Human accountability in AI-enabled knowledge work

Name the accountable decision

Start with a consequential verb: approve a payment, release a report, change a customer record, submit a recommendation, notify an external party, or escalate a control issue. Identify who has formal authority for that action and which specialists advise them. Separate business ownership, technical operation, source-data authority, risk review, and legal interpretation.

Avoid naming “the human” or “the manager” without a role and escalation route. A reviewer who can comment but cannot pause release is not the accountable decision-maker. A person who sees only a model summary cannot verify facts hidden in the source materials.

The NIST AI RMF Govern guidance calls for documented roles, responsibilities, communication lines, legal and regulatory requirements, and ongoing review. The guidance is voluntary and adaptable. It supports a structure for accountability but does not establish that a particular organization has met its obligations.

Sources: [1]

Design an accountability checkpoint

For each decision, record the proposed output, authoritative inputs, model contribution, material uncertainties, required checks, named reviewer, release authority, and stop conditions. The reviewer should be able to open the sources, inspect relevant transformations, record disagreement, request more information, and select a safe fallback.

Define what evidence proves the checkpoint occurred. An approval timestamp alone may show a click but not what was reviewed. Useful evidence may include sampled source comparisons, reperformed calculations, issue dispositions, revised wording, exception records, and an explicit release decision. Retain only evidence that is lawful and proportionate for the workflow.

The GAO AI Accountability Framework organizes accountable AI practice around governance, data, performance, and monitoring. It was designed for federal agencies and other entities and is not a universal compliance standard. Its emphasis on documentation and monitoring supports connecting a named decision to evidence and later review.

Sources: [2]

Worked artifact: a synthetic board briefing

Consider a synthetic strategy team using AI to draft a board briefing about a new market. The model compiles source notes and proposes a launch recommendation. The accountability checkpoint identifies the strategy director as release authority, the research lead as source reviewer, and finance as owner of the forecast assumptions.

During review, the research lead finds that the model combines a national market estimate with a city-level customer count. Finance discovers that the sensitivity table uses an outdated exchange rate. The recommendation is paused. The team corrects the denominator, reruns the sensitivity calculation, and changes the board language from a confident launch instruction to a staged test with explicit unknowns.

The checkpoint record retains the rejected draft, source discrepancy, corrected calculation, reviewer dispositions, unresolved assumptions, and final release approval. The scenario is synthetic. It does not show that a particular control works in production or that a business decision will succeed.

Test whether review is meaningful

Run exercises that make disagreement possible. Include conflicting evidence, a plausible unsupported claim, an urgent deadline, or a proposed action outside the reviewer's authority. Observe whether the reviewer can find the issue, access the source, understand the consequence, stop the workflow, and route the decision correctly.

Test workload as well as interface design. If one reviewer receives more cases than can be examined, formal authority may exist without practical control. Sample approvals for depth, measure unresolved exceptions, and inspect whether teams route around the checkpoint. Do not equate a high approval rate with control effectiveness.

Reopen accountability after change

Review the checkpoint when the intended use, model, source system, permissions, automation level, affected population, or downstream consequence changes. Confirm that role assignments still match actual authority and that reviewers retain access to the evidence they need. Record departures from the designed process and decide whether to narrow, pause, or redesign the workflow.

This approach makes human responsibility executable and reviewable. It does not provide a legal opinion, certify control effectiveness, or make human oversight sufficient for every use. Two Prune can observe task-bounded evidence of how a participant reviews, verifies, and escalates work in a simulation. It does not operate customer controls or make employment decisions.