Most AI governance documents get written once, filed, and never checked against what is actually happening. The board does not need a policy. It needs ten direct questions asked before any AI tool is given access to company data, and again at every renewal.
These are drawn from where the main frameworks actually converge: inventory, classification, contractual guarantees, and an audit trail. Internationally that means the US NIST AI Risk Management Framework (voluntary, published 2023) and the EU AI Act (law, phasing in from 2024). For an Australian board it also means the National AI Centre's Guidance for AI Adoption, which in October 2025 condensed the ten guardrails of the Voluntary AI Safety Standard into six essential practices, and the 2024 Privacy Act amendments. Ask them in the boardroom, not just in a policy binder.
- Where is AI actually in use across the business? Not just the tools IT approved. Internal models, AI features already embedded in SaaS you pay for, and the tools staff have brought in themselves. You cannot govern what you have not inventoried.What good looks like: One inventory that covers approved tools, AI features inside the software you already pay for, and the tools staff brought in themselves.
- What does “shadow AI” look like inside this business specifically? Most leaders have no visibility into how much sensitive data staff paste into public AI tools. Find out before deciding what to do about it.What good looks like: A measured picture of what staff paste into public AI tools, gathered before anyone decides what to do about it.
- Which data classification tiers exist, and are they mapped to AI approvals? High-sensitivity data (contracts, financials, customer records) should only ever reach AI tools that meet a correspondingly high bar. If that mapping does not exist yet, that is the gap.What good looks like: A written mapping from each data classification tier to the AI tools approved for it, with contracts, financials and customer records held to the highest bar.
- Does every AI vendor we use offer a genuine zero-retention mode? Not a marketing claim. A contractual guarantee that our data is never used to train a model outside our control, and that we can prove it.What good looks like: A zero-retention clause in each AI vendor contract, and a way to prove it holds.
- Who can see what, and does the AI tool respect it? An AI layer that ignores existing permission boundaries creates a new exposure path, even if every individual answer looks reasonable.What good looks like: The AI tool enforces the permission boundaries people already work within, and opens no new access path.
- Is there an audit trail for every AI action, not just every AI answer? Who asked what, what was returned, and what the AI did as a result. Without this, an incident review has nothing to review.What good looks like: A record of who asked what, what came back and what the AI did, complete enough to run an incident review from.
- Is there a kill switch, and has it ever been tested? The ability to immediately cut an AI tool's access when something goes wrong is only real if someone has actually pulled it once, in a drill.What good looks like: A named way to cut an AI tool's access at once, pulled at least once in a drill.
- What happens when an AI tool is wrong? A defined escalation path, not an assumption that a human will catch it. Workday's January 2026 research found that nearly 40% of the time saved with AI is lost again to rework: correcting, rewriting and verifying its output.What good looks like: A defined escalation path for wrong output, so catching errors never rests on an assumption that someone will notice.
- Which framework are we actually building against? In Australia, the Guidance for AI Adoption is voluntary, but the Privacy Act is not: from 10 December 2026, privacy policies must disclose where a computer program makes, or substantially helps make, decisions that significantly affect individuals using their personal information. The EU AI Act binds anyone selling into the EU, and NIST's framework is the common reference for risk management. Know which apply, and where the gaps are against each.What good looks like: A written list of the frameworks that apply, the known gaps against each, and privacy policy disclosures ready before 10 December 2026.
- Who owns this, ongoing, not just for the rollout? Governance that lives with one project team decays the moment that team moves on. It needs a permanent owner and a recurring review date.What good looks like: A permanent named owner and a recurring review date that do not depend on any one project team.
Where this connects to the Org Brain
A governed intelligence layer answers part of this list by construction. It resolves identity and permissions before any data is read, so it respects who can see what (question five). It keeps a record of each request, what was read and what came back, which is the core of question six. And it reads through the permissions that already exist rather than around them, so it adds no new access path.
The other questions, from the inventory to the vendor contracts to the named owner, stay with the business. Good architecture makes those easier to evidence. It does not answer them for you.