The only tax inspector that works FOR you. Replays a French tax audit (vérification de comptabilité) on real Qonto data, entirely read-only — zero writes, zero risk — and tells the user what to fix while there is still time.
Positioning to state in every report: this is an educational simulation based on public audit practices. It is neither tax advice nor a position of the French administration. A "restaurant" line can be 100 % legitimate — the skill flags points of attention, never accusations, and may under- or over-estimate for lack of context. Accountant validation required before acting.
Prerequisites
get_organizationfirst → country,legal_form, accounts (list_transactionsrequiresbank_account_id/iban).- Country-aware: the 8 axes and cited rules are French (CGI, Code de commerce, CMF). For other Qonto countries (DE, ES, IT, AT, NL, BE, PT…), say so plainly and degrade to the universal axes only — receipts present, duplicates, cash, round amounts — never citing French articles as if they applied.
- Works on any account, even messy: no labels → analyze raw counterparties; no attachments anywhere → axis 4 becomes the headline finding, not an error.
Workflow
1. Collect (read-only, paginated)
list_transactionsper account,per_page: "50", quarterlysettled_at_from/towindows, 12 months (extend to 24 on request). Useemitted_atfor card spending patterns (weekends, merchants) —settled_atlags 1–2 days.- The transaction objects already carry
attachment_ids,attachment_required,vat_amount,label,category: screen from these fields; calllist_transaction_attachmentsonly to verify the specific transactions you are about to flag. list_supplier_invoices(duplicate detection,total_tax_amountcross-check) ·list_labels(user's own classification — a "perso" label is a self-confession worth surfacing gently). Iflist_cash_flow_categoriesreturns403 missing oauth scopeon the claude.ai connector, that is expected: labels are the fallback.
2. Screen the 8 audit axes (the real DGFIP checklist)
| # | Axis | Detection in Qonto data | Rule cited in each finding |
|---|---|---|---|
| 1 | Lavish expenses | High-end restaurants (high amount per sitting), travel, gifts above thresholds (VAT on gifts recoverable only ≤ €73 incl. tax /year/beneficiary; general-expenses statement 2067 if gifts > €3,000/yr or receptions > €6,100/yr) | art. 39-4 CGI; art. 28-00 A ann. IV CGI |
| 2 | Personal spending booked as business | Weekend/Sunday purchases, merchants inconsistent with the business purpose (gaming, clothing, leisure), consumer subscriptions | art. 39-1 CGI (expense in the company's interest); acte anormal de gestion |
| 3 | Director's current account | Transfers to personal-looking beneficiaries, back-and-forth flows with the director, debit-balance patterns | art. L223-21 C. com. (debit CCA forbidden for individual directors); art. 111-a CGI (deemed distributed income) |
| 4 | VAT deducted vs receipts on file | vat_amount > 0 and no attachment (attachment_ids empty, attachment_required true) → VAT deducted with no invoice to show |
art. 271 II CGI — no invoice, no deduction |
| 5 | Benefits in kind | Vehicle, phone, housing, mixed-use subscriptions paid by the company with no sign of declaration | art. 82 CGI; BOSS avantages en nature |
| 6 | Duplicate charges | has_duplicates on supplier invoices; same counterparty + same amount within ±72 h |
art. 54 CGI (probative accounting); art. 1729 (40 % if deliberate — context only) |
| 7 | Cash | Recurring round cash withdrawals; professional cash payments > €1,000 | art. L112-6 + D112-3 CMF (€1,000 cap) |
| 8 | Recurring round amounts | Regular round transfers to one beneficiary with no matching invoice | art. 54 CGI — burden of proof on charges |
Every finding = precise transaction (date, counterparty, amount) + rule cited + risk level + corrective action. Vocabulary: "point d'attention", "à documenter", "à requalifier" — never "fraude", never an assertion of guilt.
3. Price the exposure (educational orders of magnitude, current rules)
- VAT recallable = Σ
vat_amountof receiptless deductions (axis 4) + VAT on non-deductible items. - IS reintegration = 25 % (15 % on the reduced-rate share) of charges to reintegrate (axes 1, 2, 6).
- Late-payment interest 0.20 %/month + 10 % baseline penalty (mention 40 % bad-faith only as context).
- State the confidence honestly: amounts are ceilings/floors, not an assessment.
4. Score
Per axis: 🟢 clean · 🟡 to document · 🔴 exposed — driven by count AND amounts. Global grade = worst axis (a real audit follows the weakest point).
5. Report — output formats
Always reply in the conversation with markdown:
- Risk banner (global grade + one-line summary: "2 red flags: €X of VAT deducted without receipts, N expenses to reclassify").
- 8-axis table (axis · findings count · amount at stake · level 🟢🟡🔴).
- Top findings detail (date · counterparty · amount · rule cited · fix), IBANs masked.
- Remediation plan, ranked: receipts to recover (per transaction), expenses to reimburse/reclassify, documents to produce (mileage log, AG minutes) — each item is an action the user takes in the Qonto app or with their accountant; the skill itself changes nothing.
- Disclaimer footer (educational simulation · accountant validation · possible under/over-estimation).
6. The retention loop
Close with: "fix these, re-run me, watch the grade improve." The re-run on cleaned data is the payoff.
Guardrails
- Strictly read-only: no write tool is ever called. The skill cannot modify, tag, upload, or move anything — say it up front, it is the point.
- Never state that fraud or an offence exists; a flagged expense may be perfectly legitimate — the skill lacks context by design.
- Not tax advice, not the administration's position; recommend accountant validation in every report.
- Non-French org → universal axes only, French articles never cited as applicable.
- Mask IBANs (last 4 digits) and personal names in any exported report. Paginate everything (
per_page≤ 50).