Vital City asked Dave Cole, New Jersey’s Chief Innovation Officer, what his office is made of. “About 80 full-time, plus contractors. The core disciplines are product management, human-centered design and research, content design and software engineering.”
That list is the AI strategy. Three of those four disciplines exist to decide what should be built and whether it worked, and only one of them builds anything. However, most governments adopting AI have engineers, when they have any, and nothing else on that list, often on contract, which leaves the deciding to whoever sells them the system.
Cole’s office has “avoided” public-facing chatbots, finding them more useful internally “where staff have enough domain knowledge to recognize when outputs aren’t quite right.” That is not caution about models. It is a judgment about who is in the room when the answer is wrong, which is what a digital team is paid for. But no state CIO was ever fired for employing too few product managers.
One can procure a system, but cannot procure the function that decides what it is for, because that function is the client. So here is a cheaper readiness test than any maturity index: count the disciplines on your payroll paid to decide what a service should do, rather than to build or buy it. New Jersey lists three. If your list is empty, the vendors are already rubbing their hands.
Assorted links
- Colorado rebuilt its IT office around product delivery: 173 layoffs against ninety-odd new specialist roles: these disciplines are not additive hires, they replace the generalists you have.
- Anthropic’s Economic Index finds conversations mapped to top-wage-tercile occupations burn 2.07 times the tokens of bottom-tercile ones, so a flat per-seat license misprices AI against the salary it stands in for.
- METR’s “expenditure horizon” asks at what dollar value an agent matches a human on the same budget (so far only on a coding benchmark), the procurement question no business case asks: cost per case, not capability.
- Seventy-one court dockets on algorithmic benefits determinations put the failures where statute gets translated into computational logic, a design job no engineering shop owns.