Insights
Notes on governing AI in regulated work: where controls belong, what evidence should look like, and the design choices behind Synainesi.
-
Model independence is a governance property, not a feature list
Supporting many AI models is easy to claim. Being able to approve, restrict, compare and switch them off under one set of rules is what regulated organisations actually need.
-
Evidence without surveillance: what an AI control should record
Oversight of AI use needs evidence. It does not need a central copy of everyone’s prompts. The difference shapes what a well-designed control keeps, and what it deliberately does not.
-
Agents have permissions, not trust
An AI agent should be able to do exactly what it was allowed to do, and nothing its own output talks it into. That principle is simple to state and demanding to…
-
Pseudonymisation is not anonymisation: what tokenising a prompt does and does not do
Replacing names and account numbers with tokens before a prompt leaves the device is a useful protection. Calling the result anonymous is a mistake, and in regulated work an expensive one.
-
A setting is not a control: where AI governance has to be enforced
Most AI policies are enforced by the same interface they are meant to govern. For regulated work, the rule has to be applied somewhere the interface cannot switch off.
Expression of interest
Be first to shape governed AI.
We are opening Synainesi to a first group of regulated teams. Tell us about yours.