Enterprise risk leaders and AI-enabled knowledge work

Clarify the risk leader's role

The enterprise risk leader may own the risk framework and aggregation method, influence risk appetite, challenge evidence, or coordinate escalation. They do not automatically own every AI budget, implementation, security control, legal interpretation, or business outcome. Record who owns the workflow, who owns each safeguard, who can accept residual exposure, and who must be consulted.

Use a precise unit of analysis. “Generative AI” is too broad for a useful decision. A research assistant using public sources, a customer-summary tool handling confidential records, and an agent that can change account data have different exposure paths and evidence requirements even if they use the same model provider.

Build an exposure-to-evidence register

For each workflow, record its intended decision or output, users, information classes, external services, permitted actions, affected parties, plausible failure modes, safeguards, owners, and current operating stage. Add links to test evidence, incident or exception records, monitoring results, and unresolved assumptions. State the next decision and the evidence required for it.

NIST says its AI Risk Management Framework is intended to help organizations manage risks to individuals, organizations, and society. Its Generative AI Profile is a cross-sector companion that identifies risks associated with generative AI and proposes actions aligned with organizational goals and priorities (NIST AI Risk Management Framework). These are voluntary risk-management resources, not a certification or a substitute for applicable requirements.

Sources: [1]

Aggregate without erasing context

Use common exposure dimensions so leaders can compare workflows, while retaining the evidence behind each rating. Useful dimensions may include consequence, reach, reversibility, information sensitivity, autonomy, human review, dependency concentration, and detectability. Define the scale and decision thresholds before scoring.

Do not turn unlike observations into a false portfolio average. A low-frequency, high-consequence action path should remain visible. So should a shared dependency that affects several otherwise low-risk workflows. Separate inherent exposure, implemented safeguards, observed control performance, and residual uncertainty. Unknown evidence is not the same as a poor result, and a pilot result is not evidence of performance at larger scale.

Worked artifact: a synthetic portfolio register

Imagine three synthetic workflows. A public-source market-research assistant drafts annotated briefs but cannot publish; a service-summary tool processes approved customer records and requires line-by-line review; an operations agent can prepare account changes but cannot execute them without a separate approver.

The register classifies the first workflow as reversible and low reach, with citation completeness as its main evidence gap. The second has higher information sensitivity and a concentrated dependency on access controls; a sampling test finds two summaries that omit material caveats, so expansion is held. The third has potentially consequential actions, but the enforced separation between preparation and execution limits autonomy. A control test finds one approval log that cannot be reconstructed, so the workflow remains restricted while logging is repaired.

The enterprise risk leader does not collapse the three into “amber.” They record distinct decisions: continue the first within scope, repair and retest the second, and restrict the third until evidence closes the logging gap.

Set escalation and re-review triggers

The NIST Generative AI Profile describes profiles as ways to apply Govern, Map, Measure, and Manage to a particular setting based on requirements, risk tolerance, and resources (NIST AI 600-1). Apply that contextual principle by defining triggers that reopen a decision: a new data class, broader user population, external publication, autonomous action, material model change, new vendor dependency, repeated exception, or evidence that a safeguard did not operate.

Set escalation rules around the decision, not just a numerical score. Specify who must act, what must stop or narrow, what evidence is needed, and when the decision expires. Preserve dissent and unresolved assumptions in the register so a clean executive summary does not hide material uncertainty.

Sources: [2]

Verify the portfolio view and preserve boundaries

Before presenting the register, reconcile it with source systems and control owners. Sample linked evidence, confirm that exceptions are current, test whether shared dependencies are counted, and verify that each next action has an accountable owner and due condition. Compare the executive view with the underlying workflow records to ensure that aggregation has not removed the reason for a hold or restriction.

This method can support consistent challenge and escalation. It does not prove compliance, predict incidents, establish control effectiveness, or demonstrate financial return. Two Prune can provide task-bounded assessment evidence about how a participant handles a realistic work simulation. It does not run enterprise risk management, certify production AI systems, or make employment decisions.