zero gravitySOLUTIONS
← All insightsAI & AUTOMATION / PRODUCT ARCHITECTURE

Design a human approval inbox for AI actions

QUICK ANSWER

A human approval inbox should show the exact action, supporting evidence, and affected records before execution. Bind the decision to a versioned proposal, recheck permissions and preconditions at execution, and make stale approvals and uncertain outcomes visible.

Inside this article 4 sections

An approval button does not create meaningful human control unless the reviewer knows what will happen. “Approve AI recommendation” is too vague when the action might change a customer record, send a message, or create an external transaction.

Design the inbox around a proposed change that can be inspected, accepted, rejected, or allowed to expire. The model produces the proposal; the application controls its permissions, lifetime, and execution. This distinction makes approval useful even when the underlying model changes.

Show the exact change and its context

A proposal should identify the action type, target record, current state, proposed state, evidence, and reason for review. For a subscription adjustment, show the account, existing entitlement, requested entitlement, effective date, and any customer-facing message. Do not hide a material side effect inside a collapsed technical payload.

Use a stable proposal identifier and version. Store the precise payload that was reviewed, plus the relevant source-record version and policy version. If the model rewrites the proposal after review, that is a new version requiring a new decision. Approval of an earlier summary should not authorize whatever the latest model output happens to contain.

Keep the reviewer’s choices concrete. Approve executes the displayed proposal subject to final checks. Reject closes it with a reason. Edit creates a new proposal reflecting the reviewer’s changes. Request information returns the task to a defined owner. Each choice should produce a durable event rather than a transient interface message.

An approved proposal can still become stale

CONCURRENT CHANGE / SUBSCRIPTION EXAMPLE
  1. Proposal createdAccount version 18

    AI proposes extending an active subscription by one month.

  2. Human decisionProposal approved

    The reviewer accepts the displayed change for version 18.

  3. Before executionAccount cancelled

    Another process writes account version 19.

  4. Execution checkReturn to review

    The expected version no longer matches; no extension is applied.

Approval records intent for a specific proposal. It does not remove the need to check whether that proposal still applies when execution begins.

Check current authorization and business preconditions immediately before execution. A reviewer may have lost the required role; the customer may have cancelled; the allowed adjustment policy may have changed. A long-running review must not preserve powers that the user no longer holds.

Use a conditional update or an appropriate transaction for the record change. Reading a row, checking its version in application memory, and later writing without a condition leaves a race. For PostgreSQL, transaction isolation and locking choices matter; even a stronger isolation level needs correct handling of conflicts and retries.

External actions need a different strategy because the local database cannot usually transact atomically with a remote service. Store an execution intent, use the remote system's idempotency mechanism where available, and reconcile outcomes. Do not silently retry a non-idempotent action just because the approval still exists.

Give every non-happy state a useful next step

Use distinct states for awaiting review, approved, executing, completed, rejected, expired, stale, and execution uncertain. The last state matters when a request may have succeeded remotely but its acknowledgement was lost. Calling it failed can encourage a reviewer to approve a duplicate action.

Show who owns the next step and how long the item has been in that state. An expired proposal should explain why it can no longer execute and offer a fresh proposal if appropriate. A stale item should highlight the changed business facts instead of making the reviewer compare two large raw payloads.

Prioritize by consequence and age, not just arrival order. Time-sensitive actions may require earlier attention, but bulk approval should be limited to proposals whose effects can actually be reviewed together. A checkbox that accepts fifty unrelated actions weakens the very control the inbox was intended to provide.

Workflow engines can suspend execution until a human decision arrives; callback patterns are one implementation option. The suspension mechanism does not define who can approve, what they approved, or whether the action remains valid. Those rules belong to the application and must survive restarts and duplicate callbacks.

Test the decision boundary under interruption

Exercise two reviewers opening the same item, a double click on approve, role removal during review, an edited proposal, expiry during execution, and a lost remote response. Verify that each path leaves a comprehensible record and cannot execute an unauthorized duplicate.

Measure how often reviewers change or reject proposals, how long decisions wait, and what happens after approval. A very high approval rate can mean good proposals, but it can also mean poor visibility or rubber-stamping. Review a sample of approved outcomes against the underlying evidence.

For record intake, read document extraction validation. For execution and recovery, see reliable background jobs. Our AI integration work joins these boundaries into a coherent operational workflow.

FROM INSIGHT TO ACTIONAI-native implementation

Build an approval workflow people can trust.

We help connect AI proposals to readable review screens, durable execution, and explicit recovery states in your existing application.

Explore AI implementationDiscuss your AI workflow

Start with your goals.
We’ll work through the next step together.