How AI agents handle invoice exceptions, and when they don't

Every AP automation vendor promises to reduce manual processing. The ones that have actually deployed at scale know that the hard part is not the standard invoices. It is the exceptions.

A standard invoice that matches the purchase order, has the correct supplier details, and falls within the approval threshold processes automatically in any reasonable AP system. That is the easy 60-70% of your invoice volume. The difficult 30-40% is the exception pile: invoices with price discrepancies, missing PO references, split deliveries, payment term disputes, and the long tail of one-off situations that did not fit the standard template.

How your automation handles exceptions determines whether you are actually reducing manual work or just relocating it.

Why exceptions break rule-based AP automation

RPA and first-generation AP automation tools handle exceptions the same way: they route them to a human queue and stop.

This is not a failure of implementation. It is a fundamental constraint of rule-based automation. A rule can say "if the invoice amount is within 2% of the PO amount, approve automatically." It cannot say "this invoice is 8% over the PO amount because the supplier applied a fuel surcharge that was verbally agreed with the procurement team last month, so it should probably be approved."

The second scenario requires understanding context, making a judgement, and potentially checking another data source. Rule-based systems cannot do that. So they escalate. And the exception queue fills up.

In most organisations with mature RPA deployments, the exception queue is where human AP effort is concentrated. The easy invoices are automated. The human team processes the exceptions. If exceptions are 30% of volume, the human team is handling 30% of the work. If exceptions grow (new suppliers, new product categories, more complex contracts), the human workload grows proportionally.

What AI agents do with exceptions

An AI agent in accounts payable does not route every exception to a human queue. It reasons about the exception first.

The reasoning process looks like this: the AI reads the invoice, identifies the discrepancy, checks it against the purchase order, looks for relevant context (is this supplier known for surcharges? is there a note on the PO about variable pricing?), assesses the magnitude of the discrepancy against the approval policy, and makes a decision about whether to auto-approve, request clarification from the supplier, route to the appropriate approver, or flag for senior review.

This is different from a rule-based system because the AI is not pattern-matching against a fixed set of conditions. It is reasoning about the specific situation.

Freeday's deployment at Woonbron, the Dutch housing corporation, processes 35,000 invoices annually at approximately 80% end-to-end automation. That 80% includes invoices that would have been classified as exceptions under a rule-based system but that the AI can handle because the discrepancy is within a range it can reason about. The Freeday finance agent page describes the exception handling logic in more detail.

The exception types AI agents handle well

Price discrepancies within a tolerance range. If a supplier invoices EUR 1,050 against a PO of EUR 1,000, the AI can check whether this is within the tolerance policy, whether the supplier has a history of small rounding differences, and whether the amount is within the automatic approval threshold. Most of these cases are approvable without human involvement.

Missing PO references. Rather than routing every invoice without a PO reference to the exception queue, the AI can attempt to match the invoice to an open PO based on supplier, amount, date range, and line item description. For partial matches, it can present the likely match to an approver with a confidence score rather than routing it as unmatched without context.

Duplicate invoice detection. The AI can cross-reference incoming invoices against the invoice history for known patterns: same supplier, same amount, close date range. Potential duplicates are flagged with the specific matching evidence rather than processed or silently rejected.

Currency and tax discrepancies. Invoices from international suppliers frequently have rounding differences in currency conversion or varying tax treatments. The AI can reason about whether a EUR 0.50 discrepancy on a EUR 10,000 invoice is a genuine dispute requiring supplier contact or a rounding artefact that can be posted and closed.

The exception types that still need humans

Honest about this: there are invoice exceptions that AI agents should not handle autonomously, and good deployment design makes this explicit.

Commercial disputes. When a supplier has invoiced for work that was not delivered or that the organisation is disputing, this is a commercial relationship issue. The AI can identify the discrepancy and prepare a summary, but the decision to accept, negotiate, or dispute requires human judgement and relationship context.

Contract interpretation ambiguities. When the contract terms are ambiguous about what is included in the scope, interpreting them requires legal and commercial input that is outside the AI's authority.

Fraud indicators. Invoices with fraud indicators (new bank account details, unusual supplier behaviour, amounts just below approval thresholds) should always route to a human with a clear flag. The AI can identify these patterns, but the decision should not be automated.

Senior approval thresholds. High-value invoices above the defined senior approval threshold should always involve a human sign-off, regardless of whether the AI would approve them. This is a governance requirement, not a capability limitation.

The design principle is that the AI handles everything within a defined scope and escalates everything outside it with context. A well-designed AP deployment has an explicit escalation policy that defines these boundaries before go-live.

The operational impact: what changes for the AP team

When AI agents handle the automatable exceptions, the nature of the AP team's work changes.

Instead of processing a mix of standard invoices and exceptions, the team handles only the cases that genuinely require human judgement: commercial disputes, high-value approvals, fraud investigations, and relationship-sensitive situations. The volume they touch is lower. The complexity of what they touch is higher.

Pathe's deployment processes 50,000 invoices annually at 75% automation, freeing 2.5 FTE. Those 2.5 FTE were not eliminated: they were redirected to supplier relationship management, contract negotiation support, and the genuinely complex exception cases that require experience and judgement.

This is the operational shift that matters more than the automation rate itself. An AP team that spends its time on supplier relationships and complex cases is a more strategically valuable team than one that spends it processing a queue.

What a 30-day exception analysis tells you

Before implementing AP automation, the most useful diagnostic exercise is a 30-day exception analysis: categorise every invoice that required manual intervention over the past month.

The typical distribution in a Dutch enterprise AP function looks like this: price discrepancies and tolerance issues (35-40% of exceptions), missing or incorrect PO references (20-25%), duplicate invoice risk flags (15-20%), and genuinely complex cases requiring judgement (15-25%).

The first three categories are automatable. The last category is not, at least not fully. If your exception analysis shows that 75-85% of your exception volume falls in automatable categories, you have a strong AI automation business case. If the genuinely complex cases are 40% or more of your exceptions, the business case still exists but the expected automation rate is lower.

The Freeday accounts payable solution page covers how this analysis feeds into deployment scope definition.

FAQ

What percentage of invoice exceptions can AI handle automatically?

Based on Freeday's deployment data and typical Dutch enterprise AP exception distributions, 70-80% of exception volume is automatable. Price tolerance discrepancies, missing PO references, and duplicate detection are the highest-volume automatable exception types.

What happens when the AI makes an incorrect exception decision?

Well-designed deployments include audit trail logging for every AI exception decision and regular sampling of automated decisions for quality review. Incorrect decisions are caught either through supplier communication (if an incorrect approval is disputed) or through the review sampling process, and the AI's exception handling is updated accordingly.

Does AP exception automation work with any ERP system?

Freeday integrates with the major ERP systems used in the Dutch market including SAP, AFAS, and Microsoft Dynamics. Integration with the ERP's PO and contract data is what enables the AI's exception reasoning to be accurate.

How do you prevent AI agents from approving fraudulent invoices?

Fraud indicator patterns (new bank details, amounts clustering just below approval thresholds, unusual supplier behaviour) are defined as mandatory human review triggers. The AI flags these cases and does not proceed autonomously regardless of the amount or apparent legitimacy.

What is the right approval threshold for AI autonomous decisions?

This is an organisational policy decision, not a technology question. Most organisations start conservatively, with AI autonomous approval limited to amounts under EUR 500 or EUR 1,000, and expand the threshold as confidence in the deployment builds. The threshold should reflect both risk appetite and the quality of the AI's exception reasoning.