The honest assessment, three years in
Of the 1,000 AI projects launched in France in 2023, roughly 30% made it to production. Of those in production, about half generate a measurable, defensible ROI. That’s about 15% of the initial projects. The rest were either dropped or turned into a gadget no one uses.
Not a failure. But not the promised revolution either. Above all, the success rate varies dramatically depending on the use case you pick.
What works, and works well, in 2026
- Internal document search. The simplest, most profitable, most mature case. 90%+ success rate when the documentation is properly prepared.
- Writing assistance on quotes, proposals, and reports, drafted from internal examples.
- Support deflection on recurring questions (enriched FAQ, page citations).
- Document analysis at scale (audit, litigation, compliance).
- Code assistance in technical teams, now standard practice for most developers.
What still doesn’t work
- Fully autonomous agents running chains of actions without supervision. Brilliant demos, fragile production.
- Marketing content that “converts.” Models write correct copy, rarely memorable copy.
- Deep strategic analysis. Good synthesis, hollow reasoning the moment you step off well-trodden paths.
- Full replacement of a job. In practice, wherever it’s been tried, humans have been repositioned, not replaced.
The gray zones where value depends on context
- Professional translation. Excellent for routine content, still behind the best human translators on high-stakes text.
- Image analysis. Very good at classification, fragile on contextual interpretation (medical files, judicial expertise).
- Multimodal (seeing, reading, writing in a single conversation). Spectacular in demo, not yet mature in stable production. Most cases can be handled in text/document mode — multimodal in a second wave.
The 2026 shift: everyday productivity
What changes in 2026 isn’t the rise of spectacular use cases. It’s the quiet spread into daily work. Finding a note, drafting a reply, retrieving a precedent, summarizing a long email. The individual gain is 30 minutes to 2 hours per day, no headlines.
For French businesses, the question isn’t “should we deploy” anymore. It’s “how do we scope it so it works, without putting data at risk.” See How to succeed with your AI assistant project.
For the cost-side implications, see Generative AI and operating cost reduction.
Twenty minutes to identify the two or three cases where generative AI will really deliver value in your organization, and rule out the ones that remain fragile.
Book a demo→