I have been skeptical of custom software development in government for a long time, and that skepticism comes from experience. When I worked for Sacramento County's Department of Human Assistance, we served as a pilot county for CalWIN, the predecessor to CalSAWS. The system was not ready, but the pressure to go live was intense. Our county pushed back and won a one-year delay. In hindsight, that still was not enough. During go-live, we issued a payment to a foster care provider for $84 million but fortunately stopped it before it was mailed. Imagine the headlines if we had not caught that.
We launched CalWIN knowing the system had defects, because waiting indefinitely was not a realistic option. What we underestimated was how hard it would be to fix things while flying the plane. That experience, compounded by similar ones across state and county government, shaped my caution about custom development. But recent conversations have given me genuine reasons to see this differently, and I think those reasons are worth sharing.
WHY AI CHANGES THE CALCULUS
State CIOs I respect have pushed back on my reservations in specific and credible ways. One told me his shop already has the internal capability to do its own programming and has built a structured approach to leveraging AI tools responsibly. Another described work that previously took months — writing code, reviewing it, testing it, documenting it — now taking hours with the help of agentic AI, which accelerates not just the coding, but the review process alongside it.
What makes this compelling is that the problems with CalWIN were not failures of ambition. They were failures of process under time pressure. Data conversion done hastily. UAT scripts that were incomplete. Documentation that never got written because the team was always moving to the next deliverable.
Agentic AI directly attacks those bottlenecks. It can generate and validate test scripts at a scale no human team could match, document code as it is written and flag data mapping problems during conversion rather than after go-live. For the phases of legacy modernization where time constraints and resource pressures have caused so much damage, agentic AI represents a real opportunity and shifts what is achievable.
BUILD THE BLUEPRINT FIRST
My optimism comes with one non-negotiable condition. Before any department pursues AI-assisted development, it needs a blueprint: a clear picture of what the system ultimately needs to do, how the pieces fit together and what gets built in what order.
I do not mean a rigid project plan. Think of it like the foundation and wiring of a building. The room layout can change as work progresses, but the load-bearing walls cannot move after the fact. That blueprint, owned and understood by the department's program and IT staff rather than the vendor, is what allows agentic AI to accelerate the right work rather than accelerate in the wrong direction.
That ownership question is also where my optimism runs into a hard reality. Some state IT shops are genuinely ready for what AI now makes possible. Most are not, and the danger is that departments start generating code faster than their staff can review it, test it or understand what they have built. Code that cannot be reviewed, tested or maintained by the people who inherit it is not an asset. It is a future liability.
GETTING THIS RIGHT MATTERS MORE THAN GETTING THERE FAST
For departments with mature IT leadership and the discipline to build incrementally against a clear blueprint, agentic AI opens possibilities that were not practical even a few months ago. For those departments, AI makes legacy modernization faster and less expensive without forcing the tradeoffs between speed and rigor that have derailed so many projects in the past. That is genuinely good news for a state with too many aging systems and not enough time or money to replace them the old way.
But before departments dive in, they need to decide what the system needs to do, what can wait and whether the organization is truly ready to own what it is building. Agentic AI does not make those decisions easier. If anything, it makes it easier to overlook them.
The good news is that state departments do not need to figure this out alone. The California Department of Technology can help departments honestly assess their readiness and develop the guardrails and processes needed to use these tools responsibly.
But the vendor community has a role too. The state's technology partners are already deep in agentic AI development, and the departments that leverage their knowledge and experience will be better for it. The state and its vendors do their best work when they come together. So, let’s see how we can collaborate and iterate so we both succeed in this new agentic AI environment.
