Two Ways Government Can Get AI Wrong
Most AI conversations I’m in with government seems to produce two bad instincts.
One is to freeze.
The other is to buy something.
Freezing may feel safer. Government tends to be cautious about new technology, and AI comes with legitimate questions about privacy, security, records retention, accuracy, labor, and who is responsible when it gets something wrong.
But freezing isn’t actually a neutral position. Staff are often already using these tools, and if the organization doesn’t create a legitimate way to experiment with AI, experimentation doesn’t stop. It just locks it into personal accounts and workarounds where nobody is governing it. Los Angeles had to issue a directive in March 2024 discontinuing unauthorized AI transcription and summarization tools — city staff had started using tools like Otter.ai and Fireflies.ai on their personal accounts, prompting the city's Information Technology Agency, working with the City Attorney and Mayor's Office, to bar unauthorized use of them (City of Los Angeles Information Technology Agency, report to City Council, Council File 23-1020, April 4, 2024) — which tells you how much was already going on.
The other failure mode is on the opposite side of the spectrum: getting sold.
The AI market for government is almost entirely vendor-driven right now. Some of those vendors are very good. Some are selling snake oil. All of them lead with their own product – and their product won’t solve all your problems. The good vendors will tell you so. But low AI literacy across the public sector creates problems for buyers who are still figuring out what AI can actually do and discern whether a particular AI product actually solves the problem they need to fix.
That’s the capability I think government needs most urgently. Not expertise in every AI model or product, but enough fluency to define the problem, understand what part of the work AI might actually improve, and ask good enough questions to judge what’s being offered.
And if you feel behind, you’re in good company: in a Gallup survey fielded in late 2025, 57% of public-sector employees said they had never used AI at work at all. While the market is moving very quickly, government, mostly, is still figuring out where to begin.
We don't sell AI tools, and probably never will
We’re building some cool AI stuff internally to PPI, but that’s not what we sell. We think our best value is what no product can do: figuring out where AI is a good fit for the problems you actually have, and where the underlying process issues need some other kind of solution entirely.
We're paid the same whether our clients buy an AI product or don't. We don’t take commissions. Because we don't implement AI systems, we can afford to tell you when you don’t need one.
Sometimes the answer really is that you need an integrator rather than a consultant. Sometimes a supposed legal or security blocker is mostly just uncertainty about what's allowed. Sometimes the process underneath is inconsistent enough that putting AI on top of it would just make the inconsistency harder to see and ultimately fix.
And sometimes AI really is the right tool.
The tricky part is knowing which situation you’re in.
That requires some understanding of AI, but probably more important, it requires a good understanding of your own work.
A slow permitting process, for example, might involve bad information, unnecessary approvals, inconsistent policy, poor handoffs, understaffing, or work that requires expert judgment. An AI product might help with one or two of those things, but it’s very unlikely to help with all of them.
If you can’t separate those problems from each other, it becomes very easy to buy a tool that solves the part a vendor can see and sell you on rather than the part that is actually making the service fail.
You don’t need enough technical expertise to evaluate every model on the market. You need to understand your own process well enough to say: this is the problem, this is the part technology might improve, and this is the part we have to fix ourselves.
If you can do that, then when you talk to vendors, instead of asking, “What can your product do for us?” you can say, “Here’s the problem we’re trying to solve. Here’s the part we think AI might help with. Show us what your product does there, and what would still have to change on our side.”
Empowered, not sold to
I think that’s the real challenge for government right now.
AI is moving too quickly for public agencies to wait until the market settles down and somebody writes the definitive playbook. I don’t think that playbook is coming.
But making a move in the AI space doesn’t have to mean buying something.
It means getting good enough at understanding the work, testing assumptions, and distinguishing between different kinds of problems that you can make sensible decisions as the technology changes.
And that ability can't live only with the CIO, city manager, or whomever signs the contract.
The staff doing the work know where the process actually breaks down. They know which exceptions happen every day, which rules contradict each other, which steps exist mostly because nobody has questioned them lately, and which apparently simple decisions actually depend on years of experience.
That's important for another reason: your staff are also likely to find some of the best uses for AI.
The more I use these tools, the less convinced I am that anyone can reliably predict from the outside which government work AI is going to transform and which isn't. What you can do is understand your own work well enough to know what kind of problem you’re looking at.
Sometimes AI will be the answer. Sometimes the answer will be a policy change, a cleaner process, better information, more capacity, or somebody finally deciding which of the five conflicting documents is actually correct.
The hard part isn't finding an AI use case.
It's knowing whether AI is actually the solution to the problem you have.
Comments