/

/

The exception economy: Why one invoice in five sets the cost of all of them

The exception economy: Why one invoice in five sets the cost of all of them

The exception economy: Why one invoice in five sets the cost of all of them

Executive summary

Twenty years of accounts payable projects have taught me one thing about cost: the average invoice is cheap, the exception invoice is expensive, and almost every AP cost line is the second number wearing the first number's clothes.

Ardent Partners' 2026 benchmark puts the average fully loaded cost of processing an invoice at $9.90, against $2.67 for the top quintile, and reports that 19.9% of invoices still need manual exception handling before they can be paid (Ardent Partners, 2026a). Those two facts are the same.

The $7.23 gap between average and Best-in-Class is, overwhelmingly, the cost of the one invoice in five that stops.

This paper makes a practical argument. Exceptions are not a residue of imperfect automation that shrinks as capture improves. They are a distinct economic category with their own causes, cost curve, and resolution standard.


OCR and workflow tools were never designed to clear them because clearing an exception requires three things a workflow step cannot supply:

  1. A decision about what is actually wrong

  2. A fix that someone is prepared to stand behind

  3. A named owner who will act

The article introduces a five-class taxonomy of exceptions, a cost model that shows why the blended cost per invoice hides the problem, and the "decision, fix, owner" standard against which any serious exception program should be measured.

The final section demonstrates how an agentic approach, using Fractal's Cogentiq Invoice to Pay as the worked example, can meet this standard inside the client's own tenant and on top of the ERP and AP tools already in place.

The number everyone quotes, and nobody decomposes

Ask any AP leader what their cost per invoice is, and they will give you a number. Ask them what the cost of their worst 20% of invoices is, and the room goes quiet, because the metric the function reports is a blend, and blends are designed to be comfortable.

Ardent Partners has been publishing the uncomfortable version for more than a decade. Its 2026 "AP Metrics That Matter" research, drawn from 194 AP, finance, and procure-to-pay leaders, reports: an average of $9.90 and 11.6 days to process an invoice, with 38.3% of invoices moving straight through without a human touch (Ardent Partners, 2026a).

The previous year's State of ePayables put the average at $9.84 and 8.2 days (Ardent Partners, 2025). APQC's open-standards database, built from 5,846 organizations, puts the median at $6.00 (APQC, 2026). The spread between those sources itself is instructive: cost per invoice depends on what you count, and most organizations count the clean invoices and the stopped invoices as if they were the same product.

They are not the same product. In the Ardent 2026 research, Best-in-Class organizations link 85.1% of invoices to a purchase order, compared with 46.9% for everyone else, and spend 18.9% of staff time answering supplier questions, compared with 31.1% (Ardent Partners, 2026a). Those are not captured statistics. They describe how many invoices arrive with the context needed to clear them, and how much human attention goes to the ones that don't.

Average vs Best-in-Class scorecard

Key takeaway

The gap is widest on the measures that describe exceptions, not capture.

The problem: Exceptions are a separate economy

What is an exception?

An exception is any invoice that cannot proceed to payment without a human deciding something. Note the verb: deciding, as distinct from checking or reviewing.

The decisions include whether:

  • The price on the invoice or the price on the PO is right

  • A quantity shortfall is a short shipment or a split delivery

  • A missing PO number means a maverick purchase or a contract call-off the buyer forgot to raise

  • A supplier's new bank account is a legitimate change or a compromise attempt

Each of those decisions needs information that lives somewhere other than the invoice: in the contract, the goods receipt, the buyer's email, the supplier master, last month's history for that supplier.

The AP analyst's job on an exception is to go and get that information, form a view, and either act or persuade someone else to act. That work is slow because it is investigative, and it is expensive because it is interrupted. IOFM's practitioner research has long put the handling time of a single exception at a quarter to three-quarters of an hour of concentrated effort (IOFM, 2024), and anyone who has run an AP shared service knows that figure understates the elapsed time, because the analyst is waiting on a buyer in another time zone for most of it.

Why the blended number hides it

Here is the arithmetic that most AP dashboards refuse to show.

Take Ardent's 2026 figures at face value: 19.9% of invoices are exceptions and the blended average cost is $9.90. If the 80.1% of clean invoices cost roughly what the Best-in-Class group pays for everything, about $2.67, then the exceptions must cost somewhere in the region of $39 each to produce that blend.


This is an illustration of the structure, not a published benchmark, and any CFO can rerun it with their own clean-invoice cost. The point survives any reasonable set of inputs: the average organization is paying a small amount for most of its invoices, and a very large amount for a minority, and the dashboard reports the weighted mean.

Illustrative data
Illustrative data
Illustrative data


Key takeaway

The exception minority carries most of the cost.


The consequence for investment decisions is severe. A capture or OCR upgrade promises to lower the blended number by making the clean 80% cheaper still. It does almost nothing to the 20% that is driving the cost, because an exception is not an extraction error.

Ardent's 2026 research finds that 58% of AP organizations are using or piloting AI, and that the dominant use is invoice capture and data extraction (Ardent Partners, 2026a).


That is the first place AI went in AP, and it is the place with the least cost left to remove.

Where the time actually goes

Supplier inquiries are the tell. Ardent reports AP staff spending 21.9% of their time dealing with suppliers in its 2025 research and 31.1% for non-Best-in-Class organizations in 2026 (Ardent Partners, 2025; 2026a). Most of those inquiries are the downstream shadow of an exception: a supplier asking where their payment is because their invoice is sitting in a queue waiting for a buyer to confirm a price. The exception generates the inquiry, the inquiry consumes the analyst, and the cycle time stretches to the 11.6-day average that Ardent reports. One stopped invoice creates work in three places.

The Institute of Financial Operations and Leadership's 2026 research with SAP Concur adds a sobering context point: only 19% of organizations use AI anywhere in AP today, with a further 30% planning to adopt within 12 months (IFOL and SAP Concur, 2026).


Half the market has not started, and the half that has started mostly started on capture.

The framework: Five classes of exception and one standard of resolution

Over many deployments, I have found that every exception falls into one of five classes, and that each class needs a different kind of context to clear. The classes matter because they tell you what the resolving system, human or agent, must be able to see.

Class

What stops the invoice

Context needed to decide

Typical share of volume (illustrative)

Data

Missing or malformed fields: no PO reference, wrong entity, bad tax ID, unreadable line items

Supplier master, entity structure, prior invoices from the same supplier

High volume, low value

Match

Price, quantity or unit mismatch against PO or receipt; partial deliveries; tolerance breaches

Contract price list, goods receipt history, buyer intent, split-shipment patterns

The largest cost driver

Policy

No PO where one is required, approval limit exceeded, non-contract supplier, spend category rules

Delegation of authority, procurement policy, contract coverage

Medium volume, high friction

Supplier

Bank-detail change, new remit-to address, duplicate supplier record, inactive vendor

Verification records, change history, external registries

Low volume, highest risk

Control

Suspected duplicate, segregation-of-duties conflict, sanction or fraud signal, audit hold

Cross-entity payment history, SoD matrix, fraud model output

Low volume, non-negotiable


Two observations from the field.

  1. The Match class is where the money is. It is the largest share of handling time in every AP operation I have benchmarked, and it is also the class that pure capture investment cannot touch, because the invoice was read perfectly; it simply disagrees with the PO.

  2. The Supplier and Control classes are small, but they are the classes that end careers. The AFP reports that 76% of US organizations experienced attempted or actual payments fraud in 2025 and that business email compromise affected 74% (AFP, 2026). A bank-detail change that is treated as a routine data exception rather than a supplier-class exception with its own verification step is how a $2 million payment leaves the building.

The resolution standard: Decision, fix, owner

Here is the test which can apply to any system that claims to handle exceptions. For each stopped invoice, can it produce:

  1. A decision. A specific statement of what is wrong and what should happen: Invoice price $4.12 against contract price $3.98 on item 1172; contract price applies; propose debit note for $1,540.

  2. A fix. The action that resolves it, prepared and ready to execute: the debit note drafted, the PO amendment raised, the GL recode applied, and the supplier notification written.

  3. An owner. The named person or role who must confirm the decision if it exceeds the system's authority, with the context attached so they can confirm in one minute rather than investigate for 20.

Workflow tools fail this test by design. A workflow routes the exception to a queue and waits. It can tell you an invoice has a price variance; it cannot tell you which price is right, because knowing that requires reading the contract, checking what was paid last quarter, and understanding that this supplier always invoices in cases while the PO is in units.

That is a decision, and decisions were left to people because software could not make them. That constraint has now changed.

The agentic approach: How an agent clears an exception

An agent is software that can form a view from several sources, propose an action, execute it within a mandate, and learn from whether the action was accepted. Forrester's 2026 survey of agentic AI use cases in AP names exception handling as one of six categories where agents are already delivering measurable value, alongside capture, matching, supplier communication, reporting, and fraud detection (Forrester, 2026; cited via Payhawk).

Ardent's 2026 trend research goes further and recommends that organizations pilot agentic AI first on exception management and forecasting, precisely because those are the areas where mature processes and clean data give an agent something to reason over (Ardent Partners, 2026b).

The following example traces how a Match-class exception is handled using Fractal's Cogentiq Invoice to Pay. While the example is product-specific, the underlying principles apply to any agent designed for exception resolution.

A price variance, resolved

A beverage company's co-packer invoices 2,400 cases of a finished product at $4.12 a case. The PO was raised at $3.98. The tolerance is 2%. The invoice stops.

In the conventional operation, the invoice lands in an analyst's queue in the shared service center. The analyst opens the PO, sees the variance, emails the buyer, waits two days, gets a reply that the supplier raised prices in July, asks procurement whether that increase was agreed, waits again, and eventually recodes or raises a debit note. Elapsed time, eight to 12 days. Analyst time, perhaps 40 minutes across five sessions. Supplier calls in twice.

In the agentic operation, the Matching and Validation agent reads the variance and does what the analyst would have done, in sequence and in seconds. It pulls the contract price schedule and finds that a 3.5% increase was executed in July with an effective date after this delivery. It checks the last six invoices from this supplier and finds two paid at the new price without challenge. It checks the goods receipt and confirms 2,400 cases were received. It forms a decision: the new price is contractually valid but was applied one delivery early; the correct price for this delivery is $3.98; a debit note for $336 is due. It drafts the debit note and the supplier message. Because the value is under its mandate, it executes, posts the invoice at the correct amount, and logs the reasoning. Because the pattern (supplier applying increases early) has now appeared twice, it raises a flag to the category buyer as a control observation rather than an exception.

If the value had exceeded the agent's mandate, the owner, in this case the category buyer, receives the decision, the evidence, and the prepared fix in one message, and confirms or overrides. That confirmation takes a minute, and the agent records whether its proposal was accepted so that its next proposal is better calibrated.

One exception. Two operating models.


Key takeaway

The agent compresses investigation and decision to seconds and reserves the human for confirmation.

What makes this work in practice

Three design choices determine whether an exception agent clears invoices or just produces nicer alerts.

  1. Context before autonomy. The agent in the example could decide because it could read the contract, the receipt history, the supplier's prior invoices, and the tolerance policy. That context has to be assembled before the agent is turned on. At Fractal, we describe this as a layered context architecture for production agents: policy, master data, transaction history, tolerances and controls, each made available to the agent in a governed way. An agent without that context is an OCR tool with a conversational interface.

  2. Mandate, not permission. The agent executes within a stated authority, by value, exception class, supplier segment, and entity, and escalates above it. The controller sets that mandate, versions it, and makes it auditable. It is the mechanism that lets a finance leader say yes to autonomy without saying yes to everything.

  3. Resolution that changes the upstream. The second appearance of the early-price pattern became a control observation, not another exception. This is the part most exception programs miss. Exceptions are symptoms; the cause is usually upstream in procurement, the supplier's billing process, or master data. An agent that only clears exceptions is a faster analyst. An agent that reports causes is a process-improvement engine. In Ardent's own framing, the move is from flagging errors to scrutinizing discrepancies and communicating with suppliers for resolution, with people intervening only where judgment is needed (Ardent Partners, 2026b).

Where it runs matters

The context described above is sensitive: contract prices, bank details, payment history. In regulated industries and in global capability centers serving a parent in another jurisdiction, where the agent runs and where the data goes is a control question, not an IT preference.

Cogentiq Invoice to Pay runs inside the client's own cloud tenant, on the client's data platform, with the client's choice of models; the invoice and the payment data never leave the estate. It orchestrates on top of the ERP and whatever AP or procurement suite is already in place rather than replacing them.

For the exception problem, the relevance is simple. The agent can only decide on what it can see, and clients will only let it see what stays inside their walls.

Benefits and proof points

The value of clearing exceptions properly shows up in three places, and a finance leader should insist on measuring all three.

Where value appears

Measure

What changes

Cost

Cost per exception; cost per invoice decomposed into clean and exception

Exception handling time falls from tens of minutes to a confirmation or nothing; the blended number follows

Cycle time and supplier experience

Days to pay; share of staff time on supplier inquiries

Exceptions no longer wait on hand-offs; inquiries fall because payments are not late

Leakage and control

Debit notes raised; duplicate and overpayments prevented; control observations fed to procurement

Pricing and quantity leakage recovered systematically rather than when an analyst happens to notice


When exception handling is addressed directly, clean-invoice cost barely moves, and blended cost per invoice falls sharply because the expensive minority has been dealt with. Any vendor who shows you a falling blended number without showing you the exception rate and the exception cost alongside it is showing you a capture improvement.

Exception metrics move first. Blended cost follows.

Lower exception rate

Less investigation effort

Fewer supplier inquiries

Faster invoice resolution

Lower blended cost per invoice

Conclusion

Accounts payable has spent a decade automating the invoices that were never the problem. The 19.9% that stop is where the $7.23 gap lives, where supplier calls come from, where fraud attempts enter, and where the 11.6 days are spent. Clearing them requires a decision, a fix, and an owner for every stopped invoice, and for 20 years only people could provide those. Agents can now provide them inside a controller's mandate, on the client's own infrastructure, and feed the causes back to procurement so that next month's exception rate is lower than this month's.

If you run AP for a large organization or a capability center, the next step is not another capture benchmark. Take one month of exceptions, classify them against the five classes above, and cost them. Then bring that file to a working session with Fractal: we will run Cogentiq Invoice to Pay's matching and controls agents against a sample of your own stopped invoices, in your own tenant, and show you what each one would have decided. That session is the entry point to a four-week pilot on live invoices, with a measured baseline and a go or no-go at the end.

References:

Turn AP exceptions into decisions

See how Cogentiq Invoice to Pay resolves exceptions using your own business context

Author

Mayank Ahuja

Prathmesh Thergaonkar

Global Director, Finance Analytics

Recognition and achievements

Select Fractal accolades

Leader

The Forrester Wave: Customer Analytics Services Q2, 2025

Representative vendor

Gartner Hype Cycle for Consumer Goods, 2026

Great Place to Work

Great Place to Workยฎ across four regions: India (9th year), USA (5th year), UK (5th year) and UAE (2nd year)

Recognition and achievements

Select Fractal accolades

Leader

The Forrester Wave: Customer Analytics Services Q2, 2025

Representative vendor

Gartner Hype Cycle for Consumer Goods, 2026

Great Place to Work

Great Place to Workยฎ across four regions: India (9th year), USA (5th year), UK (5th year) and UAE (2nd year)