AI risk appetite and indicators
Say how much AI risk your organisation accepts, measure it each month with key risk indicators, and see at a glance what is green, amber or red, with an owner and a fix for anything outside appetite.
The AI risk scorecard is on Business Starter, Business and Enterprise. On Business Starter it is read-only and shows the indicators Deliverd measures from approvals and the audit trail. With Compliance Mode, on Business and Enterprise, it adds the indicators that read the AI register, figures from owners, your own appetite and limits, and corrective actions.
The appetite
The appetite says what your organisation accepts: a statement, what it has no appetite for, and what it accepts in a limited, controlled way. It also turns indicators on or off and sets each one's limits, owner and your own reference. Lower is better for every indicator: up to the green limit is green, up to the amber limit amber, and above it red. Leave amber empty and the first occurrence is red.
- Every save is a new version. Nothing in force changes until the version is approved.
- Approve a version with a Deliverd approval sent to the people you name, or record that it was approved elsewhere, at a meeting, with the day and where the decision is minuted.
- Until a version is approved, figures are judged against the limits Deliverd suggests.
- A baseline date lets you record and judge figures without opening actions or alerting anyone while you calibrate the limits.
The indicators
Measured indicators are read by Deliverd from what it holds. Submitted indicators come from outside it, such as access reviews, output sampling or downtime, and their owner gives the figure for each month by the 10th of the next, with where it comes from. A figure that has not arrived by then reads not measurable, which is itself a gap to close.
| Indicator | Area | Kind | Suggested limits |
|---|---|---|---|
| AI in use that is not in the register | Governance | Measured by Deliverd | Green up to 0 · Amber up to 1 · Red above 1 |
| Live AI without a risk assessment | Governance | Measured by Deliverd | Green up to 0% · Amber up to 5% · Red above 5% |
| High-impact AI without an accountable owner | Governance | Measured by Deliverd | Green 0 · Red above 0 |
| Confidential data disclosed through AI | Data and security | Measured by Deliverd | Green 0 · Red above 0 |
| Sensitive-data policy violations | Data and security | Measured by Deliverd | Green up to 0 · Amber up to 2 · Red above 2 |
| AI access that failed a least-privilege review | Data and security | Submitted by its owner | Green up to 0 · Amber up to 1 · Red above 1 |
| Actions taken without the required approval | Autonomy | Measured by Deliverd | Green 0 · Red above 0 |
| Decision records that failed verification | Autonomy | Measured by Deliverd | Green 0 · Red above 0 |
| Materially incorrect outputs in sampled review | Reliability | Submitted by its owner | Green up to 1% · Amber up to 3% · Red above 3% |
| Material AI incidents | Reliability | Measured by Deliverd | Green up to 0 · Amber up to 1 · Red above 1 |
| Material regulatory control failures | Compliance | Submitted by its owner | Green 0 · Red above 0 |
| AI suppliers in use without due diligence | Third parties | Measured by Deliverd | Green up to 0 · Amber up to 1 · Red above 1 |
| High-impact AI without a tested fallback | Resilience | Measured by Deliverd | Green up to 0% · Amber up to 10% · Red above 10% |
| Unplanned downtime of critical AI services | Resilience | Submitted by its owner | Green up to 30 min · Amber up to 120 min · Red above 120 min |
| People using AI without current training | People | Submitted by its owner | Green up to 2% · Amber up to 5% · Red above 5% |
| High-risk approvals given without opening the request | Autonomy | Measured by Deliverd, zero tolerance | Green 0 · Red above 0 |
| Transactions outside delegated authority | Autonomy | Submitted by its owner, zero tolerance | Green 0 · Red above 0 |
| Unresolved fairness or bias findings | Reliability | Measured by Deliverd, zero tolerance | Green 0 · Red above 0 |
| Agents with unapproved privileged access | Data and security | Submitted by its owner, zero tolerance | Green 0 · Red above 0 |
| Successful prompt-injection attacks | Data and security | Measured by Deliverd, zero tolerance | Green 0 · Red above 0 |
| Model changes not yet revalidated | Reliability | Measured by Deliverd, zero tolerance | Green 0 · Red above 0 |
| High-impact AI without a tested stop | Resilience | Measured by Deliverd, zero tolerance | Green 0 · Red above 0 |
Indicators that read the AI register count what it says: an incident needs its kind set to be counted, and a high-impact use needs its stop and manual fallback tests recorded on its register record. A past month always reads as it was judged, with the limits in force then. A correction keeps the earlier figure on record, superseded, with the reason.
Months
A month is judged once it has ended. The scorecard opens on the latest whole month; the cards above it show each recent month's red, amber and green at a glance, and the arrows step back and forward. The month in progress is shown but not judged: its measured indicators appear in the "This month so far" column until the 1st, when Deliverd records them.
When a figure is outside appetite
- Amber or not measurable opens a corrective action for the indicator's owner, who says what will be done and by when.
- Red opens the action and tells the owner and the people who run the programme, in the app and by the
risk_indicator.breachedwebhook, which can post to Slack or Teams. - Red on a zero-tolerance indicator tells them at once.
- Somebody other than the owner closes the action, saying what was done and how they checked it. An action can be tracked as an objective.
Owners
An indicator's owner need not be an administrator. They find the figures they owe and the actions they hold under My risk figures, and are reminded on the 1st and the 7th of the month.
The committee pack
An owner, administrator or compliance admin exports a month as a paper in your own branding, or as a signed file. The paper carries the signed file's digest, and deliverd verify (CLI 0.17.0 and later) checks the file is as it left. The scorecard is also on the API (GET /api/v1/risk-indicators) and the MCP server (get_risk_indicators, submit_risk_indicator_reading).