Practical guide
How to operationalize human accountability in an AI workflow
“Human in the loop” is not an operating control. A person can be present yet lack the evidence, time, authority, or interface needed to change the outcome. Human accountability becomes operational only when the workflow names the decision, preserves the review object, gives an authorized person meaningful choices, records the disposition, and tests whether the control worked. Use the PAUSE.
“Human in the loop” is not an operating control. A person can be present yet lack the evidence, time, authority, or interface needed to change the outcome. Human accountability becomes operational only when the workflow names the decision, preserves the review object, gives an authorized person meaningful choices, records the disposition, and tests whether the control worked. Use the PAUSE workflow to turn a policy principle into an inspectable control: - P — Prepare the review object - A — Assign authority and conditions - U — Use a real intervention point - S — Seal the disposition and evidence - E — Evaluate operation and exceptions
P — Prepare the review object
Define what the person will review before the AI output is created. The review object should contain:
Do not ask a reviewer to approve a polished answer while hiding its sources or failed checks. For a high-consequence step, make the source-to-claim trail available in the same review surface. If the evidence package is incomplete, the workflow should support “hold for information” rather than forcing approve or reject.
- the source inputs and their versions;
- the model or service and material configuration;
- the task instruction and output;
- rule checks, calculations, or comparison results;
- known limitations and missing evidence;
- the decision requested and possible consequences.
U — Use a real intervention point
Place the review before the output causes the relevant effect. A review after a customer message has been sent is monitoring, not pre-release approval. The person must be able to choose among meaningful dispositions:
Design against automation bias. Show evidence gaps and failed checks prominently. Avoid a default-selected approval button, an unexplained confidence score, or a queue that rewards speed without surfacing consequence. Require a reason for exceptions and high-consequence approvals, but do not collect empty boilerplate on every routine item.
The workflow state should be explicit:
draft → evidence ready → review → approve / revise / reject / hold / escalate → release or close
Release is permitted only from an allowed disposition, by a person with current authority, while all required evidence is present. A later edit to the output or source set should invalidate the earlier disposition and return the item to review.
- approve;
- approve with a recorded condition;
- revise and return;
- reject;
- hold for more evidence; or
- escalate to the appropriate authority.
S — Seal the disposition and evidence
Create an append-only decision event containing:
Do not overwrite a rejection when the item is revised. Link the new review to the prior event so that the sequence remains understandable. Retention and access should match the sensitivity of the underlying workflow.
The U.S. Government Accountability Office’s AI Accountability Framework emphasizes governance, traceability, documentation, performance monitoring, and accountability across an AI system’s life cycle. It provides useful questions for testing whether an accountability record is specific enough to support review, while the organization must still determine its applicable obligations. GAO AI Accountability Framework
- workflow item and version;
- source-manifest version;
- material output version or checksum;
- rule results and visible warnings;
- reviewer identity and authority basis;
- disposition, reason, conditions, and timestamp;
- delegation or escalation path;
- released version and release event.
Sources: [2]
E — Evaluate operation and exceptions
Test the control with evidence, not a policy walkthrough alone. Select or define the population, reconcile it, and check:
Run challenge tests as well as record inspection. Remove a required source, use an expired delegate, edit an approved output, create a conflicting rule result, and attempt release from an unauthorized account. Record expected and actual behavior. A successful demonstration on a few cases does not establish ongoing control operation; define monitoring and a review date.
- Did every in-scope item reach the review state before release?
- Did the reviewer have current authority for that item?
- Were required sources, warnings, and rule results visible?
- Did the event log preserve the actual disposition and released version?
- Were exceptions routed to the right authority?
- Did revisions trigger a new review?
- Could a user bypass the control through another path or integration?
Copyable workflow control matrix
| Control element | Design record | Operating evidence | | --- | --- | --- | | Objective | decision and consequence requiring human accountability | current approved workflow scope | | Review object | required sources, output, checks, limitations | versioned evidence package | | Authority | roles, thresholds, delegation, exception rights | identity and authority snapshot | | Intervention | state transition and permitted dispositions | timestamped pre-release event | | Release rule | conditions that permit release | released version linked to approval | | Exception | hold, escalation, incident, fallback | owner, reason, due date, disposition | | Monitoring | population, tests, frequency, reviewer | reconciliation, test results, retest |
Worked example: customer refund exception
An AI assistant reviews a customer message, order history, and refund policy, then recommends a refund disposition. Routine refunds below $200 may be approved by a service lead. Policy exceptions require a regional operations manager.
For one request, the assistant recommends a $340 refund and cites an outdated policy. The review object shows the order history, output, source manifest, threshold check, and a warning that the policy version is stale. The service lead cannot approve because both amount and policy exception exceed their authority. They choose “hold for evidence,” attach the current policy, and send the item back for revision.
The revised output recommends a $180 partial refund under the current policy. The workflow does not reuse the earlier review: the changed source and output create a new version. The service lead approves with a reason, and the release event links the exact version to that disposition.
During monthly testing, the control owner reconciles 1,240 in-scope recommendations to the event log. Two items were released through an older integration without a valid disposition. They are operating exceptions even though later manual review found the amounts reasonable. The owner disables the path, records the affected scope, and schedules a retest. Human agreement after release does not erase the bypass.
Failure conditions to monitor
The workflow is not operating as intended when:
Two Prune can assess whether a participant notices missing evidence, uses AI thoughtfully, invokes the right authority, verifies a recommendation, and records a defensible disposition in a realistic simulation. It does not become the accountable owner of the customer’s workflow, determine legal obligations, or establish control performance without the customer’s operating evidence and qualified review.
- the reviewer sees only the answer and not the material evidence;
- the person cannot revise, reject, hold, or escalate;
- authority is based on a title with no threshold or delegation rule;
- an approval survives a material change to sources or output;
- the release path can bypass the recorded disposition;
- exception reasons are overwritten or inaccessible;
- monitoring excludes failed, abandoned, or manually routed items; or
- the same person designs, operates, approves, and tests a high-consequence control without an explicit independence decision.
Sources
- 1.NIST AI RMF Core · National Institute of Standards and Technology
- 2.GAO AI Accountability Framework · U.S. Government Accountability Office