Outcomes-Based AI: Start With What Government Needs to Achieve

I want AI to help government accomplish more, and I want us to be specific about what should improve. An agency can produce more analysis and answer more questions without resolving the problem that brought someone to government in the first place. Before choosing where to use AI, I want us to name what should get better.

In Extending the Reach of Government, I argued for expanding what existing teams can accomplish. The next question is where to direct that capacity. I call this outcomes-based AI: starting with a public outcome, identifying what stands in its way, and choosing and evaluating AI by its contribution to that outcome.

In housing permitting, I want a homeowner to get a sound decision sooner so construction can begin. Someone seeking permission to add a home on their property first has to understand the application process and the requirements for their project. Missing information triggers corrections; a resubmission that resolves only some of the comments triggers another round.

Give AI a specific job: guide the applicant through current requirements, flag missing information, and help them address each comment before resubmitting. Staff can use it to compare revisions with outstanding comments and prepare material for review, while remaining responsible for the permit decision. The aim is for applicants to spend less time resubmitting paperwork and staff to have more time to review the project itself.

I would judge the tool by whether applications arrive complete, applicants resolve review comments the first time, and the city reaches a sound decision sooner. Those measures track the obstacle we set out to remove. Counting AI-generated answers or reports tells us little about whether housing is getting through review sooner.

If complete applications still sit waiting for specialist review, that wait needs attention too. And a permit is not a completed home: financing and construction still have to follow. Outcomes-based AI requires us to be precise about what an agency can influence and to follow the work far enough to see whether the improvement matters.

As the former Chief Innovation Officer of San Francisco, I brought in help because there was public work we needed to get done. With AI, I want to keep that purpose in view. A homeowner able to move ahead with construction gives me a reason to value the tool; a growing count of automated responses does not tell me enough. If the homeowner is still caught in the same rounds of corrections, I would want to know why before calling the tool a success.

Written with AI assistance. Final editorial decisions are my own.

I work at OpenAI. The views expressed here are my own and do not necessarily reflect those of my employer.