Find unsupported assumptions in AI-assisted work before they become decisions

Start with the decision, then trace its dependencies

Imagine a synthetic operations exercise. A team must recommend whether to expand a customer-support pilot to a second region. An assistant drafts: “Expand now because the pilot cut handling time by 30% and the new region has the same case mix.” The supplied pack contains weekly ticket counts and a few reviewed transcripts, but no handling-time baseline and no regional case-mix comparison. The attractive recommendation rests on two different unsupported premises, not just one missing citation.

Mark each material conclusion in the proposed deliverable and work backward: what measured fact, source interpretation or policy permission would have to be true for it to hold? Separate source-backed observations from inferences and from constraints supplied by the assignment owner. A document saying “pilot users liked the workflow” does not establish that time fell; an average from one queue does not describe another queue without checking its composition. The worker should review the source itself, not ask the same assistant to certify its own unsupported statement.

NIST's AI Risk Management Framework Map guidance calls for documenting assumptions, limitations, knowledge limits and the context in which AI output will be used. The UK Government's AI Playbook says AI outputs can be inaccurate and advises a process to test them in its public-sector context. These are organizational frameworks; a particular worker's fact-check can still miss an error. The dependency trace below is an illustrative work method adapted to a bounded decision.

Sources: [1] [2]

Make an assumption-dependency record

For the synthetic expansion brief, write a compact record with one row per decisive premise. Premise A: “handling time fell by 30%.” Decision affected: expand on efficiency grounds. Claimed evidence: assistant's draft only. Required evidence: comparable before-and-after timing definitions, observation windows, denominators and a review of case mix. Current status: unverified; remove the number. Premise B: “the new region has the same case mix.” Decision affected: transfer pilot results. Required evidence: categorized cases from both regions using the same taxonomy and period. Current status: unknown; treat transferability as an open question. Premise C: “an expansion may use customer transcripts.” Decision affected: which pilot materials can be shared. Required evidence: the organization's actual approved data-handling rules and owner decision. Current status: not established by the supplied pack; do not move any transcript.

The record should also name the source location, who can resolve the question, the deadline and the fallback wording. A decision-critical unknown can justify a conditional recommendation: “Run a limited comparison after collecting baseline timing and regional case-mix evidence.” A less important assumption may be disclosed as a limitation without stopping the entire brief. This is a proportionality judgment by the accountable human, not an automatic risk score or a finding that all unknowns are equally consequential.

Verify the revised deliverable against the record

Read the final recommendation sentence by sentence. Every quantitative statement should point to the measurement that supports that exact quantity; every transferred conclusion should identify the population and context it came from; every permission claim should point to the authorized owner or policy. In the worked example, the editor removes the 30% figure and the unverified regional-equivalence claim, keeps the observed ticket counts with their period, and attaches the two unanswered evidence requests to the decision memo. A second reviewer can reproduce the trace without trusting the assistant's confidence or an inaccessible prompt transcript.

Test the method with a contrast case as well: a source that does contain a directly measured value should not be downgraded merely because an AI assistant mentioned it. Check its definition and relevance, then use it within its actual scope. The principal failure here is laundering plausible premises into facts by polishing the prose or attaching a general reference. A real source can support a broad principle while leaving the local numerical claim unsupported. NIST's Measure guidance emphasizes context-appropriate methods and documenting risks that cannot be measured; this guide's assumption record remains our own illustrative artifact, not a required NIST checklist.

The method applies to ordinary knowledge-work outputs: proposals, planning briefs, analysis notes and training exercises. It does not validate the underlying AI system, certify compliance, or predict operational outcomes. A Two Prune assessment can observe how a participant treats assumptions in a synthetic task and present that bounded evidence for human interpretation; it does not review a customer's live regional expansion or make an employment decision.

Sources: [3]