How to Build AI Agent Approval Workflows

Create three tiers based on reversibility and impact. Tier 1 (auto-execute): fully reversible, low impact. Logging a note, updating an internal record, sending an internal notification. Tier 2 (notify and proceed): semi-reversible, moderate impact. Sending a follow-up email, updating a CRM stage, scheduling a meeting. Tier 3 (approval required): irreversible or high impact. Sending client-facing communications, processing refunds, publishing content, modifying financial records.

Map every action your agent can take into one of these tiers. Be conservative initially — it's easier to relax controls than to recover from a mistake. You can always move an action from Tier 3 to Tier 2 after the agent proves reliable.

Set up a dedicated approval channel in Slack, Telegram, or Microsoft Teams. When the agent triggers a Tier 3 action, it sends a formatted message to this channel: what the agent wants to do, why it decided to do it (reasoning summary), the relevant context (customer name, amount, content), and approve/reject buttons.

The message format matters. Reviewers need enough context to make a quick decision without reading the full agent log. I structure approval requests as: 'Agent X wants to [action]. Reason: [1-sentence summary]. Context: [key details]. Approve / Reject / Modify.'

When an agent hits a Tier 3 action, it pauses execution and stores the pending action in a database table. The table holds: agent ID, proposed action, full context, timestamp, status (pending/approved/rejected), and reviewer ID. The agent sends the approval request and moves on to its next task (or idles if it's a single-task agent).

A listener watches the approval channel for responses. When a reviewer approves, the system retrieves the pending action and executes it. When rejected, it logs the rejection reason and optionally notifies the agent to try a different approach. Set a timeout — if no one responds within 2 hours, send a reminder. If still no response after 4 hours, escalate.

Log every approval decision: what was proposed, who reviewed it, was it approved or rejected, how long the review took, and the rejection reason if applicable. Analyze this data monthly.

If an action type is approved 95%+ of the time, consider moving it to Tier 2 (notify and proceed) or Tier 1 (auto-execute). If an action type has a 20%+ rejection rate, the agent's decision logic for that action needs improvement. The approval system isn't just a safety net — it's a feedback loop that makes the agent better over time.

Build fallbacks for when the approval system itself fails. What happens if the Slack webhook is down? What if the approver is on vacation? What if the pending actions table fills up with stale requests?

Design a fallback chain: primary approver, secondary approver, escalation to admin. Set a maximum queue size — if more than 20 actions are pending, something is wrong and the system should alert the operations team. Configure a nightly sweep that closes stale pending actions (older than 24 hours) with a 'timed out' status and alerts the relevant team.