Albert API: a government gateway that gives public teams secure access to generative AI models

The Problem

Adding generative AI to public digital services calls for costly, complex infrastructure. The Interministerial Digital Directorate (DINUM), the French government's central digital agency, has pointed to the barrier this creates for administrations wanting to use the technology, and to the risk of separate teams building the same things twice.

The running costs are also significant. Spending on servers, the specialised chips that power AI, model licences and upkeep grows in line with use, and without pooling, each ministry would have to fund such infrastructure itself.

Control of data was a further concern. The developers of the software behind the service, a DINUM project, set out to cut the cost of commercial AI services, avoid being tied to one supplier and let public bodies retain authority over confidential information.

The Solution

The DINUM runs Albert API as a single service that any public team can plug into. API stands for application programming interface, a standard way for one computer system to ask another for a service. In practice, a team building a digital tool, such as a complaints system or a document search, sends its request to Albert API, which passes it to an AI model and returns the result, so the team can add AI without running its own infrastructure. Behind that single point of entry sits a choice of models, ranging from commercial products to openly published models that can run on servers the government controls. They currently come from companies including Mistral and OpenAI and handle tasks such as drafting text, writing computer code, transcribing recordings and reading text in images. The software that connects teams to the models, OpenGateLLM, was written by the French state and published for anyone to reuse, so another government could start from the same code.

Teams can also draw on shared features that many projects need, so each one does not have to build them from scratch. Two examples are turning scanned documents into text a computer can read, known as optical character recognition (OCR), and letting the AI look things up in a team's own documents before answering, known as retrieval-augmented generation (RAG). The service also accepts requests in the same format as OpenAI, the company behind ChatGPT, so developers can connect tools they already use.

The models and the service run on servers supplied by a hosting company, Outscale, whose cloud holds SecNumCloud certification, a French security label for cloud hosting. According to the DINUM, the content of requests sent to the models is not stored, and user data is never sent out over the open internet. The DINUM has also published the formal security approval issued for the system.

Officials anywhere in the state civil service can experiment at no cost. Teams that need higher usage limits can request a "limited production" allowance. In December 2025, the DINUM sent other ministries a note proposing that they co-fund the infrastructure. Ministries that join as partners get higher limits, personal support, every available model and a service-level agreement, a formal commitment on availability.

Small product teams from beta.gouv, the DINUM's programme for building digital public services, could use Albert API freely. After five months of trials, they met to share what had worked and what had not.

The Impact

1. Growing use across government
At the five-month review in April 2025, the service was handling about 50,000 requests a week for 62 products. The DINUM's product page, which carries no date, now reports more than 70 public projects and over 100,000 requests a week.

2. Early uses
Teams at the April 2025 review described several products built on the service, some of them still being tested:

• The council that evaluates French schools used it to search its evaluation reports, of which it produces around 12,000 a year and holds 60,000, so it could spot trends and draft summaries. The product was still in testing.
• SignalConso, which handles complaints from consumers, used it to summarise complaints and to check, against consumer law, whether the national consumer protection authority was responsible for each case. A service that reviews quotes for home energy renovation used it for the first step of turning about 1,300 quotes, sent in by 1,000 users, into text.

3. Limits found in early use
The same review recorded problems the platform did not solve on its own. The schools council saw error rates of about 20% on some extractions, such as page numbers, and the SignalConso team flagged invented content as a risk in legal analysis. Teams also reported unstable performance on one hosting provider, weak text recognition on poor scans, and a context window, the amount of text a model can take in at once, too small for some analyses.

Lessons Learned

• Pool the costly layer and let teams build on top of it. Shared infrastructure spares each organisation from buying and running its own computing capacity. The DINUM's case for co-funding is that a pooled service uses public money more efficiently than ministries each running their own, and lets them put their effort into their own services.
• Follow widely used technical conventions. Matching an interface that developers already know lowers the effort of switching to a government platform. Albert API's compatibility with OpenAI's standards means existing development tools connect to it.
• Publish data rules and security approvals where teams can find them. Teams handling sensitive information need clear answers on where their data goes. The DINUM's infrastructure page links to its security approval and data-retention documents, and explains why the platform itself does not need certification as a health data host.
• Collect structured feedback from early users. Asking pilot teams to report back reveals where a shared service falls short. The April 2025 review produced a list of priorities for improvement, among them steadier performance, stronger text recognition on complex documents and a better grasp of administrative and legal wording.
• Test every use, even on a shared platform. A common base does not remove the weaknesses of the underlying models. Teams building on Albert API flagged invented content as a risk in legal analysis and recorded extraction errors, so each product still needs its own checks.