Your FEC Export Is Compliant. But the AI That Prepared It Isn't Auditable.
French Audit Trail Requirements for AI-Assisted Accounting Processes
---
A compliant FEC — the Fichier des Écritures Comptables, the standardized accounting data file that French companies must produce for tax audits — means the file format is correct, the mandatory fields are populated, and the data passes the DGFiP's (Direction Générale des Finances Publiques, the French tax authority) technical validation check. That is not the same as saying every entry inside the file can be traced. When tax examiners move past format validation and start asking how specific AI-generated entries were produced — which invoices, which calculation rules, which AI decisions — a clean FEC export that cannot answer those questions has a compliance gap underneath a passed format test.
Producing a compliant FEC proves the accounting entries exist in the right format. It does not prove they can be traced. French tax auditors examine both — and AI-generated entries that pass the format test can still fail the traceability test.
---
Two Different Compliance Tests
French accounting compliance contains two distinct tests that are routinely confused as a single standard.
Format compliance is the technical structure of the FEC export file: correct field order, proper encoding, all mandatory fields present. The DGFiP's validation tools evaluate format compliance automatically at export time. Every accounting software that produces a valid FEC passes this test — that is the basic function of the software.
Source documentation traceability is different. Under LPF — the Code of Tax Procedures — Article L47 A, every accounting entry in the FEC must be traceable to its pièce justificative: the source document, whether an invoice, a receipt, a bank statement, or a contract, that justifies the entry. This test is not automated. It is conducted by a human examiner during a tax audit, who selects specific entries and asks to trace them to their source documentation on demand.
Passing the format test is table stakes. Failing the traceability test is where the €5,000 minimum penalty for FEC non-compliance begins — and where the DGFiP's authority to unilaterally reconstruct the company's accounts comes into play.
---
What Source Documentation Means for AI-Generated Entries
For human-generated accounting entries, traceability is straightforward: the comptable (accounting manager) who processed the invoice recorded the entry, the invoice is in the archive, the trail is clear. For AI-generated entries, the trail must still exist — but it is only as clear as the logging architecture the AI system was designed to produce.
Under LPF Article L47 A, an AI-generated accounting entry requires five traceable elements: the specific source invoice or document reference; the date and time of AI processing; the categorization and reconciliation rules applied to generate the entry; any exception flags or corrections applied; and the final entry with a complete citation traceable to the original document. These are the same documentation requirements that apply to human-generated entries. AI cannot reduce the documentation obligation — it can only make fulfilling it faster or harder depending on its architecture.
Most AI accounting tools were designed to process invoices at volume and produce FEC-valid exports. Batch-level logging — records showing that 450 invoices were processed on October 14th and the totals match — was designed to satisfy format validation requirements. It does not satisfy the traceability requirement when an examiner selects one of those 450 invoices and asks to trace the specific entry it generated.
---
What Happened in 2024
A French mid-market distribution company produced a clean FEC export for a tax audit in 2024. Format validated. Filed within the 15-day deadline under LPF Article L47 A. The examiner then questioned a series of automatically reconciled entries — approximately 340 entries representing €2.3 million in transactions.
The AI system that processed the underlying invoices produced batch logs: reconciliation runs, totals per batch, processing timestamps. It did not produce entry-level documentation linking each questioned entry to its specific source invoice. The examiner treated the questioned entries as inadequately documented under the pièce justificative standard.
If source documentation is insufficient, the DGFiP has the authority to reconstruct the company's accounts using its own methodology — applying conservative assumptions that favor the tax authority. The total tax exposure from the account reconstruction exceeded the original FEC discrepancy by a factor of eight. A documentation gap that looked like a logging detail at software selection time produced a material tax event at examination time.
Nobody asked the right question when the AI accounting tool was evaluated. Not the comptable who confirmed it produced valid FEC exports. Not the CFO who approved the purchase. Not the tax advisor who signed the annual submission. The question that was never asked: for any AI-generated entry, can the system produce the source documentation trail that LPF Article L47 A requires? That missing question accumulated into an examination risk with every invoice the AI processed.
---
The Five-Field Checklist
Before the DGFiP calls, send this list to your accounting AI vendor and ask for written confirmation on each point:
Does the system link each AI-generated accounting entry to the specific source invoice — with the document's unique reference, not just a batch identifier? Does it log the date and precise time of AI processing for each entry individually? Does it record the categorization and reconciliation rules applied to generate each specific entry? Does it flag and document any exceptions, corrections, or overrides at the entry level? Does it produce a source citation for each entry traceable to the exact document in the archive?
Ask for written confirmation on all five. A vendor whose architecture provides entry-level traceability will confirm without hesitation. A vendor whose architecture provides batch-level logging will describe what it does produce — and what it describes will not map to the five fields.
---
Accountability Sits With the Company, Not the Vendor
Legal responsibility for FEC source documentation under LPF Article L47 A sits with the company subject to audit — not with the accounting software vendor. If an AI tool generates an entry that cannot be traced to a source document when the DGFiP requests it, the €5,000 penalty, the reconstruction authority, and any resulting tax assessment fall on the company.
Accounting software contracts typically include language limiting the vendor's liability for tax examination outcomes. The vendor's product may have produced a technically valid FEC export — meeting the contractual specification — while leaving the company without the entry-level documentation the DGFiP requires. These are different standards, and the gap between them is the company's legal and financial problem, not the vendor's.
---
How Entry-Level Logging Works in Practice
Stralevo's SightCapture document processing captures every invoice at the field level — not just the amount and VAT, but the serial number, payment reference, delivery terms, and any other field present in the source document. This field-level capture at ingestion time is the foundation of entry-level audit trail logging: the source data exists at the document level, linked to the accounting entry it generated.
An examiner asking for the source documentation trail for any entry in a FEC produced with Stralevo's accounting intelligence layer receives the complete five-field documentation package — source document reference, processing timestamp, applied rules, exception handling, and entry citation — for any entry, from any processing date in the system's history. That documentation is not assembled after the examiner asks. It was generated at the moment the invoice was processed and has been available on demand since then.
Speed compounds the benefit: when every invoice generates a complete documentation package at processing time, preparing for a tax audit is a retrieval exercise, not a reconstruction. Companies with entry-level AI logging typically produce the complete source documentation package for a DGFiP examination in hours. Those still relying on batch logs spend weeks reconstructing trails — if reconstruction is possible at all.
---
The DGFiP's Direction of Travel
Examination practice evolves in one direction: more specific, more documented, more traceable. Early FEC examinations focused primarily on format and totals. Current examinations increasingly target specific entries and require granular source documentation. As AI accounting becomes more prevalent in French mid-market companies, examination methodology will become more attentive to entry-level AI audit trails — not less.
Companies that built entry-level traceability into their accounting AI architecture before this specificity arrived will absorb more detailed examination questions without disruption. Those relying on batch-level logging will face documentation requests their systems were not designed to answer — at examination timelines that do not allow for architectural rebuilds.
Entry-level AI traceability is an architectural decision made before the first invoice is processed. After millions of invoices have been processed through a batch-logging system, requesting retroactive entry-level documentation is a reconstruction project that takes months and still may not produce documentation that satisfies the pièce justificative standard. The moment to build it in is when the system is selected — not when the examination notice arrives.