Interactive explainer
The questions to ask about your AI
Boards, regulators and auditors ask much the same eight questions about an organisation's AI. Each has a good answer only when the records behind it are kept as the work happens. Pick a question to see what a good answer draws on.
Ask the estate
Each question lights the path through an illustrative estate in Aicura that answers it, and the panel shows what a good answer draws on. Use the arrow keys to move between questions.
Pick a question to see what a good answer draws on, and how an illustrative estate in Aicura answers it.
What AI does the organisation use, and who is accountable for each system?
Asked by Boards and auditors
Every other question depends on this one. Nobody can oversee a system they do not know about or that has no named owner.
- Requests from staffNew AI use is asked for and approved before it starts.
- DiscoveryConnected sources report AI already in use, and a person decides what joins.
- The AI registerEvery system and agent with a designated owner, an accountable executive and a full version history.
60 systems and agents, each with a named owner.
Each entry also names who is accountable at executive level and a technical contact. A significant change saves a new version, and the assessments that depend on it run again.
Is any AI in use that nobody has approved?
Asked by Boards, auditors and security teams
Staff adopt AI tools faster than any approval process. The honest answer is never zero, so the question is how quickly unapproved use is found and decided.
- Your sourcesManaged devices, code repositories, IT asset records and cloud accounts report the AI they find.
- The discovery queueEach find waits for a person to add it, link it to an existing entry or ignore it.
- AI use requestsRequests show what staff have asked for and where each decision is up to.
6 finds are waiting on a person, and 3 requests are open.
Discovery proposes and a person decides. An approved request joins the register at v1 with an owner, and its risk assessment starts on its own.
Which AI systems make or support decisions about people?
Asked by Regulators and privacy teams
Decisions about credit, claims, jobs or services carry the highest duty of care, and many privacy and AI laws single them out for transparency and review.
- The AI registerEach entry records what it produces and whether it affects individuals.
- Impact assessmentsWho is affected, which rights are engaged and what mitigates the harm.
- Decision disclosuresA plain account of the decision, reviewed and accepted by a person.
9 entries make or support decisions about people.
Each has an impact assessment covering who is affected, their rights and the mitigations, and a decision disclosure that a person reviewed and accepted.
Where are our highest AI risks, and are the controls in place?
Asked by Boards, risk committees and auditors
A risk rating means little on its own. Directors want to see the controls each high risk calls for and whether they exist and have been tested.
- Risk assessmentsAn overall rating, risk items and suggested controls for each entry, as guidance for your judgement.
- Risks and controlsWhether each control applies, is in place and has been tested.
- GuardrailsControls that can be enforced become rules at the gateway.
4 entries are rated High. 23 suggested controls are not yet in place.
- Allow only the tools the entry declares.
- Escalate payment plan proposals to a person.
The rating is guidance for your judgement and approves nothing. A significant change to an entry runs its assessment again.
What have our AI controls actually stopped, and why?
Asked by Auditors and security teams
A policy that is never enforced is a statement of intent. The evidence that controls work is what they blocked or redacted on live traffic, and which rule did it.
- GuardrailsRules published as signed versions, then deployed to the gateway in your network.
- The gatewayEvery call from your systems and agents passes through it, and it allows, redacts or blocks.
- Enforcement eventsEach block and redaction with the rule that fired, kept in the history.
84 calls blocked and 131 redacted, out of about 150,000.
Each block and each redaction adds its own entry to the history. Allowed calls roll up into one entry per gateway per day.
How much can our AI agents do without a person approving?
Asked by Boards and regulators
Agents act, not just answer. Oversight depends on knowing the boundary of each agent's authority and seeing that the boundary holds.
- The AI registerEach agent declares the tools it may call.
- Rules from permissionsThe gateway allows those tools and nothing else.
- ApprovalActions you mark for approval go to a person first.
4 of 7 agents act alone, within the tools they declare.
The other 2 agents ask first too. Anything outside an agent's declared tools is blocked at the gateway, and actions you mark for approval go to a person.
When something changes in production, what happens and who decides?
Asked by Boards, regulators and auditors
AI behaves differently in production than it did when it was assessed. Oversight means knowing what is watched, what happens when a limit is crossed and which actions need a person.
- Production dataTraffic, agent decisions, spend, security tests and drift come back from your network.
- Alerts and KPIsLimits and targets you set on that data.
- Automatic responsesActions taken only at the autonomy level your team chose, and a response can re-run the assessment.
6 automatic responses this month. 2 waited for approval.
Aicura responds on its own only as far as the autonomy level your team chose. A response can re-run the risk or impact assessment, so what happens in production feeds back into the rating.
What do we show the board this quarter?
Asked by Boards and executives
Directors need the same picture every quarter, drawn from records rather than compiled by hand, so they can see what has moved and what still needs attention.
- The register, assessments and rulesReporting reads them as they stand today.
- The board reportAicura drafts it for the period. A person edits and finalises it, and the next period gets a new one.
- The report catalogue and Trust CentrePoint-in-time reports on systems, agents, assessments, incidents and security testing, and a public page built from the same data.
A Q3 board report, drafted from what stands today.
Finalising the board report and publishing the Trust Centre each add an entry to the history. An evidence pack is sealed, so anyone you share it with can confirm it has not changed.
Why these eight questions
The questions follow the life of an AI system. The first two establish what exists and whether anyone approved it. The middle three ask whether the risk is understood and whether the controls hold. The next two are about autonomy and change, which is where AI agents have moved the conversation. The last asks what the board itself sees each quarter, and it can only be answered well if the other seven can. A director does not need to know how a gateway works. They do need to know that someone can answer each question with evidence, quickly, and the same way every quarter.
The questions are not tied to one law or framework. Management system standards, risk frameworks and privacy laws word them differently, but each expects an organisation to be able to answer them, and the answers draw on the same records wherever you operate.
What makes a good answer
- It is specific. "We have an AI policy" does not answer "what AI do we use". A list of systems and agents, each with a named owner, does.
- It is current. An inventory compiled for last year's audit describes last year. A good answer comes from a register that changes when the estate does.
- It shows the control working. A rule that has never fired may be well placed or may be switched off. Enforcement events tell you which.
- It names who decided. Regulators increasingly ask who approved a decision and on what basis, especially for systems that affect people and agents that act alone. The guide to human oversight of AI covers how to design those approval points.
- It can be reproduced. If the answer lives in one person's spreadsheet, it will not survive that person's leave.
How to use the questions
If you sit on a board, ask management to answer all eight at the next meeting and note which answers took a week to assemble. Those are the places where the record is thin. Ask them in the same order each time, so movement between quarters is easy to see. The guide to AI governance board reporting shows how to turn the answers into a standing quarterly report.
If you run AI governance, treat each question as a test of your own records. Where an answer needs a hunt through email and spreadsheets, start there. The guide to assessing your AI governance maturity gives a broader view of where your programme stands, and what is AI governance covers the ground the questions sit on.
If you are choosing a platform, ask each vendor to answer the eight from a live estate rather than slides. The guide to evaluating an AI governance platform lists further questions that separate what a product does from what its website says.
How Aicura works
Aicura finds the AI you run, assesses it, enforces your rules on live traffic and reports on all of it, with a sealed history of the work underneath.
AI register
Every AI system and agent you run, found by discovery or asked for by staff. Each has a named owner, and every significant change, such as a new model, saves a new version.
Assessments
Risk assessments give each version an overall rating, risk items and suggested controls. Impact assessments set out who is affected and how. Both are guidance for your judgement.
Guardrails and gateway
Rules built from the register, assessments and permissions are signed and deployed to a gateway in your network that allows, redacts or blocks.
Monitoring and automatic responses
Alerts, targets and automatic responses that act only as far as you allow. A response can re-run the risk or impact assessment, so what happens in production feeds back into the rating.
Reporting
Board reports, the Trust Centre and the report catalogue, built from your register, assessments and rules as they stand today. Evidence packs are generated from the history.
The history. Governance decisions and enforcement actions are snapped into a sealed, timestamped history as the work happens. Nobody files anything by hand.
The register, discovery and system and agent reports are in the free Register plan. Risk assessments and guardrail authoring start at Essentials. Impact assessments, monitoring, enforcement on live traffic, board reports, the Trust Centre and the history start at Pro. Compare plans
Questions
What questions should a board ask about AI?
Eight questions cover most of what a board needs. What AI does the organisation use, and who is accountable for each system? Is any AI in use that nobody has approved? Which AI systems make or support decisions about people? Where are our highest AI risks, and are the controls in place? What have our AI controls actually stopped, and why? How much can our AI agents do without a person approving? When something changes in production, what happens and who decides? What do we show the board this quarter?
How often should a board hear about AI governance?
Quarterly suits most organisations, with significant incidents raised between meetings. The report should answer the same questions each time so directors can see movement, and it should be drawn from the governance record rather than compiled by hand. The guide to AI governance board reporting covers what to include.
What evidence do auditors expect for AI controls?
Evidence that controls exist and that they work. That usually means a register of AI systems with named owners, an assessment for the version of each system in use, the rules in force, the events showing what those rules blocked or redacted, and a record of who approved each decision and when.
Are these questions tied to a particular law or framework?
No. They apply in any jurisdiction and to any size of organisation. Management system standards such as ISO/IEC 42001 and risk frameworks such as the NIST AI RMF expect an organisation to be able to answer them, and several privacy and AI laws add specific duties for systems that make decisions about people.
Is the estate on this page real?
No. The organisation, its systems, people and figures are illustrative. What each answer draws on, from the discovery queue to enforcement events and automatic responses, is how Aicura works.
Start building the answer to the first question
Create a free account and run a discovery scan. The register it starts is what every other answer on this page draws on.