French Tax Audit Targeting Algorithm: A Practitioner's Framework

I. The algorithmic framework of French tax control: what public sources reveal

A. The method: a risk-scoring system fed by eleven databases

The CFVR project and its architecture. The Fraud Targeting and Query Enhancement project was established by the ministerial order of 21 February 2014, amended in particular by the order of 28 August 2017 which extended its scope to individuals. This automated processing operation, now run by the SJCF-1D bureau of the Tax Legal Security and Audit Service, does not automatically characterise fraud: it produces a probabilistic ranking of files by level of atypia, and operational departments then analyse the leads and decide whether to launch an audit. As the Cour des comptes noted in its 2018 public report and more recently the Senate report by Carcenac and Nougein (Senate, no. 668, 22 July 2020), the explicit objective is to bring the share of operations programmed by data analysis to fifty percent; an objective the DGFiP indicates it has exceeded for business audits since 2023.

The data sources mobilised. The eleven databases aggregated in the CFVR silo are identified by the successive orders and by the CNIL deliberations accompanying each evolution (notably deliberations no. 2014-049, no. 2017-249 and no. 2022-025). They include internal declarative applications (personal and corporate income tax, VAT, real estate wealth tax, registration duties), wealth databases (PATRIM, FIDJI, BNDP), banking files (FICOBA for bank accounts, FICOVIE for life insurance contracts), social data (the DSN nominative social declaration transmitted by employers), external economic sources (the BODACC official bulletin, the register of beneficial owners, INSEE), and international automatic exchange flows (CRS, DAC 2, DAC 6, DAC 7 and now DAC 8 for crypto-assets). The GALAXIE project, authorised by ministerial order of 11 March 2022 following a favourable opinion from the CNIL (deliberation no. 2022-025 of 17 February 2022), adds a graphical visualisation function for capital and personal links, allowing an auditor to reconstruct in seconds a chain of shareholdings or a network of common associates.

The statistical method. The processing combines, according to the DGFiP's own indications in its activity reports, two complementary approaches. The first, supervised, trains the model on files previously characterised as fraudulent, to identify in the current taxpayer population those profiles that exhibit the same statistical features. The second, unsupervised, directly seeks atypia relative to the modal behaviour of a reference population, without pre-judging the nature of the anomaly. It is this second approach that produces the bulk of new files directed to the audit services, since it allows the detection of novel schemes that supervised learning would miss.

B. The public sources available to reconstruct the grid

The regulatory texts. Without giving access to the code, the orders establishing the processing operations specify, in compliance with the French Data Protection Act and the General Data Protection Regulation, the purposes pursued, the categories of data processed, the recipients and the retention periods. These texts constitute the primary source of information on the variables exploited. The order of 14 September 2017 in particular marked the extension of the CFVR to individuals by expressly listing the categories of data examined for the detection of "behaviour presenting a risk of fraud", including in relation to tax residency. The order of 11 March 2022 establishing GALAXIE details the categories of information mapped: capital links, corporate mandates, bank accounts, real estate and movable property.

The CNIL deliberations. The French data protection authority systematically examines draft orders before publication and renders public opinions. The successive deliberations on the CFVR, on the experimentation provided for by article 154 of the 2020 Finance Act (automated collection of publicly accessible data from online platforms) and on GALAXIE are sources that describe, in accessible terms, the technical contours and the safeguards in place. The CNIL noted in particular, in its deliberation no. 2019-114 on article 154, that the mechanism only covered freely accessible content and excluded restricted content as well as private messaging, a useful precision for practitioners advising clients.

The parliamentary reports and Cour des comptes. The Senate's information report no. 668 of 22 July 2020 on the resources of tax audit, presented by senators Thierry Carcenac and Claude Nougein, remains to date the most comprehensive public synthesis on the organisation and tools of algorithmic programming. It is supplemented each year by the appendices to the Finance Bill relating to the fight against tax evasion, which provide precise indicators on yields, methods and developments. The Cour des comptes (the French Court of Audit) has also published several referees and reports on the matter (notably the referee of 6 December 2018 on fraud against mandatory deductions).

Academic literature. The article by Stéphane Créange published in the Revue française de finances publiques (RFFP no. 153, 2021, p. 21) remains the academic reference on data mining at the DGFiP, written by a direct participant in the system. It is usefully complemented by the contributions of Mohamed Kimri and Pierre Legros (Revue de droit fiscal, no. 5, 4 February 2021, § 122) on the legal regime of algorithmic tax control, as well as by the regular chronicles in Les Petites affiches and Revue de droit fiscal. This literature, cross-referenced with the regulatory texts, is sufficient to reconstruct the main families of signals exploited by the models.

II. Mapping the signals of atypia and institutional monitoring thresholds

The scoring logic is not binary. Each variable contributes, with its own weighting, to a global score that ranks files in decreasing order of risk. Detection does not rely on absolute thresholds but on statistical deviations from a peer group defined by sector of activity, size, age and geographic area. Three main families of signals structure the analysis: intra-declarative inconsistencies, gaps between filings and third-party-reported data, and indications from open external sources.

A. Intra-declarative inconsistencies and temporal breaks

Temporal breaks within the same taxpayer's file. Unsupervised models primarily detect atypical year-on-year variations. A change of more than thirty percent in declared income, up or down, without an apparent justification drawn from a change in family or professional situation declared elsewhere, constitutes a first-order signal. The same is true of the sudden disappearance of a category of income (non-commercial profits, property income, dividends) without declared cessation of activity or disposal, the durable emergence of losses after a period of stable profits, or the shift of a household into a marginal tax bracket much lower than that of preceding years. These signals are not pathognomonic in themselves: they merely orient the model towards a deeper examination of the file.

Inconsistencies between related filings. The DGFiP systematically exploits the internal coherence of the declarative file. A form 2047 declaring foreign-source income without the corresponding form 3916 declaring foreign accounts, or vice versa, constitutes an immediately identifiable discrepancy. A real estate wealth tax (IFI) declaration not reflected in the property income return (form 2044) in amounts proportionate to the declared wealth, significant charitable giving (articles 200 and 238 bis of the French Tax Code) without apparent contributive capacity, recurring tax credits for in-home services or childcare without consistent household composition, all of these are configurations the model automatically flags. The Senate report cited above confirms that these intra-declarative cross-checks form the backbone of the scoring for individuals.

Structural atypia for companies and holding owners. For entities, the model examines financial ratios and their position relative to the sector. The shareholder current account to equity ratio, the financial charges to operating profit ratio, the manager's remuneration to profit ratio, the gross margin relative to the sectoral peer group, the average payment terms of customers and suppliers, and the rate of deductible VAT recovery relative to turnover are among the classic indicators of the statistical query. An extreme ratio does not in itself characterise any anomaly: there are perfectly legitimate economic justifications for all these configurations (vendor financing in the context of a sale, start-up phase, family holding, recent owner-buy-out). But an extreme ratio triggers a verification of the supporting documentation the taxpayer must be able to produce.

B. Gaps between filings and third-party-reported data

The banking and wealth pillar. The FICOBA (bank accounts) and FICOVIE (life insurance contracts above EUR 7,500) databases allow a direct comparison between the accounts effectively held and those reflected in income filings. A creditor account whose annual cumulative flows significantly exceed declared income, or a life insurance contract whose partial redemptions do not give rise to taxation as capital gains, are first-order signals. The PATRIM and FIDJI databases, derived from notarial deeds, also allow real estate acquisitions to be matched against declared financing capacity: an acquisition whose price exceeds annual income several times over without visible bank financing constitutes a signal for an examination of personal tax situation (ESFP) under article L. 12 of the French Tax Procedure Code.

The social and professional pillar. The DSN nominative social declaration transmitted monthly by employers to URSSAF is matched against the salaries declared in the income tax return. Any significant difference, whether reflecting under-declaration of salaries or, conversely, declared income without an identified employer, triggers a signal. Fees reported on form DAS 2 by paying companies are matched against the non-commercial profits declared by their recipients. For liberal professions and the self-employed, the decorrelation between the chronology of bank receipts and the chronology of filings is one of the most mobilised indicators.

The external wealth pillar. The land registry, through the MAJIC file, is now systematically cross-checked against IFI filings and property income returns. The Foncier innovant project, authorised by ministerial order and deployed since 2021, exploits aerial photographs from the French national geographic information institute (IGN) to detect undeclared swimming pools and buildings; the first results published by the DGFiP, reporting more than 140,000 pools reassessed over 2022-2023, demonstrate the effectiveness of the approach. The register of beneficial owners (RBE), kept by the French intellectual property institute (INPI), is also mobilised to identify indirect holdings not reflected in wealth declarations.

The international pillar. Automatic exchanges of information are today the most dynamic source of targeting. Directive 2014/107/EU (DAC 2) transposed into domestic law and the OECD Common Reporting Standard (CRS) require partner administrations to transmit annually information on financial accounts held by French tax residents: year-end balances, interest, dividends, sale proceeds, beneficial owners. According to figures communicated by the DGFiP, more than seven million foreign accounts were reported on inbound flows for the 2023 financial year. Directive 2018/822/EU (DAC 6) requires intermediaries to declare cross-border arrangements bearing certain hallmarks; Directive 2021/514/EU (DAC 7) extends the obligation to online platforms; Directive 2023/2226/EU (DAC 8), which entered into force on 1 January 2026, completes the framework for crypto-assets. Each of these sources directly feeds the risk score when a discrepancy is identified with forms 3916, 3916-bis or 2042.

C. Signals from open external sources and the international dimension

Automated collection from online platforms. Article 154 of the 2020 Finance Act, since extended and made permanent, authorises the DGFiP and the customs administration to automatically collect freely accessible content from online platforms. This collection, framed by the CNIL and by the Constitutional Council (decision no. 2019-796 DC of 27 December 2019), specifically targets the fight against undeclared activity (repeated sale advertisements without business registration), false residency declarations and ostentatious discrepancies between declared income and external signs of wealth. The mechanism was subject to an interim assessment by the CNIL in 2024, the conclusions of which are publicly accessible.

International mobility as one application field among others. Public sources show that international mobility is a field of application of the models, without being the most important quantitatively. The indicators exploited are those derived from article 4 B of the French Tax Code (home, principal place of residence, principal professional activity, centre of economic interests) cross-referenced with data held by the administration: maintenance of subscriptions or utilities in the taxpayer's name, schooling of children in France, duration and frequency of stays reconstructed from available data, retained corporate mandates, receipt of French-source income disproportionate to declared foreign income. Article 61 of the 2025 Finance Act extended the statute of limitations to ten years for false tax residency in respect of personal income tax, real estate wealth tax and gratuitous transfer duties, without altering the underlying algorithmic criteria. It is one technical field among others, the treatment of which obeys the same rules of statistical atypia as other audit areas.

Filing reliability signals. Beyond substantive anomaly signals, the model also exploits indicators of formal filing reliability: recurring late filings, repeated voluntary corrections on the same items, frequent contentious refund claims, systematic reliance on filing extensions. These indicators do not in themselves reflect any fraud, but they orient audit resources towards files where the investigation has the highest statistical probability of leading to significant reassessments. This efficiency logic is explicitly assumed by the DGFiP in its strategic documents.

D. Institutional thresholds for permanent monitoring: high-stakes and very high-stakes files

An institutional dimension distinct from algorithmic scoring. Beyond the statistical targeting carried out by the CFVR and GALAXIE models, the administrative organisation of tax audit relies on a second, long-standing and stable grid which assigns each taxpayer to a competent audit service based on income and wealth thresholds. This segmentation, formalised by DGFiP circular CF1/2014/10/9600 of 9 October 2014 generalising the Income and Wealth Audit Units (Pôles de contrôle des revenus et du patrimoine, PCRP), by administrative guidance published in the Bulletin officiel des finances publiques (BOI-CF-DG-20), and by the successive annexes to the cross-cutting policy document on combatting tax evasion attached to the Finance Bill, is documented in publicly available sources. It is not intended to remain secret: it determines the allocation of audit human resources.

The three strata and their thresholds. The first category, known as "infra-DFE" (sub-high-stakes), covers tax households whose gross annual income falls between approximately EUR 180,000 and EUR 270,000. The second category, the High-Stakes Files (Dossiers à Fort Enjeu, DFE), covers households whose gross income exceeds EUR 270,000 (a threshold raised to EUR 500,000 where salaries, wages and pensions constitute the bulk of income) or whose gross assets liable to wealth tax (formerly the ISF, now the IFI in its real estate scope) exceed EUR 2.5 million. The third category, Very High-Stakes Files (Dossiers à Très Fort Enjeu, DTFE), corresponds to households with gross income exceeding EUR 2 million or gross taxable assets exceeding EUR 15 million; this portfolio has fallen, since September 2011, within the exclusive competence of the National Tax Situation Audit Directorate (DNVSF), in accordance with the orientations detailed in the annexes to the 2019 Finance Bill.

The underlying logic: salaried remuneration enjoys a presumption of reliability. The dual threshold of EUR 270,000 / EUR 500,000 for the DFE category is not a technical detail. It reflects a strong orientation of audit policy: remuneration paid by a third party, subject to withholding at source and reflected in the monthly nominative social declaration transmitted by the employer to URSSAF, carries a very low residual risk of under-declaration. Conversely, manager remuneration paid by one's own structure, non-commercial profits and capital income statistically carry a higher risk of optimisation or inaccuracy, justifying monitoring from the lower threshold. This distinction has practical implications for remuneration structuring, particularly in holding/operating company configurations where the qualification of remuneration can shift a taxpayer from one stratum to another.

The 2017-2018 shift: from systematic triennial review to risk-driven examination. Until the end of 2017, High-Stakes Files were subject to a systematic triennial desk audit conducted using the correlated income/wealth audit method, aligned with the statute of limitations under article L. 169 of the French Tax Procedure Code. The cross-cutting policy document attached to the 2019 Finance Bill expressly acknowledges this evolution: the selection of files for in-depth examination among DFEs is now driven by risk-analysis queries produced by the Mission requêtes et valorisation, which allows, in the terms of the orientation note cited in the PCRP annexes, "a controlled lightweight audit of a DFE where it has already been audited during the previous triennial period and its situation has not changed". The examination is therefore no longer systematically triennial; it remains regular, and every DFE household continues to be permanently registered in the portfolio of the PCRP of its department of attachment.

The specific case of DTFEs and the stock effect. The DNVSF's DTFE portfolio, by contrast, retains a logic of continuous monitoring, with no predetermined cadence but with a dedicated auditor who is personally familiar with the file. A notable feature of this category lies in what we describe as a stock effect: a taxpayer who entered the DTFE perimeter because of the magnitude of their wealth or income does not automatically exit when their annual flows decline. A business disposal followed by a transition to more modest investment income does not cause the qualification to lapse, since it is assessed by reference to the patrimonial stock and, where relevant, to reputation. This is particularly relevant for taxpayers in post-disposal or patrimonial transition phases, who remain attached to the Paris DNVSF for several years after the operation.

Articulation with algorithmic targeting. The two grids, institutional and algorithmic, do not conflict but overlap. The institutional grid determines which audit service is competent and ensures permanent monitoring of the households concerned; algorithmic targeting determines, within each portfolio, which files will be prioritised for in-depth examination. For a DFE household, registration in the PCRP portfolio means a regular desk audit (annual verification of declaration consistency, automatic cross-checks against third-party sources) coupled with a risk of escalation (information request, request for clarification, request for justification, or even the launch of a contradictory examination of personal tax situation) triggered by any significant atypia signal. The security of the declarative file is therefore to be assessed over time, and not by reference to a hypothetical exceptional audit.

III. Practical consequences: preventive review, voluntary disclosure and compliance strategy

A. The preventive review: anticipating the signals the algorithm will detect

The reversal of reasoning. Knowledge of the targeting grid allows the traditional reasoning to be inverted: rather than waiting for an examination notice to reconstruct the justification of a declarative situation, the taxpayer and their adviser can identify upstream the configurations likely to produce an atypia signal and prepare the corresponding documentation. This approach, sometimes referred to as preventive tax review, has nothing to do with concealment: it consists in anticipating the questions the administration will ask and documenting the answers while the memory of the operations is still fresh. The marginal cost of such a review is incomparably lower than that of an improvised response to a request for justifications addressed several years after the facts.

The recommended scope of examination. A complete preventive review covers the entire declarative file over the applicable statute of limitations, three years for ordinary income tax (article L. 169 of the French Tax Procedure Code), six years in case of undeclared activity or non-declaration of foreign assets (article L. 169 paragraph 4), and now ten years in case of false tax residency for personal income tax, real estate wealth tax and gratuitous transfer duties. It successively examines the internal consistency of filings, the concordance with accessible third-party sources (bank statements, FICOBA extracts obtained by exercising the right of access, accounts of held companies) and the traceability of atypical operations. Traditionally sensitive areas, shareholder current accounts, flows between private and professional patrimony, cross-border movements, and preferential regimes (Dutreil pact, contribution-disposal, dismemberment), deserve particular attention.

Documentation as a shield. In tax audit matters, it is not the economic legitimacy of an operation that guarantees its security, but the ability to demonstrate that legitimacy when the question is asked. A shareholder current account of one million euros backed by a duly formalised vendor's loan is unassailable; the same current account without a documented sale agreement or formalised loan convention becomes a subject of contentious discussion, even if its origin is perfectly lawful. The lesson holds for any structuring operation: contemporaneous formalisation of the operation is the first line of defence, and it is precisely this formalisation that the preventive review aims to objectivise.

B. Voluntary disclosure: an underused tool

The regime of article L. 62 of the LPF. When the preventive review reveals a significant omission or inaccuracy, the taxpayer has access to a powerful and largely underused tool: voluntary corrective filing. The mechanism of article L. 62 of the French Tax Procedure Code, extended and softened by the law of 10 August 2018 for a State at the service of a society of confidence (ESSOC), allows for a substantial reduction in late-payment interest (30 percent reduction if the correction occurs spontaneously before any audit act). This reduction is available as long as the administration has not initiated an audit procedure on the period concerned and as long as the taxpayer is not required to regularise by a prior formal notice.

The articulation with the so-called "Bercy lock". The law of 23 October 2018 on the fight against fraud amended article L. 228 of the LPF by introducing an obligation of automatic transmission to the public prosecutor of files presenting the most serious characteristics (duties exceeding EUR 100,000 with penalties excluding good faith). Voluntary disclosure made before the initiation of an audit procedure excludes this automatic transmission: this is one of the major advantages of preventive regularisation for significant files. Administrative guidance (BOI-CF-INF-40-10-10) details the conditions of this articulation.

The economic calculus of disclosure. Voluntary disclosure should, however, only be considered after a rigorous cost-benefit analysis. The cost comprises the additional duties owed, the reduced late-payment interest, and the residual risk of further examination on other years or other taxes. The benefit lies in the gap between this burden and the burden that would result from a subsequent contradictory reassessment, with potential penalties of 40 percent for deliberate misconduct, 80 percent for fraudulent manoeuvres, and the criminal risk for the most significant files. In the great majority of situations where an atypia is identified through preventive review, voluntary disclosure presents a favourable cost-benefit ratio, provided it is structured with the assistance of an experienced practitioner.

C. Practical recommendations: building a durable compliance strategy

Document in real time rather than reconstruct after the fact. The golden rule, for any taxpayer whose income or wealth reaches a significant threshold, is to document each atypical operation at the moment it occurs. A sale between relatives, a current account contribution, a transfer of residence, a Dutreil arrangement, none of these should ever rest on the sole memory of the protagonists. The internal note, the operation memorandum, the preserved email exchange, the contemporaneous attestation from the chartered accountant or lawyer, all of these are elements whose evidentiary value increases with the age of the operation. The cost of this documentary discipline is negligible relative to the security it provides.

Map one's own atypia signals. The second useful reflex is to maintain, for each financial year, a summary mapping of the configurations likely to produce a signal in the grid described above. This mapping has no vocation to be exhaustive or to conceal anything: it serves to identify the points on which a request for justifications could be addressed and to verify that the corresponding documentation is available. For a taxpayer holding a patrimonial holding, pursuing an independent activity and receiving foreign-source income, this mapping fits on one page but saves hours of reconstruction under the pressure of a procedure.

Anticipate rather than endure the evolution of the tools. The technological trajectory of the DGFiP is clearly oriented: the share of algorithmically targeted audits will continue to grow, the integration of generative AI in the instruction of files is already underway as revealed by the Senate report on AI in the public service (thematic report no. 491, March 2024), and the external sources integrated into the model are enriched at each exercise (DAC 8 on crypto-assets in force since 1 January 2026, interconnected European registers of beneficial owners, automatic exchange between social and tax administrations). The taxpayer who anticipates this trajectory by structuring compliance, rather than reacting on an ad hoc basis to administrative questions, benefits from a durable advantage.