AI Governance Playbook

Principles

• Verify first. Check a checkable premise before building on it. A premise proved false and closed with evidence is a successful run.
• Never fabricate. Not a result, a number, a date or a test. Escalating counts as progress.
• Prove by execution. "Done" means the check ran and its result is recorded. "Should work" is not a result.
• Narrow gates, wide autonomy. A system that must ask about everything gets ignored; one that asks about nothing eventually spends your money.
• Kill before add. Every new rule, task or agent replaces something or earns its place.

The approval-gate matrix

Five classes of action an AI system never takes on its own. It prepares the work, then stops and waits for a named human.

1. Money going live — payment wiring, pricing changes, payouts, refunds, new paid seats or plans.
2. Auth and secrets — never create, rotate or commit a credential; write the code that reads an environment variable and leave the value to a human.
3. External sends and posting — email, DMs, social posts, anything that contacts a third party in your name.
4. Scheduled automation — creating, disabling or rescheduling anything that will run without a person present.
5. Destructive or irreversible change — deleting data, DNS and domain changes, production migrations.

Everything outside the five classes is allowed without asking. Write the matrix down, name who approves each class, and keep the list short.

The agent (or use-case) contract

Every agent — or every approved AI use case in a team — gets a one-page contract before it runs:

No contract, no run.

The handover ledger

One append-only file is the audit trail. Whoever builds or acts writes one line per item, in the same run:

date | item ID | source | project | STATUS | note

STATUS is one of BUILT · PARTIAL · ESCALATED · KILLED. Lines are never edited or deleted. Anyone who plans work reads the ledger before re-planning; an item that silently disappears is a handover failure, not housekeeping. The same file is your usage report: count it by team, by status, by week.

Autonomy envelopes

When a gated action must run automatically, write an envelope with every edge on paper: the one case it covers, the hard daily cap, the stop-loss that halts it, and how every automated action is tagged in the ledger. Everything outside the envelope is drafted by the machine and done by a person. The envelope answers "how much autonomy?" with a boundary you can read, not a feeling.

Measuring it

Report from the ledger, not from impressions: items built, escalated and killed per week, per team, per use case. Add cost per team where the platform logs it. A number that is derived from the system that produced it cannot quietly drift.

Mapping it onto an organisation’s rollout

• Gates → group-level access and approval rules on the AI platform (who may connect data, publish, automate).
• Contracts → one card per approved use case, with an owner, the data it may touch and its review step.
• Ledger → the usage and spend report by team that leadership reads monthly.
• Envelopes → the way restrictions are loosened: one bounded case at a time, with a cap and a stop-loss.
• Principles → the first page of the AI policy everyone actually reads.

The model is platform-neutral: it works the same whether the seats are Claude, Copilot or something else.