Weflayr raises $2.5M to route enterprise AI requests to the right model at the right cost

Weflayr (Paris) has raised $2.5 million (approximately €2.2 million) in a Pre-Seed round led by 13books Capital, the London-based early-stage fund, with participation from Founders Future (Marc Menasé), Kima Ventures, Plug and Play Ventures, and Station F Venture. Roxanne Varza, director of Station F, joined as an angel investor. The company is incubated at Station F's F/AI programme, the accelerator's dedicated AI track. It was founded earlier this year by Romain Mazeries (CEO), the former CEO of Mangopay, and Noé Achache (CTO), previously head of AI engineering at Theodo Data & AI.

The routing problem

There are now hundreds of AI models available to enterprises: GPT variants, Mistral, Anthropic, Gemini, open-source models on self-hosted infrastructure, sovereign French models, and an accelerating tail of specialised models built for specific tasks. Each has different strengths, different pricing, different data residency characteristics, and different rates of improvement. Asking a development team to route each request by hand — matching query complexity to model cost to latency requirements — is not a sustainable strategy.

Weflayr inserts itself between the application and the model layer. It tests models against each client's own data, identifies which models perform best for which request types, and then routes each query in real time to the model best suited to it. A simple lookup doesn't require the same model as a complex legal analysis. A customer running under GDPR constraints requires EU-hosted inference; a research team experimenting with cutting-edge reasoning may prefer the frontier. Weflayr handles this without the application needing to know.

The company's first production deployment reduced AI usage costs by more than 55% at equivalent output quality. The mechanism is direct: many queries that enterprises route to expensive frontier models because it is the path of least resistance can be handled equally well by cheaper, faster models the routing layer selects for them.

The Mangopay parallel

Mazeries draws the analogy explicitly: payment networks solved an analogous routing problem in the 2010s — intelligently directing transactions between rails, acquirers, and processors to optimise cost and success rates. The market built sophisticated infrastructure for it; now the same infrastructure problem is appearing in AI at what Mazeries describes as "far higher speed."

The reference point is not theoretical. Stripe's reported acquisition of OpenRouter for approximately $7.5 billion this summer signals what the market assigns to a well-positioned model router with sufficient distribution. Weflayr is not competing directly with OpenRouter's US-market dominance; its differentiation is the European sovereignty layer — compliant data handling, EU-hosted inference options, and the option to route to French models where regulatory or political requirements demand it.

13books Capital — whose second known French bet this is after Aria (Paris, embedded invoice financing, July 2026) — leads. Founders Future and Kima Ventures both have extensive ProYarn-tracked records in French AI and infrastructure investments. The $2.5M funds team growth and platform development, with go-to-market in Europe and the United States targeted within the next 12-18 months.