How to develop and assess evidence sufficiency in AI-enabled work

Start with the claim, not the size of the evidence pack

The US Government Accountability Office's 2024 Government Auditing Standards requires an overall assessment of the collective evidence supporting findings and conclusions. It directs auditors to consider significance, corroboration, and audit risk, and not to use evidence that is insufficient or inappropriate as support. The standards also call for additional procedures when significant limitations or uncertainties remain.

NIST's AI Risk Management Framework applies the same discipline to AI evaluation: measurement should connect to deployment context, testing conditions should be documented, limitations beyond those conditions should be stated, and results should be interpreted in context. Neither source offers a universal document count because the question is whether the evidence supports the bounded conclusion.

In everyday AI-enabled work, a model can create the appearance of sufficiency by repeating one source in several forms. A report, dashboard, and slide may all derive from the same incomplete dataset. Three documents are then one evidence lineage, not three independent confirmations. The worker must test coverage, quality, independence, and counterevidence before strengthening the conclusion.

Use an ENOUGH assessment

Build one ENOUGH assessment for each material conclusion:

| Element | Sufficiency question | Evidence artifact | | --- | --- | --- | | End use | What decision or action will this claim support? | Named user, action, and consequence of error | | Necessary claim | What exactly must be true, and how strong is the proposed wording? | Testable claim with scope, period, and population | | Origins | Which independent evidence lineages support or challenge it? | Source map with derivation links | | Uncertainty | What limitations, missing cases, or conflicting evidence remain? | Gap and counterevidence log | | Grade | Is support adequate for a fact, estimate, range, hypothesis, or no conclusion? | Evidence grade with rationale | | Hold point | What additional evidence or review would change the grade? | Named procedure, owner, and decision deadline |

The evidence grade should control the language of the output. A directly verified fact may support a categorical statement within its scope. A calculated estimate should retain its assumptions and range. A plausible explanation with limited support should remain a hypothesis. When a decision cannot responsibly proceed, “insufficient evidence” is a useful conclusion, not a failed analysis.

Do not average weak sources into strength. Consider whether evidence is relevant to the claim, reliable for the intended use, sufficiently complete, current enough, and independent of the other support. The assessment should also search for evidence that could disconfirm the preferred conclusion.

Worked example: a synthetic support-process change

Imagine a synthetic service-operations analyst asked whether a new AI-assisted triage process reduced ticket resolution time. The evidence pack contains a dashboard showing a 22 percent reduction, a manager email praising the rollout, ten ticket examples, and a process note. This example is invented and does not describe a Two Prune customer.

An AI assistant drafts the conclusion, “The new process improved resolution time by 22 percent without harming quality.” The analyst opens an ENOUGH assessment before using that sentence.

The analyst revises the decision note: “The pilot dashboard reports a 22 percent lower median resolution time for included tickets. Because case mix, pause rules, reopened tickets, and quality outcomes are not yet comparable, the evidence is insufficient to attribute the difference to the AI-assisted process.”

A small sensitivity table then shows the result with reopened tickets included and with the two pause definitions. The effect narrows substantially. The expansion decision is not automatically rejected; instead, the note proposes a bounded second pilot with the missing checks.

The useful artifact is the claim-to-evidence assessment. It prevents a precise number from carrying more certainty than its lineage and coverage support.

• End use: the operations director will decide whether to expand the process to another team. A false positive could increase unresolved customer issues.
• Necessary claim: tickets using the new process had lower comparable resolution time during the pilot, and the observed change was not explained by case mix or measurement changes.
• Origins: the dashboard and manager email both derive from the same operational system and rollout team. The ten tickets were selected by that team. These are not three independent lineages.
• Uncertainty: the dashboard excludes reopened tickets, the pilot group handled fewer complex cases, the pre-period used a different pause rule, and no complete quality review is present.
• Grade: the evidence supports a descriptive observation about the reported dashboard period, not a causal improvement claim and not the claim about unchanged quality.
• Hold point: rerun the analysis with a stable resolution definition, comparable ticket groups, reopened tickets included, and a documented quality sample.

Practice by changing the decision threshold

Give a learner the same evidence pack for two different tasks. First ask for a low-risk internal hypothesis to guide further investigation. Then ask for a recommendation to make an expensive, difficult-to-reverse change.

The learner should not simply reuse the same conclusion. The evidence may be sufficient to justify another test but insufficient to justify a broad rollout. Ask the learner to complete an ENOUGH assessment for both uses and explain which additional evidence becomes necessary when the consequence of error increases.

Add one apparent corroborating source that actually derives from the same dataset, one genuinely independent source, and one piece of counterevidence. The learner should map origin rather than count files.

Useful practice behaviors include:

• writing the exact claim before gathering support;
• linking the required evidence strength to the intended action;
• tracing apparently separate sources to their origins;
• checking relevance, reliability, completeness, and currency;
• actively seeking counterevidence and alternative explanations;
• grading the conclusion rather than forcing a yes-or-no answer;
• naming the smallest additional procedure that could resolve the gap.

Assess observable sufficiency judgment

In an assessment, provide a realistic decision with incomplete but plausible evidence. Observe whether the participant:

Do not score sufficiency by whether the participant reaches a preferred conclusion. Strong performance may consist of narrowing a claim, deferring one part of a decision, or identifying that existing evidence supports only a test. Sufficiency judgment and subject-matter correctness should be reviewed separately.

• defines the decision and consequence of error;
• states a claim that is specific enough to test;
• distinguishes source count from independent evidence lineage;
• notices selection, coverage, definition, timing, or missingness limitations;
• looks for corroboration and counterevidence;
• scales conclusion strength to the support actually available;
• changes categorical language to a range, estimate, hypothesis, or limitation when needed;
• uses AI to organize or challenge evidence without letting fluency determine sufficiency;
• proposes proportionate additional procedures rather than requesting all possible data;
• communicates what can still be decided despite the gaps.

Verify through a claim stress test

First, ask an independent reviewer to identify the strongest alternative conclusion consistent with the evidence. If that alternative is still plausible, record which missing evidence separates the two.

Second, remove the most influential source and reassess the grade. A conclusion that collapses may depend on a single point of failure. This does not make the source invalid, but it should be visible to the decision maker.

Third, trace every material sentence in the final artifact to the ENOUGH record. Confirm that facts, estimates, hypotheses, and unknowns use distinct language. Check that the stated hold point would actually resolve the named gap rather than merely add more documents.

Evidence sufficiency is specific to a claim, decision, and evidence environment. It does not prove that a conclusion will remain supported as new information appears. In a Two Prune assessment, the ENOUGH record, source review, AI interactions, counterevidence search, sensitivity checks, revisions, and final decision note can contribute bounded evidence about Evidence Navigation, Verification & Risk Control, Judgment & Synthesis, and Communication & Delivery. They do not make an employment decision.

Sources

US GAO Government Auditing Standards 2024 Revision

• NIST AI Risk Management Framework Core

Sources: [1] [2]