How to assess uncertainty management in AI-enabled work

Treat uncertainty as a decision state

AI-assisted work often turns incomplete inputs into fluent output. The problem is not that every incomplete statement is unusable. The problem is that readers may not be able to tell which statements are well supported, which depend on assumptions, and which should block action.

NIST's AI Risk Management Framework asks organizations to document knowledge limits, human oversight, likely impacts, measurement uncertainty, and the limits of generalizing beyond tested conditions. Its Generative AI Profile adds actions for risks that can be novel to or amplified by generative systems. Together, these sources support a practical rule: uncertainty should be represented in the work product and connected to the consequence of being wrong.

This skill is different from adding a generic disclaimer. A footer saying "AI can make mistakes" does not help a programme owner decide whether to send a briefing, approve a supplier, or request another source. Useful uncertainty management identifies the exact uncertain item, its basis, its decision effect, and its route to resolution.

Use the RANGE method

Apply RANGE to every material statement or recommendation whose support is incomplete:

The output is an uncertainty register linked to the deliverable. It is not a universal confidence percentage. Numerical confidence can imply precision that the evidence does not warrant. A categorical evidence state plus a concrete resolution route is often more useful for bounded knowledge work.

• Record the proposition precisely. Avoid vague labels such as "some risk" or "probably fine."
• Assign an evidence state: observed, derived, assumed, estimated, disputed, or unknown.
• Name the consequence if the proposition is wrong, including which decision or audience would be affected.
• Give a resolution route: obtain a named record, ask an accountable owner, rerun a calculation, test a condition, narrow the claim, or escalate.
• Expose the release decision: use, use with a qualifier, hold, or exclude.

Worked artifact: an uncertainty register for a launch note

Consider an original synthetic exercise. A participant receives a product brief, a support log, a draft launch date, and an AI-generated executive note. The note recommends launching a new service on 15 November and says migration demand will be modest.

The participant creates this RANGE register:

| Proposition | Evidence state | Consequence if wrong | Resolution route | Release decision | | --- | --- | --- | --- | --- | | The launch date is 15 November | Disputed | Customer communications could promise an unavailable date | Ask the programme owner to reconcile the dated plan with the newer engineering message | Hold the date | | Fewer than 20 accounts need migration | Estimated | Support capacity could be understated | Reproduce the CRM filter and record the extraction timestamp | Use only after recalculation | | Existing contracts permit the change | Unknown | The rollout may require customer-specific review | Obtain the current contract position from the accountable commercial owner | Exclude from the note | | A two-week support window is sufficient | Assumed | Issues could outlast staffing cover | Label the assumption and define a review trigger after the first five migrations | Use with qualifier | | The revised onboarding text is ready | Observed | Low, if the cited approved file is current | Check the version and approval field in the document library | Use |

The participant then rewrites the executive note. The date becomes a decision request rather than a promise. The migration estimate includes its extraction time. The contract statement is removed. The support recommendation includes the assumption and review trigger.

All organizations, dates, counts, records, and decisions in this example are synthetic. They are not customer evidence, product outcomes, or benchmarks.

Design an assessment that reveals judgment

Give the participant a source pack in which uncertainty has more than one cause:

Ask for the deliverable and the RANGE register. Introduce one new record midway through the task. A strong participant should update the evidence state and disposition, not merely append the new source.

Observe whether the participant distinguishes incomplete evidence from contradictory evidence, connects uncertainty to decision consequence, and chooses a proportionate response. Marking everything "unknown" is not strong performance. Neither is assigning confident percentages without a defensible basis.

• two records disagree because one is newer;
• a calculation is reproducible but based on incomplete rows;
• an accountable owner has not answered a material question;
• a recommendation depends on a stated risk tolerance;
• a surprising fact is directly supported and should not be weakened; and
• one low-impact unknown can be carried without blocking the deliverable.

Keep adjacent controls separate

Uncertainty management decides how incomplete support affects action. Fact checking tests whether a specific claim matches an authoritative source. Assumption testing challenges a proposition the analysis depends on. Evidence sufficiency asks whether the source set is adequate for the decision. Escalation routes an unresolved issue to an accountable person.

These controls interact, but their artifacts differ. A fact check may prove a date wrong. The uncertainty register shows whether the unresolved replacement date blocks publication. An assumption test may show that a staffing model is sensitive to volume. The register connects that sensitivity to a review trigger and release disposition.

Score the observable work

Review five dimensions:

Before release, compare the register with the final deliverable. A held statement must not reappear as a fact in a heading or recommendation. A resolved item should carry its updated evidence state and source version. The register should also record when the assessment occurred because evidence can change.

Uncertainty management cannot prove that all relevant evidence was available or that the resulting decision is correct. It makes the limits and decision logic inspectable. NIST describes AI risk management as continuous and contextual, not as a checklist that ensures trustworthiness.

In a Two Prune assessment, the RANGE register, source navigation, updates, and final dispositions can provide task-bounded evidence about Verification & Risk Control, Evidence Navigation, Judgment & Synthesis, and Communication & Delivery. They do not certify professional competence, predict workplace outcomes, or make an employment decision.

• classification: material items have defensible evidence states;
• consequence: the effect of being wrong is specific to the task;
• resolution: the next step names an evidence object, test, or owner;
• proportionality: the disposition fits the consequence and available evidence; and
• propagation: qualifiers and holds appear consistently in summaries, tables, and recommendations.

Sources

NIST: AI Risk Management Framework Core

• NIST: Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profile
• NIST: Framing Risk in the AI RMF

Sources: [1] [2] [3]