Paris concentrates several industries that rarely share a city: the largest banks and insurers on the continent, the luxury groups, aerospace and defence, energy, and a very substantial public sector. Each produces a different AI problem, but the constraint they hold in common is that the output has to be explainable in French, to a French counterparty, sometimes under a regulator's eye.
Banking and insurance arrive with decisioning and with documents, usually both at once. A file is read, a judgement is made, and eighteen months later somebody asks how. The engineering that matters is the trace rather than the model, which is the whole subject of our risk and compliance architecture and of credit decisioning.
Luxury and retail groups have a harder version of an ordinary problem. Product search has to work across languages and across a vocabulary that the house itself defines, and the tone of any customer-facing text is a brand asset that a generic model will quietly destroy. That is a retrieval, ranking and evaluation problem before it is a generation problem — see merchandising and search.
Aerospace, energy and the public sector tend towards the deployment question first and the use case second. If the conclusion is that nothing may leave a controlled environment, then the model has to come to the data, and the sizing conversation becomes real: what runs on what hardware, at what cost, at what quality. Deployment options lays out the choices without pretending they are free.
French-language work also affects model selection. Open-weight families developed in Europe are frequently strong on French and can be self-hosted, which makes them a serious candidate rather than a patriotic one — our notes on Mistral models cover where that holds and where it does not.
