MAIF is not only presenting three new technological partnerships. The mutual is setting up an organization intended to bring artificial intelligence into its business processes, without abandoning control of its data, its rules and its skills. His case provides useful benchmarks for any company seeking to move from experiments to a sustainable AI strategy.
Finding a guarantee, reconstructing the history of a claim, preparing a response to a complaint or summarizing an appeal: it is in these operations that artificial intelligence is beginning to modify the work at MAIF.
The partnership announced in September with MISTRAL AI, SCALEWAY and DIABOLOCOM must therefore be read as the next stage of a long-standing program. SCALEWAY provides infrastructure capacity, MISTRAL AI, models and support for generative uses, DIABOLOCOM, a platform designed to gradually structure interactions with members. The idea is not to add these building blocks, but to fit them into the same strategy: who chooses the use cases, who verifies the responses, what data can circulate, who arbitrates the risks and how does the company retain the possibility of changing supplier?
A strategy that is based on history, not on an announcement effect
In 2020, MAIF already presented Mélusine, an internal tool for analyzing and directing incoming emails. Its commercial activity also relies on ZELROS, tested in 2017 then deployed in 2023 to help advisors formulate appropriate recommendations, particularly around the insurance of electric vehicles.
Created in 2017, the Data Factory was followed in 2022 by a Data and AI department which brought together 240 people in 2024. Good practice is that before generalizing assistants, a company must have a place capable of linking business expectations, technical architecture, security, legal and value measurement. Without this coordination function, each department buys its tool, supplies its own data and invents its own rules.
The MAIF portfolio now covers several types of AI: classification of requests, recommendation, documentary research, generation of reports and writing assistance. The same file can tomorrow mobilize several of these tools. The issue then becomes less the isolated performance of each person than their consistency: do they use the same version of the contractual rules? Are the access rights the same? Does the advisor know what information was used and can they correct the result?
Buy capacity, keep business assembly
MAIF chooses external partners without renouncing the internalization of the core business. At the end of 2025, the General Director in charge of IT, Nicolas Siegler described a hybrid architecture, combining two internal data centers, Clever Cloud, Azure and largely containerized applications under OpenShift. It also indicated that the application core, still partially in Cobol, was being modernized towards Java and microservices. Management then aimed for an increase in the share of internal skills, from 50% to 75%, a more realistic distribution than an opposition between “doing” and “buying”.
A company can purchase a cloud, a language model, or a customer relations platform. However, it must keep control of the data that the model accesses, the rules that govern the response, the interfaces with business applications and the validation criteria. It is this layer of orchestration that transforms a technological capability into its own asset.
For general management, the rule is simple: any AI contract must specify what remains under internal control, this includes the selection and updating of documentary corpora, the management of identities and authorizations, activity logs, quality tests, human escalation rules and the ability to replace a model. The supplier delivers a technology, when the company remains responsible for the decision made to the customer.
Putting documentary history at the center of generative AI
The MAIF identifies three generative projects in the process of being generalized: internal knowledge research, call reports and assistance in responding to complaints. At the end of 2025, it also identified around fifty experiments and around 10,000 employees had access to Microsoft Copilot.
Documentary research constitutes the most revealing test. The approach adopted combines research in knowledge bases with the generation of answers, according to the RAG principle. In insurance, a fluid response is not enough and must be based on the correct version of the contract, integrate the applicable exclusions or not confuse a general rule with the situation of a member.
History management therefore becomes a subject of governance, and not just documentation. A company must designate the owners of its content, trace their versions, organize access rights and remove outdated documents, it must also be able to show the employee the sources which form the basis of a response. This discipline simultaneously serves operational quality, compliance and team trust. It ties in with the broader issue of datasets as strategic assets for companies deploying AI.
Make customer relations a cross-functional project
The progressive deployment of DIABOLOCOM also occurs in a migration context: MAIF had to upgrade its Genesys solution after the discontinuation of its on-site version. A modernization which is therefore not exclusively triggered by AI, but also responds to a supplier life cycle constraint.
The contact platform opens up possibilities for continuity between channels, qualification of requests, assistance to advisors and, ultimately, better knowledge of the history of exchanges. But it involves marketing, customer relations and operations as much as the IT department. A client should not have to repeat their situation because they move from phone to mail or digital space, and an advisor should not receive a summary that erases important exceptions.
The practice to remember is to treat this type of deployment as a journey transformation. Metrics should not be limited to automation rate or average call duration, but should also cover first contact resolution, callbacks, corrections, complaints, customer understanding and actual time spent on complex situations.
Sovereignty: moving from intention to reversibility
MAIF has added sovereignty to the criteria of its IT purchasing policy, alongside performance, security and reliability. If this orientation is consistent with the sensitivity of the data processed by an insurer, it cannot however be reduced to the nationality of a supplier or the place of accommodation. The choice of SCALEWAY comes as the French player itself seeks to build a European cloud platform covering infrastructure, data and AI.
The right question is that of effective reversibility. Where are documents, queries, responses, audit trails and backups stored? Can the data be exported in a usable format? How long does it take to change models or platforms? What internal skills should be mobilized? The MAIF specifically includes the inability to quickly replace a service provider or reinternalize a service among the criticality criteria of its subcontracting. This logic aligns with DORA’s requirements in terms of operational resilience and management of ICT service providers.
For companies structuring their AI purchases, reversibility should be designed and tested from the start, with contract clauses, export formats, onboarding documentation and exit exercises.
Organize human control and skills development
At the beginning of May, the MAIF concluded an agreement with its six trade union organizations excluding economic layoffs linked to the deployment of AI. The text provides for training, priority reinvestment of efficiency gains in service to members and professions, as well as an AI commission of twelve members attached to the CSE.
This framework provides a useful response to an often underestimated difficulty: automating a task can also eliminate a learning moment. For assistance in drafting responses to complaints, MAIF has therefore chosen not to immediately give the tool to novice employees who have not yet mastered the task without assistance.
Human validation is only valid if it is practicable; the company must define what must be controlled, give the employee the necessary sources, allow time for verification and follow up on the corrections made. It is both a requirement for quality, legal responsibility and maintaining skills.
Measuring value at the process level
The MAIF thus gives a more interesting reading of AI than just the race for models. Its potential advantage will not come from its suppliers taken separately, but from its ability to organize them around its business rules, its documentary heritage, its obligations and its teams. For companies scaling up, the lesson is central: AI becomes strategic when it is governed as a complete operational transformation (technological, legal, marketing and human) and not as a succession of tools.
It now remains to measure this contribution over time. The next step will be less the announcement of new partners than the publication of indicators capable of linking the investments made to the quality of service, processing times, skills preserved and the gains actually obtained. It is on this ground that the MAIF strategy can be evaluated.
