There is a number that almost every vendor deck in this market puts on slide three: 99% extraction accuracy. It is true, and it is beside the point.
True, because the leap really happened. Classic OCR plateaus near 60% once it meets real documents — skewed scans, faxes from small suppliers, PDFs spat out by a decade-old printer driver. Intelligent document processing, with language models behind it, reaches 95% to 99% on the same files. Anyone who lived through the era of one template per supplier knows how big that gap is.
Beside the point, because touchless invoice processing solved the easy half.
Reading is finished. Deciding never started
Gartner has already changed the criterion by which it assesses this market. It no longer ranks intelligent document processing primarily on how accurately a tool reads text, because differentiation has moved up towards orchestration and the ability to act on what was extracted. Boris Evelson, VP and Principal Analyst at Forrester, put it more bluntly: generative and agentic AI is becoming “an equaliser” that challenges vendors’ ability to differentiate at all.
When everyone reads well, reading well sells nothing.
What remains standing is the next step, and that one never appears in a demonstration. A read invoice is a set of fields. A qualified invoice is a decision already taken: which VAT treatment applies, whether the reverse charge kicks in, which ledger account it belongs to, which cost centre, which project, which funding line, who may approve that amount for that supplier. And finally, whether the thing can proceed to payment without human eyes on it.
The place where that decision materialises is the ERP. That is where it goes right or wrong, and it is the place nobody inspects when buying a reading engine.
Forty codes, none of them fit
A concrete case, from a tax configuration catalogue we reviewed inside an international ERP. Forty reverse-charge VAT codes available. Every one for intra-community transactions. Not one for imports from outside the European Union.
This is not a software defect. It is the fossil record of a configuration decision taken when the company bought within Europe, left in place after the buying pattern changed. The invoice was read correctly. Every field came out right. The classification that followed was wrong because the rule did not exist — and no extraction engine, however good, detects the absence of a rule.
The trap is not exotic, and nobody is to blame for it. Reverse charge applies across construction, scrap metal and metallic waste, forestry products, emission allowances, intra-community services and purchases from non-resident suppliers, with the list and the required wording differing by member state. In each case the invoice arrives with no VAT charged and the buyer does what the supplier did not.
Then the rules move. Portugal cut construction VAT from 23% to 6% on qualifying works in January 2026. An organisation whose construction coding was correct in December was wrong in January — not through incompetence, but because the law changed mid-process. Every jurisdiction has its own version of that sentence.
What the numbers say, and who they apply to
Ardent Partners’ Accounts Payable Metrics That Matter is worth stopping on, because it is one of the few data sets that separates the cost of reading from the cost of deciding.
Best-in-class organisations process an invoice for roughly €2.56 in 3.1 days. The market average spends €10.02 and 10.9 days. Four times the cost, three times the elapsed time, same document. The interesting part is where that gap sits: the touchless rate — invoices reaching payment with nobody involved — is 32.6% at market average and 49.2% best-in-class, meaning that even in the best organisations half the invoices stop and wait for a person. The exception rate runs at 14% against 9%, with each exception costing three to five times more than an invoice that flows straight through. Some estimates put it at five to ten times, and I am sceptical of the upper bound: it depends far too much on who handles the exception and what they earn.
The Institute of Finance & Management supplies the piece that closes the argument: 84% of an accounts payable professional’s time goes to manual tasks, and top departments handle 6,900 invoices per person per year against 4,200 at the average.
Now the uncomfortable arithmetic. If exceptions are 14% of volume and cost five times more, they consume a share of total cost far above their weight. The expensive work is concentrated in the minority of documents that automatic extraction does not fix — and it does not fix them because their problem is not reading. It is judgement.
One honest caveat, because this sum does not work for everybody. In a company receiving two hundred invoices a month from thirty familiar suppliers against a short chart of accounts, this is a theoretical problem: two experienced people solve it better and more cheaply than any system. The maths starts to turn somewhere around a thousand invoices a month, or earlier where there are construction sites, grant-funded projects or purchases from outside the European Union. Software vendors rarely say this. They should.
The calendar that removes the slack
Here is where most commentary gets it wrong, so precision matters.
The ViDA package was adopted on 11 March 2025 — Directive EU 2025/516, Regulation EU 2025/517 and Implementing Regulation EU 2025/518 — published in the Official Journal on 25 March 2025 and in force since 14 April 2025. What that entry into force achieved was the removal of the derogation requirement: member states may now mandate electronic invoicing on their own authority. The Digital Reporting Requirements themselves only bite on 1 July 2030 for intra-EU B2B, with domestic system alignment on 1 January 2035.
So ViDA obliges nobody in 2026 or 2027. It authorised the wave of national mandates now landing, every one of them built on the EN 16931 semantic standard.
That wave is what matters operationally. Belgium mandated B2B electronic invoicing through Peppol from 1 January 2026. France requires every business to be able to receive from 1 September 2026, with large and mid-sized companies issuing from the same date and smaller ones a year later. Spain phases Verifactu across 2026 and 2027. Germany has required receipt since January 2025, with issuing obligations from 2027 above €800,000 turnover.
The practical effect is the same everywhere, and it is not a new classification obligation. It is the loss of the slack that currently allows teams to fix by hand what goes wrong. A process surviving today on 14% exceptions handled by three people who know the business by heart will get the same percentage on a much larger base — and it will still be the same three people.
Where the rule should live
The answer is not better reading. It is taking qualification out of the head of whoever has been there fifteen years and putting it somewhere it can be versioned, tested and proven.
In practice that means three unglamorous things: tax and accounting classification rules held as configuration; a record of who decided what and when, with a complete audit trail; and the ability to state, when construction VAT changes in January, exactly which documents were processed under the old rule and which under the new one.
The returns are measured, and they are not trivial. McKinsey reports 25% to 40% reductions in processing cost within 18 months for finance teams that automate with discipline. There is also an effect that almost never reaches the business case: cutting cycle time from the sector average of around ten days to under three captures, on average, 1.8% of annual revenue in early payment discounts. That is not cost saving. That is margin.
And one figure separates the two generations of automation better than any argument of ours: organisations running OCR plus basic matching sit at 7.2 days average cycle time, whilst those adding intelligent exception handling drop to 3.1 days. The reading engine is identical in both. The entire difference lies in what happens when the invoice does not fit.
Four questions for the next vendor meeting
If you are evaluating tools right now, swap “what is your extraction accuracy?” for these.
What share of our invoices reaches payment with no human touch — not how many were read correctly.
What is our exception rate, and what causes it — if most exceptions are tax classification rather than misreads, a better engine corrects nothing.
What does an exception cost us — in senior people’s hours, not in licence fees.
Where do the qualification rules live — in a system that versions them, or in the experience of whoever applies them daily.
Deloitte’s 2025 CFO survey found 49% of finance leaders prioritising the automation of routine work to free people for higher-value tasks. In accounts payable, the higher-value work is deciding the hard cases. It follows that the system should hand those cases over already framed, rather than pushing all of them into one queue at the same priority.
Reading the invoice is done. Qualifying it is what still pays salaries.
Jorge Gamito Pereira
CEO | Uniksystem