Put AI into a mid-market firm without staying at PoC
A narrow use case, clean data, EU hosting, a run budget. Without that, the copilot stays a slide.
Most AI projects we see did not fail on the model. They failed on the rest: nobody chose a file type, the corpus is a mess, hosting was “we’ll see”, and the run cost was never written.
Start with one file
Not “the company”. One process, one team, one success metric. Support tickets, a finance pack, an ops checklist. If you cannot name the Monday-morning action, it is still a demo.
Data and access
RAG on a dump of Confluence is not a product. Rights have to match the directory. Traces have to exist. Compliance, including GDPR when it applies, is not a footer.
Hosting and money
EU region if that is a criterion. Inference as a run line, not a surprise. If the unit cost only works in a slide, it will not survive the first invoice.
Then someone has to stay until people use the thing. That is often the FDE part.