Human-in-the-loop AI approval workflow
A human-in-the-loop AI approval workflow lets AI prepare work while authorised people decide whether changes proceed. For small business teams, Ai Flowi uses four action risk classes: reads run automatically, writes wait for approval by default, and money records need a second approver. Approval does not move funds.
Updated 21 October 2026; facts checked 21 October 2026
Technical review: Janice, Ai Flowi.
Step 1: separate reading from changing records
Start with the action the AI proposes. Looking at a permitted record is different from changing it, deleting it or recording a money action. Ai Flowi classifies each AI action before it runs.
| Risk class | What happens | Review boundary |
|---|---|---|
| Read | Runs automatically within the user’s permissions. | Access checks still apply. |
| Reversible write | Waits for human approval by default; an owner can set reversible writes to run automatically. | Can be undone within 24 hours. |
| Irreversible write | Always waits for human approval before proceeding. | Do not assume an undo is available. |
| Money | Waits for approval and requires a second approver. | An approved money record does not transfer funds. |
Step 2: give reviewers the right access
Set department access for the people who review work. Ai Flowi supports roles, department grants and sensitive-field grants. Salary and bank details require additional verification before viewing. Sensitive actions also use an additional identity check.
Assign work to a person or an AI staff member and track its status. Decide who should review each type of change before the team starts using it.
Step 3: review the proposed action in the queue
Risky AI actions wait in the tasks-and-approvals queue. Owners can approve or reject and record a reason. Read the proposed change and check the underlying business record before deciding.
For a money action, a second person must approve. This governs the money record; it does not pay a supplier, transfer salaries or collect an online payment.
Step 4: keep the decision history and check reversal limits
The activity log records AI decisions, writes, approvals and denied attempts. Audit and approval history cannot be edited or deleted. Use that history to review what the AI did and how a decision was handled.
Reversible actions can be undone within 24 hours. Check whether the action is reversible before approving it; the window does not make an irreversible action reversible.
Step 5: review learning proposals too
The assistant’s learning loop produces improvement proposals from outcomes and feedback. An owner can approve, edit, reject or roll back a version. Changes require approval. Good human sales replies can also become candidate improvements for owner review.
Approval features and what your team supplies
| Feature | What it does | What you need to supply |
|---|---|---|
| Approval queue | Holds risky AI actions for approval or rejection with a reason. | Your authorised reviewers; no external account for internal approval. |
| Second approver | Requires another person for money actions. | A second authorised person; no external account for the decision record. |
| Roles and sensitive access | Controls department and sensitive-field access, with extra verification for salary and bank details. | Your role and access decisions; no external account. |
| Activity and audit history | Records decisions, writes, approvals and denied attempts; audit and approval events cannot be edited or deleted. | Nothing external. |
| Reversal window | Allows reversible AI actions to be undone within 24 hours. | An eligible reversible action within the window; no external account. |
| Learning review | Lets an owner approve, edit, reject or roll back improvement proposals. | Your owner’s review; no external account for the review record. |
Example: a supplier payment record
Illustration – example setup, not a customer result.
Suppose an AI staff task proposes a supplier payment record. The money action waits for approval and a second approver. Your reviewers check the record before approving it. The activity history records the decision. Even after approval, no supplier funds are transferred by this action.
For a purchasing example, read the purchase order approval workflow guide. For conversation handoff, use the chat channel workflow guide.
Review the workflow in every core department
Start from the AI staff overview. Open each department page to choose the records and actions you want your team to review:
Read the AI agent audit trail and security guide for activity history and access controls.
Frequently asked questions
What does human-in-the-loop mean in AI workflows?
It means people review decisions at defined points in the AI’s work. In Ai Flowi, reads can run within permissions, while writes wait for approval by default. Money records need a second approver, and approval does not move funds.
What is human-in-the-loop authorization for AI agents?
It combines permission to access a record with a human decision before a governed action proceeds. Ai Flowi checks permissions and ownership, queues AI writes for approval by default and requires a second approver for money records.
What is human AI workflow?
In this context, it is work shared between people and AI staff. Ai Flowi lets you assign tasks to either, track their status and review AI changes through approvals. Owners also review learning proposals before changes are applied.
What is an approval workflow process?
It is the path from a proposed action to review and a recorded decision. In Ai Flowi, owners approve or reject queued actions with a reason. Money records require a second approver, and audit history preserves approval events.
Walk through an approval with your team
Bring one stock change, one payment-record scenario and your reviewer roles. Ask to follow the decision through the queue and activity history.