AI for B2B Credit Files in France: Automation & Human Role
By Sygnet Research, checked before publication
Key takeaways
- AI can reliably automate the mechanical half of a credit file: reading the extract, matching the SIREN, checking dates, comparing names and addresses, pulling figures out of the financial statements and flagging documents that have been retouched.
- It cannot automate the decision. Setting an encours (credit limit) still requires a human who knows the margin, the sector and the commercial stakes.
- Two checks must always end in an official register, never in a PDF: the company's registration (RNE/Kbis) and the URSSAF attestation de vigilance, which is valid 6 months and must be renewed throughout the service.
- The risk is real: French business failures rose again in 2025, to 68,602 procedures, 15% above the 2010-2019 average.
Which documents make up a B2B credit file in France?
A standard French trade credit file holds five to eight pieces: proof of registration, proof of social security compliance, recent financial statements, a RIB (bank details), signed CGV or a credit application, and often an ID document for the person signing. Nothing exotic, but each one carries a different verification logic.
The registration proof is where most confusion starts. Since 1 January 2023 the Registre national des entreprises (RNE) has been the central register for all French companies, the Kbis extract is no longer requested by the French administration, and a free INPI/RNE extract containing the same information can be downloaded from the Annuaire des Entreprises. For a third party wanting a certified Kbis, Infogreffe charges 3.06 € online in 2026. And many of your prospects simply have no Kbis at all: since the RNE was created, artisans, liberal professions and commercial agents hold an RNE extract instead.
The attestation de vigilance is a separate obligation, not a nice-to-have. It is mandatory when the contract concluded with your donneur d'ordre is worth 5,000 € excluding tax or more.
What can AI verify automatically, and what still needs a human?
AI handles everything that can be expressed as a comparison: text against text, a figure against another figure, a document against a register. Take a 40-person building materials distributor opening 60 trade accounts a month. The machine reads each incoming PDF, identifies the document type, extracts the SIREN, the legal name, the address, the officers, the issue date, then cross-checks those fields against the official registry data and against the credit application the prospect typed in. Mismatched name, expired extract, address that differs from the register: flagged in seconds.
What it should not do alone is weigh the result. A 2-year-old company with thin equity may still deserve 15,000 € of credit because the founder is a known contractor. That judgement is yours.
| Check | Automatable? | What it relies on |
|---|---|---|
| Document type and completeness of the file | Yes | Classification of the uploaded PDF |
| SIREN / legal name / address match | Yes | Comparison with RNE and Sirene data |
| Extract age (the customary "less than 3 months") | Yes | Date on the document |
| Insolvency proceedings mentioned on the extract | Yes | Register lookup |
| Authenticity of the URSSAF attestation | Yes, via the security code | URSSAF verification service |
| Key ratios from the financial statements | Yes | Table extraction from the liasse fiscale |
| Signs of tampering (retouched PDF, altered figures) | Partly | Fraud detection, then human review |
| The credit limit itself | No | Human decision |
The machine's job is to make sure every field in the file is true; yours is to decide what the truth is worth in euros.
How does AI check a registration extract?
By reading the document and then ignoring it in favour of the register. A Kbis or RNE extract is a snapshot, and the only question that matters is whether it still matches reality. No legal text sets a validity period for the extract: it reflects the register on the day it was issued, and the three-month rule is simply usage, applied by most banks, landlords and contracting parties.
So the sequence is: extract the SIREN from the PDF, query the register, compare. All the data displayed on the Annuaire des Entreprises is available through INPI's Registre National des Entreprises API, which means the comparison can run without anyone retyping a nine-digit number. Three things matter most for credit: the legal representative (is the person signing the credit application actually authorised?), the registered address, and any ongoing collective proceedings such as redressement or liquidation, which appear on the extract.
One caveat worth knowing before you build alerting on registry data: a company can restrict public broadcast of its personal data, so names of natural persons, address and geolocation may be hidden on request.
Can AI authenticate an URSSAF attestation de vigilance?
Yes, and this is the clearest win in the whole file. Every attestation carries a security code, and the donneur d'ordre must verify validity and authenticity at contract signature and then every 6 months, by entering the 15-character security code into URSSAF's verification tool. Reading that code off a PDF and submitting it is exactly the kind of task that should never occupy a credit analyst.
The alternative route removes the step entirely: attestations obtained through API Entreprise come directly from URSSAF and are therefore certified, so the code is not needed to verify authenticity. The API also returns the average monthly headcount computed by URSSAF from DSN data and the payroll total from the last contribution statement, two useful sanity checks on a prospect's declared size.
Skipping this is expensive. A donneur d'ordre whose subcontractor used undeclared labour can be held liable for that subcontractor's taxes and social contributions in proportion to the turnover generated.
Falsified attestations are common enough that checking the security code is not paranoia, it is the minimum.
What does AI read in the financial statements?
It reads the tables: turnover, equity, net result, debt, trade payables and receivables, then computes the ratios your credit policy actually uses. The liasse fiscale and annual accounts are structurally hostile to manual entry (dense numeric grids, multi-page, inconsistent layouts), which is precisely why table extraction quality matters more here than raw text accuracy. If you are comparing tools, Sygnet's note on table extraction metrics explains how those tables are scored.
Two limits deserve honesty. First, many accounts are stale by the time you read them, and a large share of French companies file them confidentially. Second, financial statements say nothing about payment behaviour, which is the better predictor. In 2025, 86% of French companies still reported suffering late payments according to Coface, with delays averaging 49.7 days. The Banque de France reads the same signal: it downgraded the rating of 1,290 companies in 2025 because of their payment behaviour.
Watch the profile too. In 2025, failures of SMEs, mid-caps and large companies ran 68% above their 2010-2019 average, while microenterprises, 92% of all failures, were only 12% above.
How do you turn verified documents into a credit limit?
By writing the rules before the file arrives, then letting the system apply them. A workable policy for our distributor: file complete and all identity checks green, turnover above 500 K€, equity positive, no collective proceedings, attestation valid, then auto-approve up to 10,000 € or 5% of last-year turnover, whichever is lower. Anything above, or any red flag, routes to a human with the discrepancies listed.
Pin the payment terms to the law while you are at it. Article L441-10 of the Code de commerce caps terms at 60 days from issue or 45 days end-of-month; absent a clause, it is 30 days after receipt or performance. Breaching it is costly: administrative fines up to 75,000 € for an individual and 2 million € for a legal entity, doubled on repeat within two years, and enforcement is tightening, with fines rising from 19.1 M€ in 2022 to 35.8 M€ in 2024 and the average fine from 94,000 to 165,000 €.
Late payment also triggers automatic remedies: penalties plus the 40 € flat recovery indemnity are due as of right, with no formal notice.
What are the compliance constraints?
Two, mainly. Identity documents and payslips are personal data, so a credit file built with AI falls squarely under the GDPR: lawful basis, retention limits, minimisation (do you really need a full ID scan, or just a name match?) and the right to a human review of any automated refusal. Sygnet's GDPR checklist for document AI projects covers the practical points, and the KYC document collection checklist maps which pieces to ask for in the first place.
The second is traceability. Keep the verification evidence, not just the answer: which register was queried, when, with what result. Attestations should be kept at least 5 years. E-invoicing will make this easier and stricter at once, since issue and receipt dates are time-stamped on the state platform; Sygnet, which builds document extraction tooling, covers the calendar in its guide to e-invoicing in France. For altered documents specifically, see the document fraud detection guide.
FAQ
Is an RNE extract enough, or do I need a real Kbis?
For your own internal credit checks, the free RNE extract is usually enough: it contains the same information as a Kbis and is the version in force within the administration. But it is an informative document without the court registry stamp, and some banks and administrations accept it while others still require an official Kbis. Decide once, write it into your credit policy, and stop debating it file by file.
How often should I re-verify an existing customer?
Align on the URSSAF rhythm: every six months, because the attestation is valid 6 months and one collected at the start of a long contract will expire before the contract ends. Registry checks can run more often, since they are cheap through an API. Any significant limit increase should trigger a fresh check regardless of the calendar.
Can AI detect a forged payslip or a doctored attestation?
Partly. Tampering leaves traces (inconsistent fonts, misaligned figures, totals that do not add up, PDF metadata), and those can be flagged automatically. Authenticity itself, though, is only proven by the issuing source: the URSSAF security code, the register, the bank. Treat AI as a filter that sends the suspicious 3% to a human, not as a verdict.
One email when we publish something worth your time.