OCR
Accuracy and confidence score: certainty is checked field by field
Invoice OCR accuracy (optical character recognition) cannot be read off an overall rate. What matters is certainty on the field the payment depends on, document by document: a score per field, the exact position of the value in the page, a review threshold you set yourself.
The question, badly put
Why an overall rate means nothing
The market displays percentages. They are reassuring and they tell you nothing, because an overall rate mixes the easy fields with the fields that decide the payment. It is the same confusion set out in why OCR is not enough: reading is not checking.
Take a tool that reads invoice numbers, dates and company names perfectly. Those fields are numerous, regular, and they pull the average up. The same tool can go wrong regularly on the unit prices of the lines at the bottom of the page, where the print is fainter and the layout tighter. The rate on display stays flattering; the field the payment depends on is wrong.
The reasoning holds without a single figure from us. Take a rate of correct fields, as high as you like, and apply it to an invoice of forty lines each carrying six fields. That is two hundred and forty fields, of which a residual share is wrong: several errors per document. A rate of correct fields says nothing about the number of correct documents.
There is a substantive reason for that reticence. A rate only becomes information if you know the set of documents measured, the definition of a correct field, the share of degraded documents in the sample and the way empty fields are counted. Two honest tools can display very different figures on the same corpus, simply because they are not counting the same thing.
An honest measurement would require a set of real documents, a published method and a declared scope. That is a project, not a line of copy. So we publish no rate: we publish the means of verifying every value.
What we give instead
Certainty is not declared in aggregate. It is displayed per field, where it is used to decide.
- A score per field
- Every extracted value carries its own level of certainty. The invoice number and the unit price on line thirty-two are never summarised by the same mark.
- The exact position in the page
- Every field points back to the area of the document where it was read: page, box, line. Checking takes two seconds and does not mean reopening the PDF alongside.
- A human review threshold you set
- You decide the level below which a field goes for review, and which fields that threshold applies to. The unit price can be stricter than the due date.
- Consistency checks, not just reading checks
- Sum of lines against total, quantity against unit, VAT rate against base: a value can be read with certainty and still be inconsistent. The two checks are distinct.
- Doubt is flagged, never filled in
- An unreadable field stays empty and flagged. No value is completed by inference in order to lift a statistic.
- A trace of what was reviewed
- A human correction is recorded with its author and its date. Quality is then measured on your documents, over time, rather than on a promise.
The threshold deserves one more word, because it is what sets the economics of the whole arrangement. Set it high and many fields go for review, consuming human time; set it low and uncertain values pass on fields that decide the payment. The right setting is not the same for a stationery invoice and for a construction payment application, and it is not decided for you.
All of it can be verified on your own documents, during a framed diagnostic. A measurement on your invoices is worth more than a percentage on ours.
What it changes
The useful question is not “how many fields are correct”, it is “am I entitled to pay this invoice”.
- Human checking concentrates where it adds something: the flagged fields, not the full re-reading of a document that was already correct.
- The threshold is set per field and per document type: a timesheet does not carry the same requirements as a framework agreement.
- An error found is tied to the value, to its position and to the person who corrected it: it becomes a measurable fact.
- Progress is measured on your own flow, period after period, rather than on a rate announced at the point of purchase.
It is also what makes the conversation with an auditor possible. A value, its position in the document, its level of certainty, the threshold in force and the trace of the review: all of that can be shown. An overall rate opens onto nothing. And your ERP (enterprise resource planning system) remains the source of truth: approved values are returned to it, never imposed on it.
Capture
Gather and consolidate all your procurement data.
Collection by dedicated mailbox, upload, SFTP, scan, API or ERP export. Reads every format (native PDF, scan, photo, spreadsheet, structured feed), then extracts line by line: supplier, references, quantities, unit prices, discounts, taxes, terms.
The end of re-entry. A procurement history you can finally query.
What Capture detects
- Long documents and multi-page tables
- Unknown layouts
- Several languages and currencies
- Poor scans and handwriting
The subscription covers extraction and its confidence scores. Checking your negotiated terms is the next step and is sized with you: the extraction tiers and their price per page are published plainly.
Nothing is lost. Everything can be checked, everything can be proven.
- The contract clause and the invoice line, highlighted side by side.
- Every extracted value stays linked to the exact place in the document where it was read.
- The same case produces the same decision, today as in six months: the rules are applied deterministically.
- No discrepancy is set aside in silence. Anything that matches no rule is raised, with its reason.
- The agent records what it did, in the order it did it: who, what, how much, when.
Frequently asked questions
What is your accuracy rate?
We publish none, and that is deliberate. A defensible figure would require a measurement on a set of real documents, with a published method and a declared scope. Instead, every field carries its score, its position in the document and its review threshold, verifiable on your own invoices.
How is the human review threshold set?
It is set per field and per document type. Below the threshold, the value goes for review before anything is done with it; above it, the value carries on. The setting is yours and can be changed without us.
What happens when a field is unreadable?
It stays empty and flagged, with the area of the document concerned. No value is inferred to fill a gap: a flagged blank gets handled, an invented value gets paid.
How do I check a Zylio conclusion?
Every discrepancy opens onto its evidence: the contract clause and the invoice line, highlighted side by side, with the calculation shown.
Measurable impact in every environment
More than 5 million procurement documents analysed
Between 1 and 7% of margin recovered
on the scope analysed
From 15 to 45% of time given back to teams, per FTE
depending on the scope and on data maturity
Zylio fits into your existing ecosystem.
The ERP runs the process. Zylio handles the exception and recovers the value that escapes it: invoices without a purchase order, line-by-line price discrepancies, duplicates and overbilling, off-contract spend.
Your data under high security.
Zylio meets the most demanding standards, and nothing is committed without your approval.
- Certifications
- Hosting
- Encryption
- Access
Continue on OCR
- The fields of a supplier invoiceWhat is extracted, and what is then checked against the contract.
- Line-by-line extractionControl happens at line level, so extraction has to as well.
- Traditional OCR or AI OCRWhat one transcribes, what the other understands.
- API and integrationIntegrate without a project: API, SDK, webhook and a dedicated inbox are included in every plan.
See what this looks like on your own data
Twenty minutes, on a spend category of your choosing. We show you what the agents detect, with the evidence behind it.
- No commitment, on your own data
- Result in 3 weeks
- 20 minutes, no sales pitch
- Your data stays hosted in France
- No change of tool or process