On the record
On the record.
Plain notes on governing AI at work, from gopal joshi. What the risk really is, how enterprise AI differs from the chat tools you already use, and what an Inspector does about it.
on the record · Oct 2026
The model is now a dropdown. The proof is not.
Enterprise AI platforms have made switching models almost free — one account, one API, every major provider behind it. What they have not made free is knowing the new model behaves the same on your work. Here is why that gap is wider at work than in the chat tools you already use, and the test to run before the next switch.
on the record · Sep 2026
Your doctor pressed record. Where did the conversation go?
An AI scribe in a clinic is the clearest example of a question every enterprise AI system has to answer — where does the input go, how long does it stay, and who can read it. Here are the words you will hear, in plain English, and the question behind each one.
on the record · Sep 2026
You’ve been briefing a consultant. Most decisions needed an instrument.
For four years, every AI decision in your business has been made the same way: ask in words, get words back, have someone read them. A second shape of model has arrived. Here is what it is, where each one fits, and what it does not fix.
on the record · Sep 2026
Confidence is the new bottleneck
Ten years in enterprise AI pre-sales — the technical chair, not the selling one — and what replaced the education problem. Literacy got solved. Permission inverted. Calibration got worse, and that is now the expensive one.
on the record · Aug 2026
Copilot is licensed. That doesn't mean it's governed.
A Copilot licence is a buying decision, not a governance one. The gap between having Copilot and proving what it touched is where the risk lives.
on the record · Mar 2026
Before you govern AI, you have to see where it already lives
Most enterprise AI risk isn't in a lab. It's in the ordinary workflows people already run. How to find it before you decide what to do about it.
on the record · Feb 2026
What an "LLM checker" actually has to do in a regulated workflow
Checking an LLM's output is the easy 10%. The real job is proving what went in, what came back, and that nothing left your walls — on record.
on the record · Jan 2026
Where AI already lives in your enterprise — and what you can't prove about it
Before you govern enterprise AI, you have to find it. It is rarely where the org chart says, and most of it runs today with no record at all.