Domain Atlas / Benefits navigation & public-facing chat
Albert France Services
Albert France Services was a sovereign, in-house generative AI assistant built by DINUM with ANCT to help France Services counter advisers answer citizens' benefits and procedure questions from a curated base of official documents, presented by the Prime Minister as a sovereign French AI in April 2024 and, in a demonstration before him, giving a wrong answer on identity-card cost. Piloted from an initial panel of about sixty volunteer advisers to roughly eighty advisers across more than forty counters (forty-eight at final count per AFP) in six departments over three iterated versions, it was, per a January 12, 2026 AFP dispatch, formally not going to be generalized 'in its current form,' a decision DINUM announced on January 9, 2026 while stating that the majority of Albert-brand projects are sustained and fully operational. No error rate, usage volume or override count for the tool was ever published; AFP reports DINUM's annual AI budget at about 1.2 million euros since 2024 with Albert France Services a minimal share, a figure distinct from and not directly comparable to the union Solidaires Finances Publiques' separate claim of a roughly 1.3 million euro project cost.[4]
What happened
Albert France Services was an in-house, sovereign generative AI assistant built by DINUM (the Direction interministerielle du numerique) with ANCT (the Agence nationale de la cohesion des territoires) to help front-line advisers in the France Services one-stop-counter network answer citizens' questions about administrative procedures and benefits. It was a retrieval-grounded assistant: an adviser typed a natural-language question describing a citizen's situation and received a sourced draft answer built from a curated knowledge base of service-public.fr practical sheets and the documentation of the national operators France Services fronts for (taxes, pensions, family benefits, health insurance and others), together with related questions, practical links and simulators. It was advisory only and human-in-the-loop by design: the adviser was expected to verify, modify and validate every generated answer before relaying it to the citizen, and could ignore the tool entirely, which many reportedly did, preferring an ordinary search. From an initial panel of about sixty volunteer advisers, the experiment expanded to roughly eighty advisers across more than forty counters (forty-eight at final count in AFP's January 2026 reporting) in six departments (Vienne, Deux-Sevres, Rhone, Allier, Meurthe-et-Moselle and Var), iterating three successive versions from adviser feedback journals and qualitative interviews. The surrounding network context is large: more than 2,750 counters assisting nearly 800,000 procedures a month.
DINUM presented Albert as the State's free and sovereign generative AI, created by and for public agents, and received the "innovation" prize at the Victoires des Acteurs publics on 7 February 2024. On 23 April 2024 Prime Minister Gabriel Attal publicly launched it at the Sceaux France Services counter as part of a "debureaucratization" roadmap, calling it a "100% sovereign" French AI that would "revolutionize" public services and promising that AI would free agents for higher-value work rather than replace them (the roadmap targeted 3,000 France Services counters by 2027, from a network reporting 96% satisfaction and about 10 million procedures a year). In a demonstration before the Prime Minister, Albert answered that renewing an identity card is free, when the cost that applied in the demonstrated case was 25 euros — an error that became emblematic of the tool's reliability problems and dominated later coverage. (French ID-card fees vary by case: renewal of an expired card is free, while a 25 euro fee applies to a replacement after loss or theft, so this is best read as a wrong answer in the demonstrated case rather than a settled account of the fee schedule; the sources do not hard-date the error to the 23 April launch itself.)
What the record does not contain is any measurement. No error rate, usage volume, adoption frequency or override count was ever published for Albert France Services. In place of an instrument, its failures surfaced through the operator side: several unions documented recurring technical malfunctions and plainly wrong answers, an investigative-television broadcast in April 2025 (per the union Solidaires Finances Publiques) featured unenthusiastic France Services agent testimony, and advisers reported that Albert often answered less well than an ordinary search. According to Solidaires Finances Publiques, the project had in fact been discontinued by September 2025 — a fact the union says it learned incidentally in a ministerial AI-strategy working group where Albert no longer appeared among the presented projects, with no press release or announcement — at a claimed project cost of about 1.3 million euros, and the union called it a top-down failure developed without adviser consultation. These specifics (the September 2025 stop, the 1.3 million euro figure) are the union's claims and are carried as such.
On 9 January 2026 DINUM formally announced that Albert, as experimented at 48 France Services counters, would not be generalized "in its current form," citing the pilot record. DINUM simultaneously disputed the failure narrative, stating that "the majority of experimental projects grouped under the Albert brand are now sustained and fully operational," and AFP reported DINUM's annual AI budget at about 1.2 million euros since 2024, with Albert France Services a minimal share of it (a different and not directly comparable figure to the union's project-cost claim). The broader Albert program survives as interministerial infrastructure: an Albert API model-access point for ministries, aggregated public reference bases (Albert Data), and a conversation product that evolved into a successor adviser tool ("Assistant IA"), in test with about 10,000 public agents across ministries through June 2026, that integrates models from the French vendor Mistral AI. A full evaluation of that large-scale experimentation is due in summer 2026 and must notably establish the cost of a generalization — an explicit cost-accounted go-or-no-go gate following the France Services withdrawal. Alongside the January 2026 decision, DINUM migrated the Albert API's model aliases away from the "albert-" branding and removed the web-search functionality, retiring the legacy aliases by 15 February 2026 — an infrastructure-level rescoping accompanying the front-line product's withdrawal. As of mid-2026 the summer evaluation had not yet published, so the successor's fate remained open. The sovereignty and "free and open" descriptions throughout are the government's own characterizations, and no AI model identifiers appear here; the successor's model vendor is named (Mistral AI) but no specific model is.
The sociotechnical reading
The Atlas already carries the two cases this one sits between, and it is unlike both. The NYC MyCity chatbot was told it was wrong, in public and then in a formal audit, and kept answering anyway: exposure is not correction. GOV.UK Chat built a staged gate that could say "not yet" and published its whole accuracy learning curve, gate by gate: a gate is only a control if it can fire. Albert is a third thing. It was neither kept online in defiance of a finding nor gated on a published number, because there was no number. No error rate, no override count, no usage figure was ever measured or released. That absence is the case. Strip out every metric a governance story usually turns on and ask what is left to detect failure, and the answer here is the only one left: the workforce. Advisers who found the tool worse than an ordinary search, unions who wrote the malfunctions into a dossier, and one televised wrong answer to a Prime Minister were the entire error-detection apparatus. The distinct lesson is that when a deployment ships with no instrument pointed at itself, its operators become the sensor of last resort — and that is a detector which fires slowly, in public, and last. Albert was not overruled by an audit; it was walked away from, and the walking-away, aggregated by a union, was the evaluation.
Two structural features make that legible and separate this case from the verify-before-use copilots it superficially resembles. First, the override cost was near zero, and it cut both ways. Because the tool decided nothing and an adviser could simply ignore it, contamination was bounded whenever advisers declined a wrong answer — but the same near-zero cost is exactly what let adoption silently collapse. The copilots in this domain worry about a caseworker over-trusting a fluent answer; Albert's advisers under-trusted it and abandoned it, and abandonment leaves no error log. A tool that is safe because people can ignore it is also a tool that can die without anyone deciding to kill it. Second, the store was protective, so the usual contamination loop never had to be the story: the assistant answered only from a curated official base it did not write back into, and every answer pointed the adviser back to the cited source. That is why the failure that mattered was not a poisoned record but a legitimacy set by a single demo error and a workforce that quietly opted out. The governance question the case makes precise is not "which control failed?" but "what do you build when nothing was ever watching?" The leverage sits off the model entirely: a review cadence that is the first instrument the tool ever had, an external go-or-no-go bar so a version clears a standard before a Prime Minister launches it rather than after, and a standing check inside the system to replace the reputational one that fired too late. The coda is the sharpest part. The program did eventually install a real gate — but only for the successor, and the number that gate turns on is not an accuracy figure. It is a cost. The one metric this metric-less deployment finally produced is the price of scaling the next tool, which is an honest place to end and a telling one: the thing that governs an AI adviser here in the last analysis is not how right it is, because that was never measured, but what it costs.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.