California has spent at least $5.8 million on AI pilots for state government. Half of them are already dead.
In 2023, Gov. Gavin Newsom signed an executive order launching eight generative AI projects across state agencies, from a call center assistant for tax questions to a tool meant to predict recessions. The pitch was efficiency without cutting services. Roughly a year and a half later, four of those projects are gone, including one that by all accounts was actually working.
The California Department of Tax and Fee Administration paid $445,000 for a yearlong contract with SymSoft Solutions to build a custom AI call center tool. Powered by Anthropic's Claude, it listened in on employee calls with taxpayers and searched through more than 16,000 pages of tax policy to suggest answers in real time. According to the department's own assessment, it reduced the time it took to handle an average inquiry and freed up staff for other work.
Then the department killed it.
Government Operations Secretary Nick Maduros, who ran the department before joining the governor's cabinet, told the Sacramento Bee that "even though the solutions worked in practice, they didn't save as much time as hoped." The real problem was that off-the-shelf tools from Amazon Web Services had caught up to what SymSoft built custom, and the department already had an AWS contract. Paying $445,000 for a bespoke tool when the generic version does roughly the same thing stopped making financial sense.
That's a dynamic a lot of organizations are about to run into. The tool didn't fail. The market just moved faster than the vendor could justify its price.
The other three kills were more straightforward. The Department of Public Health shelved a facility inspection tool because funding dried up. The Employment Development Department abandoned an AI recession forecasting project, deciding an internal build was the better path. And a housing initiative tracker at the Department of Housing and Community Development never made it past the pilot.
Four projects survived, though the reviews are mixed.
- Caltrans has the biggest spend at at least $3.5 million across contracts with Deloitte and Accenture for traffic management and road safety analysis. Both ongoing.
- The Department of Finance has a $565,000 contract with Authorium for a tool that analyzes legislation and estimates costs. Feedback from 40 testers was split, with one calling it "a very promising step towards reducing bill analysis workload" and another saying it "misses a lot of things or has inaccuracies, so it still needs to be double-checked."
- California Health and Human Services is in a two-year $1.3 million contract with Smartling for AI-assisted translation.
The state workers' union isn't exactly enthusiastic. SEIU Local 1000 president Anica Walls said members have reported AI tools actually adding to their workload, and that "these AI programs can't handle the complexity of the public's needs."
Maduros framed the entire effort as a win regardless of the body count. "The idea of this is to actually start working and trying things and see what works in a way that's cost-effective, rather than spending years and millions and millions of dollars planning," he said.
That 50% survival rate sounds rough until you look at the private sector. A MIT report covered by Fortune found that 95% of generative AI pilots at companies fail to deliver ROI. California beat that curve by a wide margin, and it did it by doing the thing most organizations struggle with, which is actually pulling the plug.
The CDTFA kill is going to be the most common AI procurement story of the next 12 months. Custom vendors are racing against platform players who keep building the same capabilities directly into AWS, Azure, and Google Cloud. Knowing when to cut a working tool might end up mattering more than knowing which tool to buy in the first place. And if you want one detail that captures the whole situation, the department that killed its own AI forecasting pilot is EDD, the same agency responsible for tracking California's job market.
