Insights
What to read before you sign an AI contract
What we learn while building Jean, written for the people who have to sign off on it. Arguments and evidence, not announcements.
More from Insights
Pieces by the people who build Jean, on measurement, regulation and the research we are reading.
// METHOD
What fixing the kilogram taught us
Why an evaluation set that is its own reference drifts, and what we build instead.
- Aiga Andrijanova
- 23 April 2026
- 5-minute read
// COMPLIANCE
Your AI tools are already breaking the law
The EU AI Act puts non-delegable duties on the company using AI, not the one that built it.
- Daniel Deak
- 19 March 2026
- 9-minute read
// RESEARCH
Train the tools, not the agent
Freeze the reasoning core, train the tools around it, and the numbers move a long way.
- Aiga Andrijanova
- 3 February 2026
- 7-minute read
Start with the decision in front of you
Each piece answers one call a board, a risk committee or a CTO has to make about AI. Read the one closest to yours.
-
Whether your AI is an asset or a rental
Sovereign AI is used for three different things: a national model, a hosting arrangement and an ownership question. Only the third changes what your company is worth. The explainer sets out the four tests we run on any deployment, and what an ontology built from your own records needs before it can sit on the balance sheet as an IAS 38 intangible asset.
What sovereign AI actually is -
Who carries the duty when an AI tool breaks the rules
Buying a compliant product from a compliant vendor does not move the burden. Article 26 of the EU AI Act gives the company deploying a high-risk system its own duties: competent human oversight, monitoring against the provider’s instructions, and logs. Article 4 adds AI literacy for the staff who use it. Daniel Deak takes each duty in turn, and the layer most deployments are missing. Our own answer is on the trust and sovereignty page.
Your AI tools are already breaking the law -
How you know the answers are right
An evaluation set built from your own records is its own reference, so it drifts and nobody can tell. For a century the kilogram had the same flaw. Aiga Andrijanova explains how Synthia fixes the truth first, as a graph of facts, then generates a synthetic company around it, so every answer is scored against a key planted in advance.
What fixing the kilogram taught us -
Why a bigger model is not the fix
Researchers compared training a whole agent end to end, which took about 170,000 examples, with freezing the agent and training a lightweight searcher to serve it, which reached comparable retrieval performance on about 2,400. The piece reads that survey, and why Jean keeps a frozen, auditable core with small adaptive mates around it.
Train the tools, not the agent
See Jean in a demo
Thirty minutes with Alex or Daniel, on your systems and your constraints. Bring the hard question.
Book a demo