On this page
Key takeaways
- Give each AI agent its own identity with the least access it needs; approvals only mean something if you can see which identity acted.
- Review the actions that are costly, sensitive or hard to undo, and let routine work run.
- Every approval queue needs a named reviewer, a named substitute and a timeout decision, or it becomes a second backlog.
An AI agent approval workflow pauses an agent before selected actions, sends the proposed action to a person, and only continues when that person approves. For a small business the hard part is not the pause. It is making sure the right person sees the request in time, without approvals piling up behind one busy owner.
In a small business the quickest way to automate something is to hand the automation someone's login. It works on day one. By month three nobody can tell whether a refund, a deleted record or an email came from a person or a workflow.
Identity comes before approvals
An approval gate answers "should this happen?". It cannot answer "who did this?" if the agent and a person share an account. Give every agent its own service identity, scoped to the systems and actions it genuinely needs.
Three rules cover most small businesses:
- Scoped access: an agent that answers order questions can read orders. It cannot touch payroll or delete customers.
- Named approvers: refunds over a threshold, messages to key accounts and any deletion go to a specific person, not a shared inbox.
- One audit log: every action, by every agent, in one place you can search.
Decide which actions need review
| Action | Typical treatment in a small business |
|---|---|
| Reading records, drafting replies | Runs without review |
| Routine replies from approved templates | Runs, with a weekly sample review |
| Refunds or credits under a small threshold | Runs, logged, reviewed weekly |
| Refunds above the threshold | Approval by the finance owner |
| Messages to key accounts | Approval by the account owner |
| Changing customer or supplier records | Approval by the record owner |
| Deleting records or exporting customer lists | Not allowed for the agent |
Start strict. Review what the agent proposes for a few weeks, then raise thresholds where its decisions have been consistently right. Trust built on evidence lasts longer than trust granted up front.
Design the queue so it does not become a backlog
In a team of five, the finance owner is also the person on holiday next week. An approval design that ignores that will stall.
| Element | Example |
|---|---|
| Reviewer | Finance owner for refunds above the threshold |
| Substitute | Operations lead, automatically after the response window |
| Response window | Four working hours for refunds, one working day for record changes |
| Timeout decision | Notify the substitute; if still unanswered, return the case to manual handling and tell the customer when to expect an answer |
| Never on timeout | Execute the action anyway |
| Daily summary | Pending approvals, how long each has waited, who holds it |
Make each request quick to decide: the action, the evidence, the amount or change, and approve, edit or reject on one screen. A request that needs three systems to check will wait.
A worked example
A hypothetical online shop lets an agent handle delivery complaints. The agent checks the order and courier, and proposes a £60 refund for a lost parcel. The threshold for automatic refunds is £30, so the request goes to the finance owner with the order, the tracking history and the customer's messages attached.
The finance owner is away. After four working hours the request moves to the operations lead, who approves it. The agent issues the refund, replies to the customer and writes the decision, approver and outcome to the log. The weekly review shows that most refunds between £30 and £60 were approved unchanged, so the owner raises the automatic threshold to £45 for the next month.
Measure the queue
- Time from request to decision, by approval type.
- Requests that timed out and went to the substitute.
- Approved unchanged, edited and rejected, by type.
- Actions completed after approval.
If most requests of one type are approved unchanged, the threshold is probably too low. If many are rejected, the agent needs better rules or data.
For a complete list of reviewer, timeout, changed-input and failure checks, use the approval workflow automation checklist. GOFTUS builds agentic AI systems with approvals designed for the size of your team. Talk to us.
Frequently asked questions
It can, but you lose the ability to tell its actions from a person's. Give it its own scoped identity.