Quick answer: Put the approval decision between the AI-generated recommendation and the external action. An actual approval step must stop execution until a named person accepts, rejects or edits the proposal.
Draw the boundary before choosing software
A useful diagram for a two-person consultancy looks like this: new lead form → input validation → AI summary → owner review → permitted action → audit entry. The expensive mistake is to put “send email” before “owner review.” An approval button after the email is sent is not a control.
A human reviewer needs enough context to disagree. Show the original message, the proposed output, any source records used, the reason for the recommendation and the downstream consequence. If the interface only says “AI recommends approve,” it encourages rubber-stamping.
An approval queue with four explicit states
| State | Owner action | Automation behavior |
|---|---|---|
| Pending | Review input and proposal | Hold all outward actions |
| Approved | Confirm specific operation | Execute once with unique request ID |
| Rejected | Record reason or correction | Stop action and retain minimal log |
| Expired | No decision by deadline | Escalate; never silently approve |
For example, a UK trades company might require approval before sending a quotation. The reviewer checks that the address is within its coverage, that prices come from the current approved list and that the message does not make a guarantee the firm cannot keep.
Build it without depending on a special agent feature
Use a workflow tool that can pause execution or create a durable approval item. Store a unique task ID, a payload hash and the current state. On approval, verify both state and reviewer authority. Make retries idempotent: repeated clicks must not cause repeated emails or duplicate CRM updates.
Set deadlines, notifications and an owner for unresolved items. Do not let the AI choose its own approver. For sensitive information, provide masked summaries in the notification and let authorized staff open the secure record.
Test what happens when things go wrong
- The approver rejects the draft: no external action should occur.
- A staff member double-clicks approve: execute once.
- The workflow loses connectivity: retain pending state and retry safely.
- The input is changed after review: require a fresh approval.
- A reviewer is unavailable: escalate to a second authorized person, not the AI.
When human review is essential
Treat financial transfers, binding offers, employment decisions, account permissions and public claims as high-impact actions. Legal requirements vary across the US, UK and EU. The UK’s data regulator also stresses that a review must be meaningful, not merely someone clicking through a recommendation.
Short answer to a frequent question
Can I automate low-risk confirmations? Potentially, after testing fixed text and permissions. Keep approvals for messages that could mislead customers or expose their information.
References and further reading
These references provide context and official guidance. The sample workflows above are illustrative designs, not vendor performance tests.
Published by: FlowHarbor Editorial Desk. Read our sourcing and review policy.