How to develop and assess assumption logging in AI-enabled work

Treat assumptions as decision inputs

The NIST AI Risk Management Framework describes mapping context, intended purpose, users, impacts, and limitations before measurement and management. The UK government's Orange Book treats risk as the effect of uncertainty on objectives and emphasizes transparent, information-based decision-making, changing conditions, confidence, dependencies, and review. These principles support a practical habit: assumptions should be attached to the decisions they influence and revisited when evidence or context changes.

An assumption log is not a list of every unknown. It focuses attention on uncertain propositions that can materially alter the chosen action, required control, cost, schedule, interpretation, or confidence.

Build a STAMP assumption record

Use STAMP to turn an implicit belief into a testable working record:

| Field | Question | Example entry | | --- | --- | --- | | Statement | What exactly are we treating as true for now? | “The source export contains one row per active account.” | | Trigger | Which decision, calculation, or workflow depends on it? | Denominator for the adoption rate | | Authority | What evidence or person currently supports it, and how current is that support? | Data dictionary version and owner response | | Materiality | What changes if the statement is wrong? | Rate may be overstated; segment comparison may reverse | | Plan | How and by when will we test, monitor, revise, or escalate it? | Reconcile account IDs before analysis; owner and date named |

Add a status such as open, supported, contradicted, superseded, or accepted for a bounded decision. “Accepted” should mean the team consciously proceeded within a stated limit, not that the proposition became permanently true.

AI can help surface hidden assumptions by challenging a plan, tracing dependencies, or proposing counterexamples. The worker should convert suggestions into specific statements and verify them against task evidence. A model's confidence, repetition, or fluent explanation is not authority for the record.

Worked example: a synthetic adoption analysis

Suppose a programme owner asks for a synthetic analysis of AI-tool adoption across three departments. The pack includes a licence export, an activity table, an organization chart, and a message saying that contractors “should be excluded.” This example is invented and does not report a Two Prune customer outcome.

An initial AI-assisted draft calculates active users divided by licensed users and concludes that one department has the strongest adoption. Before accepting the result, the participant creates four STAMP records:

The first two assumptions trigger reconciliation checks. Duplicate and unmatched identifiers are found, so the denominator and join logic are revised. The third is supported by the data dictionary. The fourth remains open because the message states a desired rule but does not show how contractors are tagged.

The final artifact presents the reconciled rates, labels the contractor exclusion as unresolved, and avoids ranking the affected departments. The log identifies the data owner, the field needed to resolve the issue, and the decision that can proceed meanwhile: compare within-department changes using the consistently identified employee records. The useful output is not merely a more cautious paragraph. It is a visible chain from uncertain statement to affected decision and next check.

• Each licence row represents a current person rather than a reassigned or service account.
• The activity table's user identifier maps one-to-one to the licence export.
• “Active” means at least one qualifying event in the same thirty-day period for every department.
• Contractors are absent from both numerator and denominator, not only from the organization chart.

Practice by changing one hidden premise

Give a learner a short task with one deliberately underspecified premise, such as the reporting period, target population, currency basis, approval authority, or meaning of success. Ask for an initial plan and assumption log. Then reveal evidence that contradicts one material assumption.

The learner should identify which calculations, claims, dependencies, and recommendations need revision. This makes assumption management observable. Merely adding “subject to assumptions” at the end of a report does not demonstrate the skill because it does not show what was assumed or what changes when the premise fails.

A strong practice cycle includes:

• extracting uncertain propositions from the brief, source pack, and AI output;
• separating assumptions from known facts, constraints, decisions, and questions;
• ranking assumptions by decision impact and ease of verification;
• assigning a check, owner, and time boundary;
• revising dependent work when evidence changes;
• preserving the old state and reason for change rather than silently overwriting it.

Assess the log and the response to evidence

Use a realistic simulation in which the participant must act before every uncertainty can be resolved. Score the behavior, not the number of rows. Relevant evidence includes the STAMP records, source review, questions asked, AI interactions available inside the assessment, calculations, revisions, and final deliverable.

Look for specificity, materiality, evidence routing, and update discipline. A good statement is falsifiable enough to check. The trigger names the affected choice. The authority field distinguishes evidence from hearsay or model output. Materiality describes a concrete consequence. The plan is proportionate and has a responsible party or decision point.

Also inspect prioritization. Logging twenty harmless unknowns while missing the denominator assumption that drives a recommendation is weak performance. Treating every assumption as a blocker is also weak when a reversible action can proceed within a stated boundary. The task is to manage uncertainty in service of a sound decision.

Verify closure without erasing uncertainty

For each material record, ask four questions before delivery: Was the assumption tested? Did the evidence actually address the statement? Were dependent calculations and claims updated? Is any remaining uncertainty visible beside the decision it limits?

Review contradictions and superseded entries rather than deleting them. A short change history lets another reviewer understand why the analysis moved. If the check cannot be completed, narrow the recommendation, choose a reversible action, request the missing evidence, or escalate according to the task's stakes.

Assumption logging is a context-framing behavior, not a complete measure of risk judgment. A participant can maintain a neat log and still choose poor assumptions or weak checks. In a Two Prune assessment, the log and response to new evidence may contribute bounded evidence about Context Framing, Verification & Risk Control, and Judgment & Synthesis. They do not prove general competence, future outcomes, or suitability for employment.

Sources

NIST AI Risk Management Framework

• HM Treasury Orange Book: Management of Risk principles and concepts

Sources: [1] [2]