Albert France Services

Albert France services was a generative AI assistant piloted at France Services, France's one-stop public service counters. DINUM, the French state's digital directorate, built it with ANCT, the national agency for territorial cohesion. Advisers typed a question about a citizen's procedure or benefits, and it drafted an answer from official documents, citing its sources, for them to check.

How it was used

France Services is France's network of one-stop counters. There, advisers help citizens with administrative procedures and benefits. The network has more than 2,750 counters, assisting nearly 800,000 procedures a month.

Albert searched a curated base of documents. They were practical sheets from service-public.fr, the government's public information site, and documentation from the national operators France Services represents. These cover taxes, pensions, family benefits, health insurance, and other services. Practical sheets are the site's information pages on each procedure. Each answer came with related questions, practical links, and simulators, which are online calculators.

Albert was advisory and decided nothing. The adviser was expected to verify, modify, and validate every answer before relaying it. The adviser could also ignore the tool entirely, and many reportedly did, preferring an ordinary search.

The pilot

The pilot began with a panel of about sixty volunteer advisers. It grew to roughly eighty advisers at more than forty counters in six departments, France's administrative areas: Vienne, Deux-Sèvres, Rhône, Allier, Meurthe-et-Moselle, and Var. AFP, the French news agency, counted forty-eight counters in its January 2026 reporting.

A joint DINUM and ANCT experiment team produced three successive versions from advisers' feedback journals and qualitative interviews.

The launch and the error

DINUM presented Albert as the State's free and sovereign generative AI, created by and for public agents. Sovereign here means under French state control. It won the innovation prize at the Victoires des Acteurs publics, a public sector awards event, on 7 February 2024.

On 23 April 2024, Prime Minister Gabriel Attal launched it at the France Services counter in Sceaux. He called it a 100 percent sovereign French AI that would revolutionize public services. He promised that AI would free agents for higher-value work, not replace them. The sovereignty descriptions are the government's own.

In a demonstration before the Prime Minister, Albert answered that renewing an identity card is free. The cost that applied in the case shown was 25 euros. The error became emblematic of the tool's reliability problems. The sources do not date it to the launch itself.

What was never measured

No error rate, usage volume, adoption frequency, or override count was ever published for Albert France services. An override is an adviser setting aside the tool's answer. The sources describe no instrument that measured whether its answers were right.

How its failures came to light

Its failures came to light through the people who used it. Several unions documented recurring technical malfunctions and plainly wrong answers. Advisers reported that it often answered less well than an ordinary search. According to the union Solidaires Finances Publiques, an investigative television broadcast in April 2025 featured unenthusiastic testimony from France Services agents.

The case file reads this as the case's lesson. When nothing measures a tool, the people using it become the error detector of last resort. That detector works slowly, in public, and late.

How it ended

According to Solidaires Finances Publiques, the project had stopped by September 2025, with no announcement. The union says it learned this in a ministerial AI strategy working group, where Albert no longer appeared among the projects presented. It puts the project's cost at about 1.3 million euros. It calls the project a top-down failure, built without consulting advisers. These are the union's claims.

On 9 January 2026, DINUM announced that Albert, as tested at 48 counters, would not be generalized in its current form. Generalized means extended across the whole network. DINUM cited the pilot's record. It also disputed the failure framing. It said most of the experimental projects grouped under the Albert name are sustained and fully operational. Albert is also the name of the state's wider AI program, described below.

AFP reported DINUM's annual AI budget at about 1.2 million euros since 2024, with Albert France services a minimal share. That figure is not directly comparable to the union's cost claim.

What came next

The wider Albert program continues as shared government infrastructure. It includes the Albert API, an access point to AI models for ministries, and Albert Data, which gathers public reference bases. The state's successor tool is Assistant IA, French for AI assistant. It grew out of a conversation product in the wider Albert program. It is a newer adviser tool for agents across ministries, tested with about 10,000 of them through June 2026. It uses models from the French vendor Mistral AI.

A full evaluation of that test is due in summer 2026. It must notably establish what extending the tool across government would cost. As of mid-2026 it had not been published.

Alongside the January 2026 decision, DINUM removed the web-search function from the Albert API. It retired the old Albert-branded model names by 15 February 2026.

What this case asks

The usual question is which control failed. Here the question is what to build when nothing was watching. The case file points away from the assistant itself. It names a review rhythm, an outside bar a version must clear before launch, and a standing check inside the system.

What this network is drawn from

This network follows the pattern the case file describes. It does not reconstruct the actual tool or any of its versions. It shows the assistant, its search, the knowledge base, the advisers, the experiment team, and the union and press reports. Citizens are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A closed pathway is one that mistakes stop passing along. The work along it may go on.

This case has a budget of 8 units. Explore (No Targets) sets no targets. Under Service Targets Only, the targets are met before any tool is used or pressure added. Every combination of tools within the budget keeps them met.

Service and Safety Targets and All Governance Targets also ask you to close every failure pathway. Before any tool is used, one is open: Draft answers to advisers. The targets can be met at both levels within the budget.

Escalate checks is the only tool on offer that closes that pathway. Once trouble is flagged, it has advisers check the assistant's answers more closely against the sheets they cite. The sources describe no monitoring on this deployment to flag it. On its own it meets the targets at both levels for 2 units. Every combination that meets them includes it.

All Governance Targets also asks for the assistant to be clearly helping the work. A few combinations with Escalate checks fall short of that. One is Review on schedule, Understand the system, and Escalate checks, which costs all 8 units.

Understand the system costs 3 units under Explore (No Targets) and Service Targets Only, and 4 under the two higher levels. Its stronger setting costs 6 units. While it is on, Vet connections, Keep prompts neutral, Keep skills sharp, and Upgrade model each cost 1 unit less. At its stronger setting they cost 2 units less, never below 1.

More is not better here. Every tool at its strongest setting costs 31 units, nearly four times the budget. That meets the targets at none of the three levels that set them, because the assistant then adds too little to the work.

No tool on offer adds the check this deployment lacked: an error measurement or independent accuracy audit before each version.