rather than a shrug.
Ask the team: when AI output looks wrong, whom does a person ask, and is that written where the work happens?
Written Criteria governs what good output means
The rule: vague instructions get you vague results. The control is a one-sentence goal written before each AI workflow is built, and the judgment criteria written down: thresholds, filters, exclusions and tone, rather than carried in one person's head.
Ask the team: can someone who did not build the workflow read its goal and its criteria and say what good output looks like?
Reversibility governs the way back
The rule: nothing is ever deleted, and recovery is "undo that." The control is a bar on AI tools deleting or overwriting anything without a human yes, plus a known undo path such as versions, drafts or backups.
Ask the team: if an AI action turned out wrong this morning, how would you put things back?
Ownership and Skills governs who keeps the AI agent
The rule: you own everything you build. The control is organizational ownership of the prompts, AI agents and credentials, so nothing leaves with a vendor or a departure, and training on these rules for everyone who touches the tools, not only the person who set them up.
Ask the team: if the person who built this AI agent left on Friday, who owns it on Monday? The page on keeping an AI agent running after its builder leaves covers the account side of that handover.
From the score to the next step
BNEDai's guide to the score bands says what each band means and the next step for it. For the definitions behind the dimensions, see What Is AI Governance? and What Are AI Operating Rules?. If you would rather run this inside one department, the pilot playbook puts the same rules in the order a pilot needs them.
Frequently asked questions
How many dimensions does the Index cover?
Eight, across 18 questions: The Yes Gate, Truth in Every Draft, Scoped Access, Graduated Trust, Stop and Ask, Written Criteria, Reversibility, and Ownership and Skills.
What are the four bands?
Ad hoc is below 40, Watched is 40 to 64, Gated is 65 to 84, and Governed is 85 and above.
Does the Index compare my team with other companies?
No. It is a self-assessment of how your own team runs AI, and full credit on a question means you could show where the rule is written down.
Reviewed on November 20, 2026.