AI governance: an operating register and practical use-policy template
AI governance is how an organisation assigns responsibility, sets boundaries and reviews decisions about AI use. Start with a register of specific uses, the data each may receive, an accountable person and evidence for approval. A practical use policy tells staff what to do when a request falls outside that recorded scope.
By Ai Flowi · Published
Scope: static educational worksheets. The examples below are fictional organisational decisions, with no recorded deployment, certification or product-enforced governance capability.
1. Register a use, not just an AI tool
A tool name is insufficient to describe what was authorised. Drafting an announcement from public information and analysing confidential employee notes are different uses, even if they involve the same model. Give each use a stable reference and describe the task, audience, permitted inputs and intended output.
Before filling the worksheet, gather the proposed task, the source material, its handling restrictions and the people responsible for the work. No software account or API key is needed for this manual exercise. If the eventual use involves a provider, record its identity and the approved handling arrangement before anyone submits real material.
Responsible AI practices become reviewable when a broad aim has a concrete decision attached. For example, “be accountable” becomes a named person who can suspend a use, a reviewer who checks its output and a record explaining the permitted scope. The register below turns those responsibilities into fields.
2. Copy the operating register
Make one record per use and retain earlier decisions when the scope changes. Square brackets mark blanks to complete; they are worksheet fields, not AI Flowi Workflow settings.
AI-use reference: [stable label] Task and intended audience: [specific purpose] Accountable person: [name or assigned role; confirmed contact] Permitted data: [source, category and approved version] Prohibited inputs: [material excluded from this use] Tool or provider under review: [identity and relevant configuration] Output and consequence: [draft, decision support or proposed action] Evidence: [reviewed samples, source references and unresolved gaps] Reviewer: [person responsible for the stated review] Approval state: [pending / allowed within scope / blocked / suspended] Decision record: [approver, rationale, scope and recorded date] Exception: [none, or separate request and decision reference] Exception expiry: [end point and person checking closure] Review trigger: [change, incident or scheduled review point] Stop and escalation route: [who to notify; where work pauses] Record handling: [access, retention decision and evidence location]
Use “pending” when evidence is missing. “Allowed within scope” needs a decision tied to the task and data actually reviewed. “Blocked” records a proposed use that may not proceed; “suspended” records a pause of an earlier permission. These are suggested manual states, not automatic routing behaviour.
3. Add a qualitative AI risk assessment
Write the failure and its consequence in words. Ask who could be affected, what information could leave the approved boundary, how an incorrect output could be used and whether a person can check it before action. Record a control and the evidence still needed for each concern. Avoid turning uncertain answers into an unsupported numerical score.
On narrow screens, scroll the tables sideways; focus a table region to use keyboard scrolling.
| Concern | Evidence to request | Possible decision condition |
|---|---|---|
| An answer invents a commitment | Compare a sample draft with the approved source. | Require source checking before release. |
| Sensitive information reaches an unapproved recipient | Identify every input category and intended destination. | Exclude the material while handling remains unresolved. |
| A draft affects a person without review | Describe who relies on it and where a decision occurs. | Keep the proposal pending until responsibility and review are agreed. |
| The use expands after approval | Compare new data, audience and actions with the decision record. | Reopen the scope decision before expansion. |
This worksheet helps structure a conversation; it is not a legal framework or a finding that a use is safe. Technical detection and tool restrictions have their own boundaries: see AI agent guardrails. Keep those implementation details in the technical evidence attached to the use.
4. Work an allowed draft and a blocked request
The following entries illustrate how the same policy can produce different decisions. Every role, source label and decision is invented for this explanation; neither entry reports a customer outcome.
| Field | ANNOUNCEMENT-DRAFT | PERSONNEL-SUMMARY |
|---|---|---|
| Task and data | Draft an internal announcement using an approved public event description. | Summarise employee case notes using a provider whose data handling has not been approved. |
| Prohibited inputs | Private contact lists, employee records and unpublished commitments. | The case notes themselves; do not attach them to the request. |
| Accountable person | Communications lead. | Personnel operations lead. |
| Evidence and reviewer | Source reference and manually compared sample; communications editor reviews each draft. | Handling evidence absent; personnel reviewer records the gap without copying the notes. |
| Approval state | Allowed within scope: drafting only, with review before distribution. | Blocked: sensitive inputs cannot be submitted under this proposal. |
| Decision and exception | Communications approver records the permitted source and review condition; no exception. | Personnel approver records missing handling approval; no exception granted. |
| Review trigger | New source category, provider change or request to distribute automatically. | A revised proposal with an approved data boundary and handling evidence. |
For the announcement, the reviewer checks the draft against the event description and removes unsupported commitments before release. If someone later adds a private attendee list, the old permission no longer covers the request. For the personnel summary, the next step is to resolve the proposal, not to upload notes to discover whether the tool accepts them.
5. Adapt this plain-language AI policy template
Purpose: [why these rules apply to this team] Scope: [people, uses and tools covered] Permitted use: Work only within an allowed register entry. Data rule: Use only the sources and categories approved in that entry. Excluded material: Do not submit [specific prohibited inputs]. Output rule: Treat generated material as a draft until [review role] checks it. Action boundary: Obtain separate approval before [distribution or other action]. Uncertainty rule: Keep unsupported claims and missing evidence visible. Exception route: Send a scope-change request to [approver] before proceeding. Incident route: Pause the affected use and notify [contact] with its reference. Review rule: Revisit the entry when [triggers] occur. Record rule: Store decisions at [controlled location]; exclude sensitive payloads.
Replace each blank with an operational instruction. Specify who receives an exception request, what evidence they need, the permitted scope if approved and when that permission ends. A pending exception leaves the original restriction in place. Urgency alone should not silently change the recorded data boundary.
6. How do I review an AI governance policy through failure cases?
- Missing accountable person: leave the use pending until responsibility is accepted.
- New confidential attachment: stop submission and compare it with the permitted-data field.
- Unsupported draft statement: hold release, record the issue and request source correction.
- Expired exception: pause the exceptional scope until a fresh decision is recorded.
- Changed provider or audience: reopen the entry rather than relying on its previous approval.
The output of this exercise is a completed register and policy with visible unresolved questions. For task-level evidence, use AI agent evaluation; a sample result supports only the case inspected. Read the editorial policy for how Ai Flowi separates documented statements from observed tests.
Do not use this starter worksheet as certification, a legal compliance assessment or permission for unattended consequential decisions. It records organisational choices; actual restrictions, incident handling and specialist review must exist wherever the work takes place.