Practical guide

What evidence shows human review is operating in an AI workflow?

A policy that says “a human reviews the output” is design evidence, not proof of operation. Operating evidence must show that an authorized reviewer received the right version and material sources, could intervene before the relevant effect, made a recorded disposition, and did so across the in-scope population. Use the REVIEW evidence test to distinguish a visible review step from a functioning.

By Two Prune

A policy that says “a human reviews the output” is design evidence, not proof of operation. Operating evidence must show that an authorized reviewer received the right version and material sources, could intervene before the relevant effect, made a recorded disposition, and did so across the in-scope population. Use the REVIEW evidence test to distinguish a visible review step from a functioning control: - R — Reconcile the in-scope population - E — Establish the required review object - V — Verify authority and timing - I — Inspect challenge and intervention - E — Evidence the disposition and release - W — Watch exceptions and retest

Define the control objective

Write the objective in terms of the consequence to be controlled. For example: “Before an AI-assisted customer response is sent, an authorized service lead reviews its material claims against the cited sources and records approve, revise, reject, hold, or escalate.”

Then define scope: workflows, channels, thresholds, systems, risk tiers, time period, reviewer roles, and approved bypass or emergency procedures. Without scope, a sample of convenient approval events cannot support a conclusion about the workflow.

NIST’s AI Risk Management Framework Core calls for documented human-AI roles, assessed human oversight, defined responsibilities, monitoring, and processes for responding to risks. It also emphasizes documenting system limits and the context in which outputs are interpreted. NIST AI RMF Core

Sources: [1]

R — Reconcile the in-scope population

Start from the event that creates review obligation, not the approval table. If the control covers every AI-assisted response, construct the population from generated or queued responses and reconcile it to sent, cancelled, failed, and pending outcomes.

Useful evidence includes:

An approval log alone can prove that some approvals occurred. It cannot reveal items released without ever reaching the log.

  • workflow-item IDs and versions;
  • creation, queue, review, release, cancellation, and failure events;
  • channel and risk classification;
  • system totals for the period;
  • documented filters and exclusions;
  • bypass, manual, API, integration, and bulk-action paths.

E — Establish the required review object

For each selected item, show what the reviewer could see at decision time:

A later screenshot is weak unless it preserves the same version and state. Prefer immutable snapshots, hashes, or versioned records. If the interface hides a failed rule or source date, the person’s presence does not establish meaningful review of that attribute.

  • the exact AI-assisted output version;
  • source documents and source-to-claim anchors;
  • relevant policy or rule version;
  • calculations, warnings, conflicts, and missing evidence;
  • the decision requested and possible dispositions;
  • material model or service information when required by policy.

V — Verify authority and timing

Match the reviewer identity to the authority matrix that was valid at the time. Test thresholds, region, information class, delegation, training or qualification conditions, conflicts, and expiry.

Compare timestamps in sequence. Review must occur before the effect it is designed to control. A later confirmation may support monitoring or remediation, but it does not convert an earlier unreviewed release into a pre-release approval.

Record clock source, timezone, system latency, and any batch processing that could change apparent order. Challenge an approval timestamp that precedes output creation or follows release.

I — Inspect challenge and intervention

Evidence that a reviewer clicked “approve” does not show whether they could or did challenge the output. Inspect whether the workflow offered meaningful options and whether the reviewer used them when evidence called for it.

Strong evidence can include:

Interpret a very high approval rate carefully. It may reflect consistently good inputs, or it may reflect weak challenge, selection bias, or an interface that makes intervention difficult. Use item evidence and designed tests rather than a target rejection rate.

  • revise, reject, hold, and escalate events;
  • comments linked to a source or failed check;
  • corrections between reviewed and released versions;
  • requests for missing evidence;
  • escalation to an exception authority;
  • challenge tests that insert a stale source, conflicting instruction, incorrect calculation, or unauthorized reviewer.

E — Evidence the disposition and release

Link the disposition to the exact output and source-manifest versions. The event should record reviewer identity, authority basis, timestamp, choice, reason where required, conditions, and any escalation.

Then link release to that disposition. Test that a material edit, source replacement, or rule change invalidates the earlier approval and returns the item to review. Preserve prior rejections and revisions rather than overwriting them.

The U.S. Government Accountability Office’s AI Accountability Framework organizes evidence questions across governance, data, performance, and monitoring. Its lifecycle and traceability emphasis supports retaining the decision sequence and the evidence available to accountable people, while the organization still determines the controls and obligations applicable to its context. GAO AI Accountability Framework

Sources: [2]

W — Watch exceptions and retest

Maintain an exception register with:

Test bypasses intentionally. Attempt release through an older integration, direct API, bulk action, expired delegation, edited approved output, and incomplete evidence package. Sample routine items and all high-consequence exceptions. Reconcile monitoring results back to the source population.

  • affected item and population scope;
  • failed criterion and observed evidence;
  • consequence and interim protection;
  • root cause and whether it is confirmed;
  • owner, due date, disposition, and authority;
  • remediation evidence and retest result;
  • remaining uncertainty and next review date.

Copyable human-review evidence matrix

| Item | Scope / source event | Review object version | Reviewer authority | Review time | Challenge or correction | Disposition | Released version / time | Exception / retest | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | HR-001 | queue event and risk tier | output + source manifest | role, threshold, delegation | before effect | source check and change | approve / revise / reject / hold / escalate | exact linked version | owner, action, retest |

Worked test: policy-summary release

An internal communications team uses AI to draft policy summaries. The control requires a policy owner to approve every summary before publication. The approval table contains 412 approvals for the quarter, all with named reviewers.

The REVIEW population starts from 419 generated summaries. Four were cancelled, 412 were approved and published, and three were published through a legacy content integration without an approval event. Of the 412 approved items, two were materially edited after approval but retained the old disposition.

Authority testing finds all named reviewers were valid. Review-object inspection finds that ten records lack the source-policy version, so the test cannot establish what those reviewers compared. A challenge test using an outdated policy triggers no warning, though the reviewer can still choose hold after manually noticing the date.

The evidence supports that authorized review occurred before release for 400 items with a complete review object and unchanged released version. It also identifies three bypasses, two invalidated approvals, and ten evidence limitations. The owner disables the legacy integration, adds approval invalidation after material edits, requires a versioned source manifest, and introduces a stale-source warning. Retest evidence covers those repairs; it does not erase the quarter’s exceptions.

Interpretation limits

Human-review evidence can show whether the defined control operated for the tested scope and period. It does not establish that every reviewer judgment was substantively correct, that the control will operate in the future, or that the workflow meets every legal requirement. Those conclusions require appropriate criteria, expertise, and additional evidence.

Two Prune can assess how a participant reviews AI-assisted work, checks sources, handles uncertainty, invokes authority, and records a defensible disposition in a realistic simulation. It does not operate the customer’s review control or decide the real-world release.

Sources

  1. 1.NIST AI RMF Core · National Institute of Standards and Technology
  2. 2.GAO AI Accountability Framework · U.S. Government Accountability Office