How to set human approval rules for AI collections
Many finance teams that try AI in collections start by having someone read every draft before it goes out. That works for a few weeks, until the review queue holds 200 emails and the team is approving reminders on $400 invoices one at a time.

Many finance teams that try AI in collections start by having someone read every draft before it goes out. That works for a few weeks, until the review queue holds 200 emails and the team is approving reminders on $400 invoices one at a time.
Approval rules work better when they’re set for each type of action. A write-off or a change to payment terms should always wait for a person, because it changes what a customer owes. Reminders on small, undisputed invoices can move to automatic sending once the drafts prove reliable. At Monk, a person can approve emails before they send and payment matches before they post.
Which actions should need approval at launch?
Start by listing what the platform can do, then decide the default for each action before go-live.
| Action | Default at launch | When to relax it |
|---|---|---|
| First reminder on a small, undisputed invoice | Approval for the first two weeks | Once reviewers rarely edit these drafts |
| Any message to a strategic account | Approval | Rarely, since most teams keep these accounts close |
| Reply to a dispute or a PO question | Approval | Not recommended, because disputes need judgment |
| Message that offers a payment plan or new terms | Approval by a manager | Not recommended |
| Cash match where deterministic matching is certain | Automatic | Already automatic |
| Suggested cash match | Approval | Once the same suggestion is accepted almost every time |
| Credit memo or write-off | Controller approval | Not recommended |
| Any write-back that changes an ERP balance | Approval, with a preview | Once previews match the final entries over several closes |
Most of the table stays gated. The items that relax over time are reminders and routine cash matches, where evidence builds up quickly. Anything that changes what a customer owes keeps a person in the loop.
How should approval rules be set?
Approval rules work best when they’re based on what could go wrong. Most teams combine four inputs:
- Invoice value. A threshold above which a person always approves, set low at first.
- Account flags. A strategic tier or an open dispute sends the item to review.
- Type of action. Anything that writes to the ledger carries more risk than a reminder email.
- Certainty. Suggested matches and drafts built on incomplete data wait for a person.
Write the rules down in one place and give them a named owner.
Who should approve what?
Match the approver to the risk, and keep approval away from the person who wrote the rule.
| Item | Usual approver |
|---|---|
| Routine collection email drafts | AR specialist who owns the account |
| Messages to strategic accounts | AR manager |
| Suggested cash matches | Cash application lead |
| Credit memos and write-offs above your threshold | Controller |
| Changes to the approval rules themselves | Controller, with a second reviewer |
Approvers should be able to edit tone and wording in a draft. Amounts and bank details come from the ERP, so any change to them goes back to whoever owns that record.
How do you keep the approval queue moving?
An approval step only helps if drafts don’t sit in it. A few operating rules keep the queue short:
- Set response times. Same business day for email drafts is a common target, with write-backs cleared before the close.
- Sort by money and age. Larger and older balances go to the top of the queue.
- Batch only routine items. Reminders on small, undisputed invoices are safe to approve in bulk. Ledger items with any amount difference should be reviewed individually.
- Escalate on time. If the primary approver hasn’t acted by a set point, the item moves to a backup.
- Never send on expiry. A draft that runs past its response time should wait for a person to act.
- Capture the reason. Every rejection and substantive edit should carry a short reason code, such as wrong contact or incorrect amount.
Those reason codes are the most useful data the approval step produces. A cluster of “wrong contact” rejections for one segment points to stale contact data, and fixing it removes a whole category of review.
What happens if something changes while a draft is waiting?
Drafts can go out of date while they wait for approval. The platform should handle these cases automatically:
- A payment posts. The pending reminder should be cancelled and the approver told why.
- The customer replies. The draft should be rewritten from the new thread before anyone approves the old version.
- The customer raises a dispute. Follow-up should pause, and the draft should move to whoever owns the dispute.
- A reviewer edits the draft. Any send already scheduled from the earlier version should be cancelled.
Ask a vendor to show each of these during a demo. If the system can’t, a customer could get a reminder after they’ve already paid.
How are write-back approvals different?
Write-backs change the books, so they need a preview and a way back. Before anything posts to the ERP, the approver should see the proposed entry alongside the payment and the reason for the match. Once approved, the entry should be reversible with a counter-entry if a customer later disputes the allocation.
Credit memos and write-offs belong in a separate queue with controller sign-off, however much else is automated. Our guide to auto-applying one payment to multiple invoices covers confidence thresholds and sandbox write-backs for cash application in more detail.
When should you relax approval?
Relax approval one segment at a time, and let the review data decide. Track the share of drafts that reviewers edit or reject for each segment. When that share stays low for several weeks and collections on the segment hold steady, move it to automatic sending and keep measuring.
Overrides deserve a second look. If a reviewer changes the agent’s recommendation and the customer later pays exactly what the original draft said, record it. A segment where reviewers are often overruled by the outcome is a strong candidate for less review.
For email specifically, our post on whether AI can email your customers safely lays out the stages from observation through to sending within written rules.
When should automation be paused?
Agree on the triggers before go-live, so nobody has to debate them during an incident. Any of these should move every segment back to full approval:
- an email states a wrong amount or references the wrong customer
- a write-back doesn’t reconcile with the payment it was meant to apply
- a legal or regulatory complaint about collection messages arrives
- a review of approved drafts finds a rising share that were wrong
When a trigger fires, export the affected messages and entries for the controller, and fix the cause before relaxing approval again. Write-back errors are corrected with counter-entries that the controller approves.
How does Monk handle approvals?
Your team designs the playbooks Monk’s agents follow, including tone and when to escalate. You can require approval on every draft for a segment or let it send automatically. Either way, a deterministic check runs before anything reaches a customer, and anything uncertain is drafted for review instead of sent. Rules can apply to every customer or to one account, and you can see which rule the agent used on a given message.
In cash application, payments that match exactly and pass deterministic checks apply automatically. The AI layer only suggests matches, and none of them post until someone on your team confirms. Clean suggestions can be bulk approved when the amount matches exactly, while any over or short difference has to be opened individually. About 80% of payments match automatically, rising to 95% once your team accepts suggested rules.
Monk’s audit log records who changed each record and keeps actions taken by the agent separate from those taken by your team. Our guide to auditing what your AI collections agent sent covers how to review that activity once it’s live.
Where should you start?
List the actions that would be hardest to undo. Write-offs and new payment terms usually top the list, and those should need approval from day one. Then pick the least sensitive segment you have and approve every message for two weeks. By then, your reviewers’ edits will show which messages can send without approval.
To see how approval works on your receivables, book a demo.
Frequently Asked Questions
Does an approval step slow down collections?
Only if the queue backs up. With response times set and routine reminders batched, most drafts are approved the same day. The share needing review also falls as segments are relaxed.
What happens if an approver is on leave?
Items should move to a named backup before the response time runs out. Nothing should send because the primary approver was away.
How many approvers does a small AR team need?
Two is enough to start: a primary and a backup for each queue. The controller covers anything that reduces what a customer owes.
Do approvals need to happen inside the AR platform?
They should. Approving in email or chat breaks the audit trail, because the record of who approved what ends up outside the system.
Who should audit the approvals themselves?
Someone outside the AR team, often the controller or internal audit, should review a sample of approved items each month to check that approvals aren’t becoming automatic.
How long should approval records be kept?
Follow your document retention policy for financial records, and keep each draft together with its edits and the approver’s name.