Practical guide
How an enterprise risk leader can build the case for applied AI capability evidence
An enterprise risk leader can build a credible case for capability evidence by showing which material decisions lack visibility, what current evidence misses, which options exist, and how a bounded investment would reduce decision uncertainty. The case should not assume that assessment is automatically valuable. It must connect a specific evidence gap to objectives, risk exposure, accountable.
An enterprise risk leader can build a credible case for capability evidence by showing which material decisions lack visibility, what current evidence misses, which options exist, and how a bounded investment would reduce decision uncertainty. The case should not assume that assessment is automatically valuable. It must connect a specific evidence gap to objectives, risk exposure, accountable owners, alternatives, cost, and a reversible next step.
Mandate and claims boundary
HM Treasury’s Orange Book describes risk management as integral to governance, options appraisal, monitoring, reporting, and organizational objectives. The NIST AI Risk Management Framework says risk tolerance is contextual and use-case specific, and its Core includes defining business value, considering viable non-AI alternatives, documenting residual risk, and assigning responsibilities. These sources support business-case reasoning. They do not assign budget authority to every risk leader or establish the benefit of a particular assessment product. HM Treasury: The Orange Book NIST: AI Risk Management Framework Core
Sources: [1] [2]
Build the case with PROVE
### P — Problem and decision consequence
Define the decision that current evidence cannot support. Examples include whether to expand an AI-enabled workflow, prioritize learning, strengthen review controls, or investigate recurring verification failures.
Describe the consequence of uncertainty, not just the absence of a metric. A risk register may list “insufficient AI capability,” but the business case needs the affected workflow, plausible event, decision owner, population, and timing. Separate a capability hypothesis from an observed incident or control defect.
### R — Roles, rights, and existing evidence
Classify the enterprise risk leader as owner, influencer, reviewer, or collaborator for each action. Service owners own workflow outcomes; learning may own support; technology and control functions own their responsibilities; delegated executives accept residual risk.
Inventory existing evidence before proposing new collection: access logs, training records, work quality samples, incidents, assurance results, manager observations, tool telemetry, user research, and assessment evidence. State what each source measures, its coverage, and its limitations. Do not assume that activity data reveals judgment, source checking, or safe escalation.
### O — Options and opportunity cost
Compare at least four credible choices:
For each, show what decision it informs, cost, time, burden, risks, evidence quality, affected parties, and reversibility. Include the cost of diverting people from delivery and the risk of collecting data that will not change a decision.
### V — Value hypothesis and verification plan
Write the value claim as a testable hypothesis. For example: “A task-specific simulation may reveal recurring evidence-navigation and verification weaknesses that current completion and usage data cannot show, enabling owners to target workflow support.”
Define the baseline, eligible population, sample, comparison, missingness rules, measures, review method, decision thresholds, and stop conditions before launch. Separate the value of better evidence from any later effect of training, process repair, or technology change.
### E — Economics, evidence gate, and exit
Show direct and indirect costs: design, source materials, participant time, review, administration, accessibility, data handling, remediation, and governance. Use ranges where inputs are uncertain. Name who validates the financial model.
Set the next gate. A pilot may justify continuation only if it reaches the intended population, produces interpretable evidence, meets control and accessibility requirements, and changes a real decision. Define what would lead to repair, pause, consolidation with another method, or exit.
- continue with current evidence;
- improve existing operational measurement;
- conduct a bounded work-sample assessment;
- combine methods in a staged pilot;
- delay, restrict, or stop the relevant workflow.
Business-case evidence table
| Claim | Evidence now | Gap | Proposed test | Decision if supported | Decision if not supported | | --- | --- | --- | --- | --- | --- | | verification failures are material | incident and quality records | causes and population unclear | representative work simulation plus record review | target workflow controls and support | improve incident classification first | | current measurement is incomplete | usage and completion data | observed practice absent | compare sources on the same bounded group | retain complementary measures | stop duplicate collection | | new evidence can change action | named owner and decision | threshold not agreed | preregister decision rule | proceed to bounded pilot | do not collect |
Worked business case: synthetic procurement review workflow
An organization plans to expand AI assistance for supplier due-diligence briefs. The enterprise risk leader sees incidents involving missing sanctions sources and unsupported ownership claims. Existing evidence covers tool access, training completion, and nine quality incidents, but cannot show whether failures arise from source navigation, AI reliance, unclear procedure, or review design.
The PROVE case names the decision: the procurement service owner must choose in eight weeks whether to expand, restrict, or redesign the workflow. Enterprise risk coordinates the integrated evidence view and challenges assumptions; procurement owns the workflow; compliance owns sanctions requirements; privacy and security own their controls; learning owns support design.
Five options are compared. Continuing current evidence is low cost but leaves the practice gap. Adding more tool telemetry shows activity, not judgment. Manager observation is available but inconsistent and may miss private checking. A bounded simulation can expose source use and verification in a controlled task but may not transfer directly to live work. A combined pilot uses the simulation to form hypotheses, then tests selected controls in live quality review.
The case recommends the combined pilot for 24 volunteers across three procurement teams. Costs include case design, participant time, independent review, accessibility checks, and analysis. The sample is not presented as a workforce benchmark. Non-participation is recorded separately, and no individual employment decision uses the result.
The verification plan preregisters three questions: whether participants locate the required authoritative sources, whether they challenge unsupported AI statements, and whether they escalate conflicts. The live review then checks whether the same failure classes occur in a representative case sample. The pilot stops if restricted data enters an unapproved tool or evidence cannot be separated from identifiable operational records.
The next decision gate requires usable evidence coverage, reviewer agreement above the organization’s preselected threshold, no unresolved high-severity control incident, and at least one actionable workflow or support change. If the simulation adds no decision value beyond existing quality review, the organization does not scale it.
The artifact contains the decision statement, role map, current-evidence inventory, option appraisal, cost range, hypothesis ledger, pilot protocol, missingness rules, control review, decision thresholds, and exit plan.
Objections to test
Two Prune provides realistic AI-enabled work assessments and evidence about observed participant performance. It may contribute one evidence layer to a bounded business case. It does not determine risk appetite, certify controls, establish financial return, or make employment decisions.
- “We already have adoption dashboards.” Test whether they observe the source use, checking, and escalation decisions that matter here.
- “Managers can judge this informally.” Define which behaviors are observable, how observations are recorded, and what remains unseen.
- “An assessment proves workplace impact.” Keep observed performance, transfer, operational change, and business outcomes as separate claims.
- “More evidence is always safer.” Consider collection burden, privacy, interpretation risk, and whether a decision will change.
- “Enterprise risk should own the programme.” Preserve the authority of workflow, control, learning, finance, and executive owners.
Sources
HM Treasury: The Orange Book
- NIST: AI Risk Management Framework Core
- NIST: Framing AI risk and risk tolerance
Sources: [1] [2] [3]
Sources
- 1.HM Treasury: The Orange Book · UK Government
- 2.NIST: AI Risk Management Framework Core · National Institute of Standards and Technology
- 3.NIST: Framing AI risk and risk tolerance · National Institute of Standards and Technology