Publishing your denominator costs nothing and almost nobody does it. A number with no sample size beside it is not evidence. It is marketing that survived a meeting. https://t.co/j3K3YDSFeJ
11.17%. That is the share of the bench the Québec court ran its AI pilot on. 22 volunteer judges. It published the percentage. It did not round up, and it did not call 22 judges "the judiciary". Then it published that 19 of the 22 answered the final survey.
Every approval figure in that report arrives with its own denominator, so you can discount it yourself. Compare what reaches a board in financial services. A percentage. No sample, no date, no conditions. Usually the vendor's, usually the best cohort.
It is less, because every decision is small and answerable. Banks keep buying the one general copilot, then spend a year trying to write a control for it. The control was the scope. You chose it at procurement. https://t.co/j3K3YDSFeJ
Nine agents. Not one assistant that does everything. Nine narrow ones, each scoped to a named task. The Québec court chose specialisation over capability. That is why it could say what each one was allowed to do. A general assistant has no boundary you can write down.
Its scope is whatever the user asks for today. You cannot audit that, and you cannot refuse it, because there is no list of permitted work to check against. Nine narrow agents give you nine scopes, nine owners and nine places to say no. Looks like more governance.
Drifting into production is not one of them. So: what is the date, and who decides on it? If nobody can name both, that is not a pilot. It is a rollout with a softer word in front of it. https://t.co/j3K3YDSFeJ
98 days. The Québec court's AI pilot had an end date before it had a result. 8 December 2025 to 16 March 2026, fixed at the start. Most AI pilots have a start date and a hope. An end date is a control, not an administrative detail.
A pilot that can run indefinitely never has to answer the question it was created to answer. It just quietly becomes the system. The date forces a choice. Expand, stop, or run it again with the conditions changed. All three are legitimate.
Great to chat with experienced Founder & CTO @martinjwells on his exp in building companies to +50M users, having a #1 app in 22 countries and his decision making process when selling 3 of his companies https://t.co/cgDZeM5LIl #startupaus#podcast#startup#tech#business#growth