How AI Invoice Processing Handles the Invoices It Cannot Read

A well-built AI invoice system does not try to read every invoice. It reads the clear ones, flags the unclear ones and sends those to a person with the reasons attached. The design of that exception path matters more than the headline accuracy figure on a sales page.

When owners ask whether AI can process their invoices, the real question is what happens on the bad days. Every finance team knows the awkward cases. A system that handles them badly creates more work than it saves, and a system that handles them well earns trust quickly.

Why do some invoices fail?

Usually for ordinary reasons. A scan that is skewed or low resolution. A supplier who changed their layout without telling anyone. A handwritten note across the total. A missing purchase order number. Two invoices combined into one PDF. A credit note that looks like an invoice. None of these is exotic, and all of them happen every week.

A small business also has a long tail of suppliers who send one invoice a year in a format the system has never seen. The system will meet unfamiliar documents constantly, so unfamiliarity has to be a normal case and not an error condition.

What does the system do with them?

It scores its own confidence on each field. If the supplier name, invoice date, total and tax figure all pass their checks, the invoice moves on. If any field is doubtful, the whole invoice goes to a review queue instead of being passed through with a guess.

The checks that matter most are simple ones:

• Do the line items add up to the stated total?
• Does the tax figure match a rate you would expect for this supplier?
• Is the supplier one you already know, with a matching name and address?
• Has this invoice number been seen before, which might mean a duplicate?
• Does the amount fall within the usual range for this supplier?

These are arithmetic and lookup checks, and they catch many problems with no AI involved.

What should the reviewer see?

Not just the document. They should see the original side by side with what the system extracted, with the specific field it was unsure about highlighted and a short reason: "total does not match line items", "supplier not recognised". A reviewer who has to hunt for the problem will stop trusting the queue and start rubber-stamping or avoiding it. A reviewer who can resolve most items in thirty seconds will keep using it.

What should it never do?

Pay an invoice. In an SME setup I would keep approval and payment as a human step. The system prepares the entry and flags anything unusual, and a person approves it.

The case I worry about most is changed bank details. An invoice from a known supplier arrives with a new account number. That is a classic fraud pattern, and it deserves a person's attention and a phone call to a known number every time. A good system treats "known supplier, new bank details" as a mandatory flag, never as a routine update.

What about accuracy figures?

Be cautious with a single percentage. "98 per cent accurate" can mean 98 per cent of fields across clean test documents, which says little about your messy real-world inbox. Ask instead:

• What share of invoices go straight through without human touch?
• Of those, how many contained an error that nobody caught?
• How long does the average flagged invoice sit in the queue?

The second question is the important one. Errors in the flagged pile are visible and fixable. Errors in the straight-through pile are the ones that cost you.

How do you improve it over time?

Log every correction. Patterns show up quickly. If one supplier causes half the exceptions, a template or a conversation with that supplier fixes it. If the system is repeatedly wrong on one field, that is a signal to adjust the setup, not to ask staff to work around it.

Review the log monthly for the first quarter. Most of the improvement happens in that window, after which the correction rate tends to settle.

What is a realistic target?

Not a perfect score. A realistic aim is that most clear invoices go straight through, the rest reach a person quickly with a clear reason attached, and nothing is paid without a human check. If your team spends less time keying data and more time resolving the genuinely odd cases, the system is doing its job.