Ask an AP manager what eats the team's week and the answer is rarely the invoices that match. It's the exceptions: the invoice with no PO number, the quantity that doesn't agree with the receipt, the unit price that drifted 3% above contract. Most finance automation tools now say they handle these. The useful difference between them is how a person gets pulled in, and what that person sees when they are.
What finance automation software supports human review for exceptions?
Tools that route finance exceptions to a human
Tool | What sends an exception to a person | How the reviewer works it | Source |
|---|---|---|---|
Stampli | Approval routing by amount, department, cost center and vendor, with enforced authority limits; unreviewed invoices escalate automatically | Approval and discussion happen on the invoice itself, so the context is kept as evidence | |
Tipalti | Configurable exception approval rules on PO matching; exceptions above a threshold go for approval | Buyers approve or reject by email in one click; AP can dispute the bill or update the PO in the ERP | |
Medius | Workflows that adapt to business rules, invoice attributes, supplier behavior and exception conditions | Exceptions are resolved inside the workflow; approvals follow policy and decisions are kept in audit-ready records | |
HighRadius | Agentic AI auto-resolves pricing, tax and missing-item exceptions; the rest go to exception resolution workflows | Workflows coordinate the stakeholders needed to clear an exception; HighRadius claims 90% touchless processing | |
Vic.ai | Confidence thresholds you define; low-confidence predictions go to a person | Reviewer edits or confirms, and the AI learns from the change | |
Automation Anywhere | Validation against POs and receipts, plus duplicate flags | Exceptions are routed with guided resolution in natural language | |
Zamp | Approval gates set by entity, amount, account, region or exception type | Escalation arrives with the source documents, findings, attempted actions and a specific question; approved corrections update the operating procedure |
Vendor claims above are each vendor's own. Test them against your own exception mix before you rely on them.
For a wider comparison of finance platforms beyond AP, see best AI platforms for end-to-end finance operations. For where invoice exceptions land during the close, see month-end close automation tools.
What good exception review looks like
A review step only helps if the reviewer can make a decision quickly and the decision is recorded. Four things separate a useful exception queue from a second inbox.
- Routing by exception type. A missing receipt belongs with receiving, a price variance with the buyer who owns the vendor, and a missing PO with the business unit that placed the order. A single AP queue for everything is where invoices go to wait.
- Evidence attached to the request. The reviewer should see the invoice, the PO, the receipt, the contract rate and the specific mismatch in one place, plus a clear question to answer.
- Approval thresholds set in advance. Decide which variances clear automatically and which need a person, by amount, entity or exception type, before the tool goes live.
- An audit trail of the decision. Record the input, the sources checked, the action, who approved it and any correction, so a controller or auditor can reconstruct the case later.
For a broader view of where human judgment still belongs in automated work, read human-in-the-loop AI.
Why AP exceptions happen
Most teams file exceptions under "data quality," which is too vague to fix. In practice they cluster around four causes.
Three-way match failures
This is the most common one. The invoice quantity doesn't match the purchase order, or the receipt hasn't been entered yet. Small price deviations, partial shipments, backorders and unit-of-measure mismatches (boxes versus single units) all raise the same flag: the numbers don't reconcile.
Missing or malformed reference data
Vendors submit invoices without PO numbers, with the wrong PO number, or with vendor codes that don't exist in the ERP. This is common with smaller or international suppliers who use different invoicing systems. The invoice is legitimate. It just can't be matched because the linking data is wrong or missing.
Duplicate submissions
The same invoice arrives twice, sometimes with a modified invoice number. Weak detection turns this into real money at volume. See how AI agents detect duplicate payments for the patterns to watch.
Approval workflow stalls
Invoices that need departmental sign-off sit in inboxes. Some wait for weeks when an approver is on leave, the right contact has changed, or there's no escalation rule.
How a digital employee works the exception queue
Zamp handles exceptions as a tiered pipeline, each stage applied to a smaller set of invoices.
Stage 1: intake and extraction
Every invoice enters the same intake path, whether it arrives as an email attachment, a portal submission, an EDI feed or a scan. The AI employee extracts header fields and line items, checks the vendor against the vendor master, checks the PO number format against ERP conventions, and confirms the invoice date falls inside the submission window. Problems caught here are cheap to fix. A missing PO number can often be inferred from the vendor and the approximate amount.
Stage 2: matching against PO, receipt and contract
Invoices that pass intake go to three-way matching. The AI employee pulls the PO and goods receipt from the ERP, whether that's Oracle Fusion, SAP S/4HANA or Coupa, through an API where one exists or through the ERP's own screens where it doesn't. It applies the tolerances the finance team has set. For example, a team might clear a unit-price variance under 2% automatically and send anything above 15% to review. Partial shipments clear when the receipt confirms the delivered quantity matches the invoice. Contract rates come from the procurement system.
The result is three groups: invoices that go straight through, minor exceptions resolved inside policy, and real exceptions that need someone.
Stage 3: classification and routing
Real exceptions are classified by type, because a missing receipt and a pricing dispute need different people.
- Missing receipt: receiving gets a notification with the original PO and expected delivery date. Once the receipt is confirmed, the invoice resumes without AP touching it.
- Price variance above tolerance: the procurement contact who owns the vendor gets a task with the contract rate, the invoiced rate and the variance already filled in.
- Missing PO reference: the business unit that placed the order, identified from vendor and spend category, gets a task to link the invoice to the right PO.
- Suspected duplicate: the invoice is held for human review, with the possible duplicate shown side by side.
Stage 4: escalation and SLAs
Exceptions that sit past a set window escalate. A team might escalate an unconfirmed receipt to the department manager after 48 hours, and an untouched pricing dispute to the category manager after five days. These are tracked tasks with an audit trail, so the AP manager sees where every exception sits. The job shifts from chasing individual invoices to reviewing a list of SLA breaches.
How to measure the result
Use your own baseline, not a vendor's modeled savings. The measures that show whether exception handling improved are work completed, time to completion, straight-through rate, exceptions resolved, corrections, approval time, records at risk and human minutes per case. Count them before go-live and again after.
The AP team's work changes too. Time that went into chasing exceptions moves to vendor relationships, early-payment discounts and working capital analysis.
The integration question
The usual objection is ERP integration. "Our Oracle instance is too customized," or "SAP won't give us the API access we need." These are fair concerns, and they're less of a blocker than they used to be.
Zamp's AI employees can work through ERP screens the way a person would, without API access. They can read PO data from an Oracle screen, enter invoice data into a Coupa form, and post a journal entry in SAP. This is slower than an API call and sensitive to UI changes. Where full API integration isn't feasible soon, it's a workable path that doesn't need a multi-year IT project.
Where to start
You don't need to automate every exception type at once. Start with the highest-volume, most predictable categories, usually missing receipts and small price variances, and expand from there. When a reviewer corrects a decision, that correction is added to the AI employee's operating procedure, so the same exception isn't flagged the same way next time. See how Zamp's AI employees learn from corrections.
The AP team stays. Their effort moves from exception chasing to spotting recurring vendor discrepancies, flagging contracts where pricing drifts, and using cleaner data in procurement decisions. If your team spends more than a few hours a week on invoice exceptions, talk to the Zamp team about your current volume.
