# Human approval policies for AI agents: what to approve and when https://palnau-systems.vercel.app/insights/human-approval-policies-ai-agents Published 2026-11-10 by Oliver Palnau and the Palnau Systems team. Prepared with AI assistance. Audience: Managers responsible for AI actions and customer commitments. An effective AI approval policy names which actions can run automatically, which require a person and what information that person must see. Base the rule on consequence, reversibility and the authority already granted to the workflow. Approval should cover a specific action with a specific scope. A generic “proceed” button does little good if the recipient, amount or underlying record can change before execution. ## Key takeaways - Tie approval to the exact proposed action and its evidence. - Keep routine checks automatic and route actual judgment to a person. - Expire approvals when material inputs change or the approval window closes. ## Classify actions by consequence Build an action inventory before choosing approval thresholds. Reading a permitted knowledge article, drafting an internal summary, changing a customer record and issuing a payment have different consequences. Consider what could go wrong, who would be affected and whether the action can be reversed reliably. Have the relevant business owner set the policy rather than asking the model to decide its own authority. Use a small number of categories people can understand. For example: prepare freely within approved data access, execute routine updates within a defined boundary, request approval for external commitments and stop for prohibited actions. The categories should map to actual controls in the connected systems. ## Make the approval packet useful Show the proposed recipient or record, the action, the relevant amount or terms, the source evidence and the reason approval is needed. Give the reviewer an edit path and a reject path. Asking someone to inspect a long chat transcript creates unnecessary work and makes important details easier to miss. In a hypothetical equipment company, an agent prepares a customer credit adjustment but cannot post it. The manager sees the original invoice, service issue, proposed amount and existing adjustments. If the invoice balance changes before approval, the proposal is refreshed. This prevents a stale decision from being treated as permission for a different transaction. ## Design the waiting and rejection states A workflow must know what happens while approval is pending. Set an owner, backup, due time and escalation route. Avoid continuing downstream steps that assume approval will eventually arrive. If a reviewer rejects a proposal, capture the reason and return the case to a clear state rather than letting the agent repeatedly request the same permission. Use rejection reasons that improve the process: incorrect facts, policy mismatch, incomplete evidence or business judgment. Keep sensitive comments within the appropriate access boundary. A good approval queue becomes a source of workflow improvements without turning every managerial preference into an automatic rule. ## Test approvals as part of the product Run tests where the amount changes, the customer record is ambiguous, the reviewer is unavailable or two people approve at nearly the same time. Verify that one approved action produces one execution. Check that the system can revoke a pending action and that an expired approval is not reused. Copy-ready policy brief: “For this action, show the exact proposed change and the evidence supporting it. Route to the named approver. Execute only the approved version. Recheck material inputs immediately before execution. If those inputs changed, pause and request a new decision.” The implementation should enforce these conditions rather than relying solely on instruction wording. ## First steps 1. List every external action the workflow can take. 2. Assign approval categories based on consequence and reversibility. 3. Design a concise evidence packet with edit and reject controls. 4. Set ownership, expiration and change-detection rules. 5. Test stale, duplicate, rejected and unavailable-reviewer cases. ## What to measure Approval turnaround: Measure elapsed time from a complete approval request to a recorded decision. Material correction rate: Track approvals that require changes to facts, terms or recipients before execution. Unauthorized execution count: Audit actions executed without a valid matching approval where policy required one. ## Should every AI action require approval? That often creates a queue people cannot manage. Automate deterministic checks and reserve human attention for actions that need judgment or carry meaningful consequences. The correct boundary depends on the workflow, existing authority and the quality of its controls. ## Can approval be given in a chat message? It can be part of the interface if the system reliably identifies the approver, binds the decision to the exact action and records the result. An ambiguous message should not become permission for a broader set of actions. ## Sources and further reading - [OpenAI: Guardrails and human review](https://developers.openai.com/api/docs/guides/agents/guardrails-approvals) - [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) ## Work directly with Oliver and his team Work directly with Oliver and his team to design approval queues that make AI workflows useful while keeping business authority clear. Free introductory call: https://calendly.com/oliverpalnau/30min Services and scope: https://palnau-systems.vercel.app/services Current pricing: https://palnau-systems.vercel.app/pricing