In 2024, 1.8 million people sent a request to a departmental disability office (MDPH), and the average processing time reached 4.8 months, according to the Key figures on autonomy support 2026 from the Caisse nationale de solidarité pour l'autonomie (CNSA, France's national solidarity fund for autonomy). The fund has tested in MDPH offices an AI responsible for the first stages of processing: splitting the scan of a file into documents, recognising each one, then checking the admissibility of the request, i.e. verifying that it is complete and in order before examining the substance.
According to the CNSA press release of 15 October 2026, the results of this experiment, conducted in 11 MDPH offices with a second use case, a legal chatbot, are "expected mid-October 2026", to be "shared with all MDPH offices". "The evaluation method will also be published", adds the CNSA. As of 20 October, cnsa.fr publishes neither the results nor the method. The fund is entering the final phase of its AI roadmap 2025-2026, of which 29 of 36 actions have been deployed in nine months: "deploy and sustain, starting to generalise the most promising use cases".
The test framework, a public CNSA document, states what the experiment will be able to establish. It limited the AI to paper files, required files of deceased persons as test data and excluded any writing to the MDPH information system. The test therefore measures first the ability of a tool to sort real documents. The guarantees for use on pending applications, from the acceptable error rate to the correction of a document wrongly rejected, remain to be set by the MDPH or the department that will purchase the tool.
At most ten MDPH offices, 250,000 euros and a jury to award the tools
The protocol appears in the adhesion framework of the call for expressions of interest (AMI), the call for applications addressed to MDPH offices, published by the CNSA and dated 4 November 2025. It implements measure 17 of the Tour de France des solutions, launched by the minister delegate for Autonomy and Disability. Its title: "Experiment with artificial intelligence to speed up processing times and implement a dedicated chatbot [...]".
The November 2025 launch mentioned by the CNSA corresponds to this call. The tests were to last three months, "from April to June 2026", and each site was to submit its report "no later than 30 September 2026". If this schedule was met, the CNSA has had the reports since that date.
The framework provides for "a maximum of 10 MDPH offices": four for the legal chatbot, six for document management. The press release counts 11, a discrepancy that no public document explains.
| AMI measure 17 | Legal chatbot | Document management |
|---|---|---|
| Sites at most | 4 | 6 |
| Grant per site, paid in advance | 10,000 euros | 35,000 euros |
| Envelope at most (ActuIA calculation) | 40,000 euros | 210,000 euros |
Across ten sites, the envelope reaches at most 250,000 euros. No publisher is imposed. Each candidate proposes one to three systems per use case, from the market or developed by itself, and the CNSA jury ranks the applications. The candidate obtains its first choice if a higher-ranked candidate has not already obtained it. The framework finally requires a tool hosted in France or Europe and not subject to the US Cloud Act or Patriot Act, which rules out providers falling under those laws.
A chatbot for assessors, an AI that splits and checks files
The legal chatbot, based on generative AI, answers "professionals in charge of assessment" and must "cite precisely the documentary sources" of its answers. It "is not intended to replace human expertise", specifies the framework.
The second use case concerns the processing itself. The tool must separate into documents the single scan of a paper file, recognise the type of each document, name it and classify it. Five documents are expected: the application form (cerfa no. 15692*01), the medical certificate (cerfa no. 15695*01), an identity document, proof of address and, where applicable, the certificate of legal protection judgment.
Then come the "checks to establish the admissibility of the request". The form must be dated and signed, the medical certificate signed by the doctor and dated less than one year ago, the identity document valid. The tool finally extracts the identity and contact details of the applicant, the date of submission and the nature of the request.
An option goes beyond formal checks: the "production of a fraud risk note on each document with alerts". The framework specifies neither its use, nor its recipient, nor the guarantees.
A test on closed files, without writing to the information system
Three clauses separate this test from a deployment. The first keeps the information system where MDPH offices manage files at a distance: "The injection of data into the SIH is not included in the scope of this experiment." The second excludes online applications: "The experiment concerns only scanned paper files." The online MDPH service nevertheless received 241,795 files in 2025, 20% more than in 2024, according to the Key figures 2026.
The third concerns test data. Only MDPH offices able to provide "at least" 100 application files "of deceased persons (not subject to GDPR) from more than 2 years and less than 5 years before the date of the experiment" are eligible. Recital 27 of the GDPR excludes such data from its scope, while leaving member states the option to regulate their processing. The framework does not say whether pending applications were also processed.
On closed files, an error does not delay any application. The preamble sets for the experiment "a clear framework: ethics, data security, mandatory human supervision".
Generalisation: five points the buyer will have to decide
The preamble announces the next step: "Ultimately, if the results are conclusive, it will be generalised across the whole territory." No timetable or criterion is given. The test book and indicators are not published, and the framework sets no success threshold.
The document mentions neither a data protection impact assessment, which the GDPR requires in particular for large-scale processing of health data (Article 35), nor information to applicants. The CNIL (France's data protection authority) appears only regarding the data of applicants to the call. The protocol partly explains these absences: a test on files of deceased persons has no applicant to inform.
On real applications, the GDPR applies to each file. The points left open by the test will become clauses that the purchasing MDPH or department will have to write:
- the error rate per document type, isolating files declared incomplete in error;
- writing to the information system and logging of the tool's proposals;
- correction by an agent when an applicant disputes a document deemed missing or invalid;
- information to applicants and impact analysis;
- activation of the fraud note, its recipient and its follow-up.
Tools of this type are already offered to departments: in May 2025, Provence.ai claimed deployments of its MDPH file verification tool, including fraud detection, in more than 15 departments.
AI Act: a qualification to document, with deadlines in 2027 and 2030
Neither the press release nor the framework qualifies these tools under the AI Act. According to ActuIA's reading, Annex III classifies as high-risk systems used by public authorities, or on their behalf, to assess the eligibility of persons for essential public assistance benefits and services. The rights processed by MDPH offices appear to fall under this.
Article 6, paragraph 3, provides an exit for a narrow procedural task or a task preparatory to an assessment. Splitting, naming and classifying documents resembles the former. The provider must then document its assessment before placing on the market or putting into service, and register the system (paragraph 4).
Two points weaken this exit. An Annex III system that performs profiling of natural persons is always high-risk: the fraud note, if activated, would raise the question. A check that declares a file incomplete also delays the opening of a right, which makes its qualification as a narrow task debatable.
According to the AI Act timeline review published by ActuIA, high-risk requirements apply to Annex III from 2 December 2027. High-risk systems intended for public authorities, including those already in service, must be compliant by 2 August 2030 at the latest (Article 111, paragraph 2). Public bodies deploying a high-risk system must also conduct a fundamental rights impact assessment (Article 27).
In Bavaria, AI checks documents during the online application
On 12 October 2026, the Landratsamt of Kelheim, the administration of this Bavarian district, presented in a press release a pilot project of its foreigners office. The online service Aufenthalt Digital has AI check the documents of residence permit applicants. According to the press release, it must alert them, during the online application, to incomplete files and erroneous or contradictory documents or information. The tool detects, among other things, whether all necessary documents have been uploaded.
Kelheim places AI on the channel that the French test excluded, the online application, and at another point in the procedure: the alert goes to the applicant during their process, whereas the tool tested in MDPH offices works for the agent on files already received. The press release names neither provider nor model and gives no figures.
The adhesion framework of measure 17 runs until 31 December 2026. As of 20 October, neither the list of participating MDPH offices nor the selected tools are published.
Our articles will then appear first in Google Top Stories.