Practical guide
How to report AI capability evidence to an enterprise risk leader
Report AI capability evidence to an enterprise risk leader as decision-ready risk information, not a leaderboard or training update. The report should connect a bounded business objective to observed evidence, risk exposure, control ownership, assurance, and an explicit decision. Enterprise risk can integrate and challenge that view; it should not absorb the accountability of workflow owners or.
Report AI capability evidence to an enterprise risk leader as decision-ready risk information, not a leaderboard or training update. The report should connect a bounded business objective to observed evidence, risk exposure, control ownership, assurance, and an explicit decision. Enterprise risk can integrate and challenge that view; it should not absorb the accountability of workflow owners or control functions.
Role and evidence boundary
HM Treasury's Orange Book says risk reporting should be timely, accurate, useful for decisions, adapted to stakeholder needs, clear about assumptions, connected to objectives and appetite, and supported by assurance. The NIST AI Risk Management Framework calls for documented roles, monitoring, measurement, uncertainty, feedback, and risk treatment across the AI lifecycle. These sources define useful reporting disciplines. They do not show that an enterprise risk leader owns the budget for an assessment programme, that an assessment score predicts workplace outcomes, or that any control is effective. HM Treasury: Orange Book NIST: AI RMF Core NIST: Measure playbook
Sources: [1] [2] [3]
Classify the stakeholder before building the report
An enterprise risk leader is usually an influencer and reviewer of AI capability evidence and may be an owner of the enterprise risk method. The workflow executive remains accountable for operating results. Privacy, security, legal, model-risk, internal-audit, learning, and people teams retain their own mandates. A board or executive committee may own the consequential decision.
Confirm the actual mandate, delegated authorities, committee route, risk taxonomy, appetite statements, reporting calendar, and assurance model. Do not infer decision rights from a title.
Use the BRIDGE report
### B — Business decision and boundary
Name the objective, decision, owner, deadline, alternatives, affected workflow, population, AI use, assessment use, and consequence of uncertainty. State whether the request is to approve a pilot, continue a restriction, fund a control, commission assurance, or accept residual risk.
Define what the evidence can support. Assessment data may describe performance in a controlled task. Training completion describes participation. Tool telemetry describes recorded use. Incident data describes known events. None alone establishes real-world capability or business impact.
### R — Risk pathway and appetite
Express each principal risk as cause, event, and consequence linked to an objective. Map existing controls, accountable owners, indicators, thresholds, and appetite or tolerance references. Include opportunity risk: excessive restriction can also prevent useful learning or service improvement.
Show inherent exposure, control evidence, residual uncertainty, and trend separately. Do not convert a capability score into a risk rating without an approved mapping and evidence that the relationship is meaningful.
### I — Information and evidence quality
For each claim, report the population, period, source, definition, coverage, missingness, method, reviewer, and version. Distinguish observed fact, analysis, inference, forecast, and recommendation. Reconcile assessment results with incidents, quality review, complaints, operational telemetry, and control tests where permitted.
Flag selection effects, non-completion, technical failures, scenario limitations, reviewer disagreement, model or rubric changes, and small subgroups. Use “directional” or “insufficient evidence” when the record cannot support a precise comparison.
### D — Decisions, controls, and assurance
Show the credible choices: proceed, proceed with conditions, repair, restrict, defer, or stop. For each, give the expected contribution, risk change, cost, reversibility, dependencies, affected parties, and evidence still needed.
Name every control owner and test method. Separate management monitoring from second-line challenge and independent assurance. A policy, training rate, or dashboard is not evidence that a control operated across the relevant population.
### G — Gaps, adverse evidence, and escalation
Place material failures on the first page: unsupported outputs, sensitive-data events, accessibility failures, complaints, assessment evidence gaps, overrides, untested routes, and control exceptions. Give containment, owner, due date, escalation threshold, and residual uncertainty.
Do not hide missing evidence in an appendix or average a severe failure into an overall score. Report disagreements and any request to use results beyond their approved purpose.
### E — Exit, expiry, and next review
End with the requested decision, decision authority, conditions, monitoring, stop trigger, evidence owner, next review date, and expiration. Evidence should expire after material changes to the workflow, AI system, data, assessment, rubric, population, control design, or risk appetite.
One-page evidence map
| Decision question | Evidence available | Limitation | Owner | Next evidence | | --- | --- | --- | --- | --- | | is the task interpretable? | completion, reviewer notes, missing evidence | controlled scenario only | assessment owner | adjudicated case review | | are critical practices observed? | source checks, corrections, escalations | opportunity-dependent | workflow owner | live quality sample | | are controls operating? | logs, exceptions, reperformance | period and population bounded | control owner | independent retest | | is expansion within appetite? | risk pathway, indicators, assurance | appetite decision remains human | accountable executive | committee decision record |
Worked report: synthetic customer-communications workflow
A communications executive must decide whether to extend an AI drafting workflow from one product team to four regional teams. The enterprise risk leader reviews the integrated evidence. The proposal includes assessment results from 28 participants, training records, six weeks of workflow telemetry, editorial corrections, privacy-control tests, and incident logs.
The report states the decision and refuses a broad “AI-ready workforce” claim. The controlled assessment tested whether participants opened authoritative sources, rejected a superseded brief, recalculated a changed denominator, marked uncertainty, and escalated an unsupported public claim. Twenty-four submissions are usable; two sessions failed technically and two lack a complete evidence trail. Reviewers agree on 88% of coded critical actions before adjudication. These are task-bounded observations, not a validated workforce benchmark.
Operational evidence shows 412 assisted drafts, 71 material editorial corrections, nine unsupported quantitative claims caught before publication, and one sensitive note entered into an unapproved field. The privacy owner contains the event and disables that route. The source-link control operated for the standard path but was not tested for two regional content systems. No external harm is recorded, but complaint monitoring is too incomplete to support an absence-of-harm claim.
The risk leader maps the evidence to three principal risks: unsupported public claims, inappropriate data entry, and over-reliance by reviewers. The executive retains decision ownership. Communications owns claim controls; privacy owns the data route; learning owns support; enterprise risk challenges the residual view; internal audit is not asked to endorse the pilot.
The options are full expansion, two-region conditional expansion, repair and retest, or stop. The decision is repair and retest: keep the current team, require source links and reason-coded overrides, close the sensitive-data route, test the two regional systems, and sample live corrections for six weeks. Expansion returns to committee only when the named evidence is complete. The report does not claim compliance, capability transfer, or financial return.
Product boundary
Two Prune supplies realistic AI-enabled work assessments and evidence about observed participant performance. It can contribute one controlled evidence layer to enterprise risk reporting. It does not set risk appetite, certify controls, audit an organization, prove workplace performance or business value, recommend employment decisions, or replace accountable human judgment.
Sources
HM Treasury: The Orange Book — Management of Risk
- NIST: AI Risk Management Framework Core
- NIST: AI RMF Measure playbook
Sources: [1] [2] [3]
Sources
- 1.HM Treasury: Orange Book · UK Government
- 2.NIST: AI RMF Core · National Institute of Standards and Technology
- 3.NIST: Measure playbook · National Institute of Standards and Technology