Practical guide

How to Report AI Capability Evidence to an AI Governance Leader

Report AI capability evidence to an AI governance leader as a decision pack, not a league table. The pack should show what was observed, what evidence is missing, which governance or programme decision it can inform, and what remains outside scope. It should also identify whether the leader is the owner, reviewer, influencer, or collaborator for each requested action.

By Two Prune

Report AI capability evidence to an AI governance leader as a decision pack, not a league table. The pack should show what was observed, what evidence is missing, which governance or programme decision it can inform, and what remains outside scope. It should also identify whether the leader is the owner, reviewer, influencer, or collaborator for each requested action.

Start with the decision route

Write the reporting purpose in one sentence: “This pack supports [owner] deciding [action] by [date] for [population and workflow].” Name the governance leader's role in that decision and list the other owners for learning, operations, privacy, security, domain quality, procurement, and budget.

NIST's AI RMF Core calls for documented roles, communication lines, monitoring, training, and accountability throughout AI risk management. Use that structure to route findings and actions; do not assume the governance function owns every programme or operational decision. NIST AI RMF Core

Open with a compact decision header:

If no action is supportable, say so plainly and name the evidence needed next.

  • decision requested and accountable owner;
  • evidence period, cohort, task, and system version;
  • recommended action and alternatives;
  • material limitations and missing evidence;
  • next review trigger.

Sources: [1]

Separate the evidence layers

Use distinct sections for participation, access, use, task-bounded assessment, workplace observation, control operation, and business outcomes. Each section should state population, denominator, source, time window, missingness, and the decisions it can and cannot support.

For example, course completion can show participation in a defined learning event. Tool activity can show recorded use under a particular configuration. A realistic simulation can show observed behaviour in that case. None alone establishes organization-wide capability, workplace transfer, control effectiveness, or business impact.

The U.S. Government Accountability Office organizes AI accountability around governance, data, performance, and monitoring. Its framework also emphasizes documented methods, metrics, outcomes, and corrective actions. Use those four lenses to check whether the report covers oversight needs without presenting them as a validation of Two Prune. GAO AI Accountability Framework

Sources: [2]

Report distributions, evidence strength, and missingness

Avoid a single average that hides different tasks or evidence quality. Show counts and distributions for the bounded cohort, pillar-level or behaviour-level observations, common failure modes, reviewer agreement where available, and the share of records with complete evidence.

Attach an evidence-strength label to each finding:

Missing evidence is not poor performance. A participant who never encountered an ambiguity did not demonstrate the escalation behaviour; a broken event stream does not show that the behaviour failed to occur.

  • Direct: observed in the retained task artifact or event evidence.
  • Corroborated: supported by more than one independent evidence source.
  • Limited: observed, but the opportunity, sample, reviewer consistency, or instrumentation constrains interpretation.
  • Unavailable: the required evidence was not captured or the behaviour was not elicited.

Convert findings into governed actions

For every material finding, provide the affected decision, proposed action, action owner, evidence threshold, due date, and closure artifact. Keep participant development actions separate from system, policy, source, or control repairs.

Use a decision-and-action table:

| Finding | Evidence boundary | Decision | Owner | Closure evidence | | --- | --- | --- | --- | --- | | source conflicts often missed in the synthetic case | 18 advisers, one task | revise verification practice | capability lead | approved exercise and rerun record | | escalation recognized but route unclear | task artifact plus reviewer notes | repair operating guidance | service owner | published route and user check | | event stream incomplete for four records | no workflow conclusion for those records | repair instrumentation | product owner | tested capture and missingness report |

Do not route every problem to training. Access, policy, interface, source quality, workload, and review design may be the operative cause.

Include objections and reporting boundaries

Answer the likely objections directly: the cohort may not represent other teams; a simulation differs from live work; reviewers may disagree; some actions happened outside recorded tools; and a result may change with task, system, assistance, or source conditions. Report these as decision constraints, not footer boilerplate.

State whether the report includes personal data, who can access participant-level evidence, retention, and whether aggregated reporting suppresses very small groups. A governance summary should normally lead with programme and system actions; participant-level access should follow the authorized purpose and applicable policy.

Worked quarterly pack: policy-assistant capability

A governance leader receives a quarterly pack for an internal policy-assistant programme. The leader reviews programme risk and reports to a steering group; the service owner controls deployment, the capability lead owns learning, and the privacy lead owns retention review.

The pack covers 24 invited advisers, 22 completions, and 20 records with complete event evidence. It reports that 15 of those 20 selected the current policy source, 12 caught an unsupported exception, and nine used the required escalation route when ambiguity appeared. It does not turn these counts into an organization-wide readiness score. Two incomplete event streams are marked unavailable for process findings, while their submitted artifacts remain reviewable for limited output evidence.

The strongest finding is a repeated source-version error across both task artifacts and reviewer notes. The pack recommends replacing the source-selection exercise and correcting a stale link before another cohort. The escalation result is limited because only 14 cases actually presented the trigger. The steering request is therefore to approve the source repair, retain the pilot scope, and review a rerun—not to expand the system based on a positive average.

The appendix contains the task specification, rubric version, evidence dictionary, cohort flow, missingness table, distributions, reviewer notes summary, source register, calculation method, and action log. Participant identities are excluded from the steering version.

Apply a release check

Before circulation, verify that every number has a denominator and reproducible source, all claims match the task and population, missingness is visible, comparative labels use an appropriate benchmark, actions have owners, participant access is authorized, and the report date and system version are current. Require a governance reviewer and the relevant operational owner to sign off on their own sections.

Two Prune can provide structured, confidence-aware evidence from a controlled assessment. It does not establish legal compliance, control effectiveness, production quality, organization-wide readiness, business value, or an employment recommendation. The governance leader should combine the bounded assessment record with current operational and jurisdiction-specific evidence.

Sources

  1. 1.NIST AI RMF Core · National Institute of Standards and Technology
  2. 2.GAO AI Accountability Framework · U.S. Government Accountability Office