Government for the AI era

What happens to government when the world moves faster than it can? AI is bringing that question into sharp focus.

Jennifer Pahlka sees developments in AI as part of the larger challenge of building a government that can adapt to a fast-changing world. She’s spent two decades on this challenge in civic tech, the White House, and across government service delivery, giving her a deep understanding of why government struggles to keep up.

Jen founded Code for America, served as US deputy chief technology officer in the Obama administration, and wrote Recoding America on why government struggles to deliver in the digital age. She now chairs Recoding America, a field catalyst for state capacity, and writes Eating Policy. She is among the most influential voices on state capacity and government reform.

In our conversation, we cover:

• The government use cases for AI that Jen finds most exciting

• The institutional barriers preventing agencies from using powerful new tools

• What talent government most needs to deliver on the promise of AI

• How to bridge the divides between technology and policy work in government

• Jen’s advice for people hoping to start careers in public service and government capacity

If you enjoy this interview, you might also enjoy working at Horizon! Our mission is to bring much-needed technology expertise into government, and we’re growing rapidly to meet the scale of that opportunity.

We’re hiring for nine roles across policy, events, comms, operations, and data. We welcome a wide range of candidates, from exceptional recent graduates to mid-career applicants. Apply here by August 23, 2026.

Remco Zwetsloot: You recently wrote that “We are just starting to understand how government might use AI. We don’t really understand how AI will change the work government needs to do.” Let’s start with the first piece. What are we just starting to understand about how government might use AI, and what are the big open questions for you?

Jennifer Pahlka: One underappreciated use is regulatory simplification. Take unemployment insurance. We had failures during the pandemic, and with workforce disruptions caused by AI, we will have them again. I do not think those systems are fixed. And the real reason they’re so fragile is not that there is COBOL in the computers. It’s that in one state, the regulations governing the program, just from the Department of Labor, added up to 7,119 pages. Try administering a program that has to scale to 10 or 20x the number of users when you have that many requirements to meet. People can and should use AI to modernize old government systems. But AI can also be used to do the hard policy work of simplifying the rules themselves, not just complying with them.

There’s such a push to take very complex policy and regulatory frameworks and say, people don’t really understand this anymore, and the way to comply with all of this stuff is to let the AIs do it. But if no person actually understands it, we truly are just having the AIs talk to the AIs and then trying to adjudicate that. It’s just not going to work.

We can choose to use AI to simplify and pare down, maybe the word is rightsize, the volume of regulation that governs government programs. Or we can use it to more easily comply, at least in the short term, with those large regulatory burdens. I think the latter is a terrible idea. We absolutely have to simplify, so that a human being can still understand what’s actually going on.

I’m also very interested in using AI to change the dynamics between governments and their technology vendors. Say you’re a mid-level person in a state health agency right now, implementing the new Medicaid work requirements from HR1. Every state is paying Deloitte or another vendor roughly $50 million to build that module. It’s nothing against the vendors, but we’ve created systems that make this expensive. Instead, you could fire up Claude Code, point it at your own codebase, and ask it for the most efficient and sustainable way to build that module. You’d be in a very different position when you put that work out to bid than you are today.

We did not have these tools until recently. And coding agents ought to do what the internet era was supposed to do for government, which was disrupt it. We didn’t disrupt it then. We actually added layers, pulling the problem even further away from the solution. I’m really hoping AI finally brings that positive sense of disruption to government, on both the policy and the technology side.

Regulatory simplification is one area, vendor dynamics another. If you had more people who could think creatively about how government could use AI, are there other areas you’d point them to? Places where we don’t have enough thinking to even know what we don’t know?

Jennifer Pahlka: Enormously. We are so early in the AI revolution, relative to how long it took the internet to have its disruption, that there’s just a lot we don’t know about how this is going to change business and society. I don’t think we yet know how much it’s going to change the public’s expectations. How long will it be before going to a website and signing up for a benefit feels really archaic? And as people like Sam Hammond have said, we have no idea what the public sector workforce is going to look like.

There are many things that could have been automated 10 years ago, before we had AI, and still aren’t. Some of that stuff does not need AI to be automated. I always say, please don’t use AI to do what a spreadsheet can do. It costs a lot of tokens to add some numbers together, and regular old software does that just fine.

But there are a lot of things where AI could be really helpful, especially if it continues to get better at the rate that it has. There’s a big difference in the attitudes of people who talk about AI, between those who’ve been using it and see the improvements, and those who tried it once and said, oh, this isn’t any good. If you’ve been using it, you really see that it could be better than humans at certain things that are still pretty low-value tasks. This is a small example, but when you’re doing a benefit determination, it’s really common for humans to misunderstand a claimant’s income. You read it as monthly instead of weekly, or you read it wrong. AI can use its judgment to do that in a way that is probably better than a human. We all want eligibility determined accurately, but I wouldn’t call that super high-value work.

There’s a big difference in the attitudes of people who talk about AI, between those who’ve been using it and see the improvements, and those who tried it once and said, oh, this isn’t any good.

AI may be able to make more consequential decisions than that, though. And if that’s the case, what does the public sector workforce need to look like if our first commitment is to serving the public? Even as we keep in mind we should not use these tools in an irresponsible way where we have devolved power to machines whose thinking we don’t understand.

The second piece of what you wrote was how AI will change the work government needs to do. What’s the initial direction of your thinking there?

Jennifer Pahlka: We don’t understand the workforce disruptions, and that’s gotten plenty of ink. We also don’t really understand the societal disruptions. My friend Tom Loosemore has a definition of digital that we used when we were trying to update government for the internet era: applying the culture, processes, business models, and technologies of the internet era to respond to people’s raised expectations. People did change how they lived and worked and what they expected of institutions and systems once the internet came along. That same thing is going to happen with AI, in ways that are very hard to predict.

So I want us to think not just about, we have a program right now that helps people when they’re out of work and we think more people will be out of work. What fundamental needs and expectations are going to change that go far beyond scaling the programs we have today?

In the UK, they’ve already started challenging not just how a program would operate in an AI world, but why we have this program at all. They took their six social safety net benefit programs and combined them into one, to make it simpler to administer, to smooth out negative effects like benefits cliffs, and to give themselves a chance to totally reboot how they run it. And they did a fantastic job building that program with test-and-learn frameworks. That’s probably still insufficient for the disruption of the AI era, but it’s a good starting place. We don’t do that here in the US. We just say, we have this program, how could we use AI to administer the program we have today a little bit better? I think that’s a mistake.

Do I know exactly how those things should change? No, because I don’t really understand the ways AI is going to change people’s needs and expectations, beyond just workforce disruption. But if we don’t get out of the habit of “how do we use AI to make this program run better” and into thinking far more deeply, probably with AI as a help, about how to meet the public’s needs and expectations as they change even more rapidly than they did during the internet era, I don’t think we will be able to maintain our democracy.

Some of AI’s effects on government work will be about how the public interacts with government, and programs for the public. But government obviously has many other missions as well. Post-Mythos, for example, we’ve seen government asking how do you keep critical systems and infrastructure secure when cyber security is being rapidly changed by AI capabilities?

Jennifer Pahlka: Yes. The clear thread between the cybersecurity threats, Mythos, and social programs is speed. The good news is that speed is something we have to fix anyway if we’re going to maintain our democracy. It’s been pointed out that if government can’t keep up with the advance of the models, especially as they relate to security threats, we simply lose the ability to govern. I think that’s absolutely true.

But I think in a certain sense, we’ve already lost the trust and faith that we can govern. Not only are we very slow to adapt to technological changes like the capability of a new model, it takes us about five years to propose and promulgate a new rule. That breaks democracy if you have a new party every four years, which we kind of do now.

The idea of democracy is that you elect somebody on the basis of them being able to do something, and then you can decide whether you reelect them on the basis of the impact of what they did. If nothing they did has any impact until after they’re out of office, what is voting even for? People see this polarization as a dissatisfaction with both parties, and I’m certain that it is to some extent. But it is also a dissatisfaction with a system in which it is not actually possible to have a feedback loop unless you speed up action that matters.

That’s true in cybersecurity, and it’s true in policy generally. And I believe, in a sort of mystical way perhaps, that that’s what AI is here for: to finally push this to the point where we’re no longer the boiled frog. It’s happening so fast that we must act. And what I’m saying is that we must act faster.

Much of what you’re describing is premised on an assumption that AI is going to be transformative. There was recently a friendly, constructive back-and-forth within the civic tech community that you were part of, about whether AI is really more of a “solution in search of a problem”. You wrote that with technologies like blockchain, that attitude is very good, and it’s reasonable to be skeptical. But AI feels different. In response, Mikey Dickerson wrote that this is the fundamental disagreement: “Jen is optimistic on the potential for genAI systems to improve to the point that they are capable of leapfrogging our current web 1.5 status quo … I am not.” Where do you think these different intuitions about AI capabilities come from? And do you think there are ways to resolve those disagreements?

Jennifer Pahlka: I hope so. I don’t want to put words in Mikey’s mouth, but I think there’s a place where he and I very much agree. Assume, for instance, that agentic coding tools could help with the problem I mentioned earlier, the states that have to build that new Medicaid module. A couple of things need to be true. The person in that state health and human services agency would need access to their own codebase. There would need to be somebody there who felt confident trying that experiment, which means they would need to overcome a fair number of both real and perceived barriers. You can just imagine all the ways someone could say, “you can’t do that.”

If I’m pessimistic on AI, it’s not because I don’t think the tools can be wildly helpful. It’s that the human, institutional, and legal barriers aren’t being addressed at the same rate as the technology is developing. Claude Code could be literally perfect next year, and it won’t matter if we can’t use it.

What has become scarce is not the ability to, say, write code or analyze something very quickly. That is becoming abundant as AI gives us powers we didn’t have before. What is scarce is the ability to work within very constrained bureaucratic environments to make that stuff have a positive impact. The fear that it will have a negative impact is one of the barriers, but it is not the only barrier. Even if no one were afraid that it would get something wrong, or that it’s a bad tool to use because it concentrates power, there are still institutional reasons why these tools might not be brought to bear on problems that really need them.

What has become scarce is not the ability to, say, write code or analyze something very quickly. What is scarce is the ability to work within very constrained bureaucratic environments to make that stuff have a positive impact.

Fixing the institutions so that we get the best of these technologies and not the worst of them has been the work of the past fifteen years for me, and remains the work of this larger community. You’re never going to get all of the good and none of the bad. We need to get some of the good and minimize the bad.

I also think we need some real wins that speak to the values of the skeptical community, wins that show enormous benefit to the people and issues they care about. If you care about a social safety net, let’s show those wins. If you care about a stronger, safer nation that can deter its adversaries, let’s show that. Then we can more accurately weigh the benefits against the costs, instead of focusing exclusively on what the costs might be. And I don’t think there are no costs. There are costs to using this in a lot of ways.

But even since that dialogue happened, I have found more and more evidence that folks from the progressive left, who tend to be more skeptical, are diving in. Mayor Mamdani recently announced something called the PIT Crew, deploying in-house technical talent teams in New York City. This is a democratic socialist mayor who is being very tech-forward, and my understanding is that they appropriately use agentic coding tools in their work. That’s the kind of signal that should help others come along.

And I want to tell that story in the context of the Trump administration also using those tools, because we really need to make sure that the entire state capacity agenda, of which I see use of AI in government as a critical but not exclusive part, is not owned by any one party. It is fundamentally a bipartisan thing—it is something you do if you care about government working. Plenty of people will have beliefs on both sides about whether the other party actually wants government to work. But let’s judge things by the concrete ability to see progress and value for the American public.

One example you gave in that exchange was new to me: Ukraine has a national AI agent that lets you describe a need in plain language and then completes the workflow end to end, instead of routing you through forms. Is that the kind of win you’re imagining?

Jennifer Pahlka: Exactly. And how little ink does that get? Why aren’t we talking about that?

I’d love to transition to what this all means for what people should do, whether that’s government, field-building organizations, or individuals. Suppose some of these more significant implications come true, on whatever timeline. What do you think are the biggest changes in government’s talent needs?

Jennifer Pahlka: There’s such a wide variety of needs. We tend to use “AI talent” as an umbrella that includes really wildly different needs. If we try to do a NASA for AI, or a DARPA for AI, or a public option for AI, we’re going to need AI researchers, and those are going to be really, really hard to get. But I can see a whole wide range of things where we don’t actually need AI researchers. What we need are new versions of the roles we’ve had before, but that understand the world we are moving into and can meet the needs using the tools.

That goes back to my point about the policy and the constraints of the bureaucracy often being the barrier. To get public servants who can work within that environment and get the best out of the technology, you do not need a $10-million-a-year AI researcher. You need somebody who can see the bureaucracy for what it is, envision what it can be in a new world, and then map from A to B. That will involve a deep understanding of what the technology can do, for sure, but also how we actually bring it to life in the best possible way within our public sector systems.

You need somebody who can see the bureaucracy for what it is, envision what it can be in a new world, and then map from A to B. That will involve a deep understanding of what the technology can do, for sure, but also how we actually bring it to life in the best possible way within our public sector systems.

I think that’s just the new version of policy wonks and bureaucracy hackers, but with a little power-up: a vision for what I’ve talked about as the third horizon, the world we’re trying to get to. Whether you’re on the outside, in advocacy organizations or talent pipeline organizations, or on the inside in an agency, how do we collectively start interrogating everything we do with a lens of: is this making today a little bit easier but tomorrow more painful, or is this actually building toward the world we need to be in? Because that world is coming so soon.

So, more next-gen product managers, next-gen policy analysts, roles that are going to need wildly different job descriptions than they have today. Hopefully we can task Scott Kupor and his team at OPM1 with that and share it with every state, or maybe states are doing it and we can flow it up through them. We’re entering a phase where all the old skills are still relevant, as long as they acknowledge their role in being part of a transformation. Just getting things done is no longer going to be sufficient.

You’ve written about the division between civic tech and policy work, where technologists are often told to get back in their lane if they venture into policy. Speaking from personal experience, at Horizon we think about getting technology experts into public service, and the skills we teach are quite policy-oriented, things like memo writing, coalitional dynamics, and convening. There are lots of synergies with civic tech-flavored work. We have fellows who help their host offices adopt technology, and the hiring and talent reforms we care about are often very similar. But the policy and civic tech communities can also feel a bit siloed. I’m curious if you have a sense for how we could break down some of these barriers, while recognizing there are real differences? What does a good future look like, where technologists are crossing those technology and policy boundaries in a more positive way?

Jennifer Pahlka: I think organizations like yours and mine, and the larger community we both work with, need to be advocates for the test-and-learn feedback loops. We talk about this new model government needs to move to to achieve its policy goals. It has four things in it. You need the right people, so we have to do civil service reform. They have to be focused on the right work, so we have to do procedural reform. They need the right tools, so we have to reform how we build and buy technology. And they need incentives for outcomes over process, which means we have to create a feedback loop between policy and delivery, largely between the executive and legislative branches. We have to reinvent oversight.

It’s very hard for each individual to fight on a one-on-one basis: I know I’m a technology person, but this is actually a policy discussion. But we can advocate for a broader understanding that the model government works on is stuck in an industrial era and needs to move into an AI era. That model requires a feedback loop that braids technology and policy in a fundamental way, not because this is a technology we’re trying to regulate, but because everything government does requires technology to do it, and that technology is changing, and the needs we’re trying to meet are changing.

This is a very tall order, I get it. But if we can advocate for that shift in thinking, to a new model in which feedback loops become the norm instead of the waterfall method we have today, then every individual you’ve ever supported through your fellowship, every person fighting the fight of “I’m stuck in a silo here and the outcome will be bad because we don’t have a feedback loop,” will have air cover. That needs to become the norm, and I think that requires a narrative this whole field has been lacking, and a field agenda this whole field is now in the process of building. It’s just so hard to fight one-on-one.

And to return to the point about speed: it’s not just speed of rulemaking, and it’s not just speed of delivery. It’s speed of the whole system. As we all know, a system runs at the speed of its slowest bottleneck. And Congress is going to quickly become that slowest bottleneck. I guess it already is. I think that’s another role for AI, and I don’t just mean legislating on AI.

It’s not just speed of rulemaking, and it’s not just speed of delivery. It’s speed of the whole system. As we all know, a system runs at the speed of its slowest bottleneck. And Congress is going to quickly become that slowest bottleneck.

When we talk about Congress and AI, everybody goes to regulation of private sector AI. What I mean is that Congress is going to have to learn to legislate in a fundamentally iterative, test-and-learn way, where what they write is more about what they hope to achieve and what the outcome should look like, and then allows for testing and learning about how we actually get there. That’s what the AI world is going to demand. Otherwise, we’re going to keep trying to meet the needs coming our way and failing, because that’s what we’ve done in the past: we try something, it’s not quite right, but we have no ability to adjust it, because we’ve hardcoded every detail of it.

If instead we’re in that new world where we’re focused on outcomes, with lots of flexibility in how we get there, that will require a constant feedback loop between a whole set of actors that don’t talk today. And that will be the world Horizon fellows live in: being involved across what are now silos on a daily basis.

Let’s bring it down to the individual level, because this movement ultimately consists of individual actors who have to make choices about where they work and where they can contribute. Imagine you were 25 today, with your current convictions about everything we’ve talked about—but without your network or experience, so you’re at the starting line. Where would you start? What kind of work would you do, what kind of studying or learning?

Jennifer Pahlka: Such a good question. I both regret and totally don’t regret everything in my past, in the sense that I can go back and see what parts of it were wrong, and don’t regret it because that is how you learn. So the first thing I would say to anyone who is 25: if you learn from what you do, you did not make a mistake in what you chose. Your question should be “what am I trying to learn” more than “where should I work.”

If it were me, I would go try to solve a particular problem I could get my head around, with the new tools and new approaches that are available, and run into all the barriers I’m going to run into. That would teach me what I needed to know about what I should be advocating for. And then I would try to be a voice among many others. What we need is a chorus, a very, very large chorus of voices all singing in harmony. They don’t need to say the same thing. I would want to find my place in that chorus, so that it sings loudly enough and beautifully enough that we make the changes we need to make.

What’s an example of a particular problem you could tackle?

Jennifer Pahlka: There are so many. Let me give you a shorter-term one and a longer-term one. Here’s a vivid example: somebody needs to get out there and figure out how we get far more granular and timely data on workforce disruption, now. We may or may not choose to respond to the workforce disruption AI brings. But if we don’t even really understand how and where it’s happening, we don’t even start with that choice. And it’s not that the data isn’t there or couldn’t be. As my friend danah boyd would say, data are made, not found. Our current systems give us data far too delayed and far too high-level to actually help you navigate this from a policy perspective—you could make that data in a way that becomes actionable.

The longer-term example is maybe a meta one. We know that the way government builds and buys technology is still stuck in this old, very requirements-heavy, big-bang, vendor-driven way that is not getting good outcomes. And we know the internet era taught us something called the product model that does work better. We should have adopted that 10 years ago. We are adopting it slowly, in certain places, not quickly enough.

So: what does the product model look like in an AI era? The product model is something we built in the internet era, by contrasting it with the industrial era and learning from that. How do you retain those lessons, but project them into the future? Because the temptation is to drag government into the mid-2010s, and that’s not a good idea. How do we define what the model should look like now, so that we can help government adopt it and skate to where the puck is going to be?

People will say, well, there are companies doing AI technology development everywhere. Yeah, that’s true. We know that. That’s not the problem. The problem is how it fits in with how decisions get made in government, with the way we fund things, the way we staff them, the way they are overseen. This technology bleeds beyond technology. To have it actually play out in a positive way in government, you have to change the whole model. That is much more than just how somebody makes code come into existence.

For someone who is new and still exploring whether these are issues they want to work on, are there key resources you would point them to? Programs, books, essays, podcasts, conferences?

Jennifer Pahlka: Some great articles to get you started would include:

• Nicholas Bagley’s “The Procedure Fetish”

• Steve Teles’s “Kludgeocracy in America” and “Minoritarianism is Everywhere”

• My own “AI Meets the Cascade of Rigidity”

Some books:

• The Unaccountability Machine by Dan Davies — systems/cybernetics lens on institutional dysfunction

• The Fifth Risk and The Premonition by Michael Lewis

• Crisis Engineering by Marina Nitze, Matthew Weaver, and Mikey Dickerson

• Abundance by Ezra Klein and Derek Thompson for political-economy framing

• If you want to read something over 50 years old but still oddly compelling and relevant try the book with the longest subtitle ever: Implementation: How Great Expectations in Washington Are Dashed in Oakland; Or, Why It’s Amazing that Federal Programs Work at All, This Being a Saga of the Economic Development Administration as Told by Two Sympathetic Observers Who Seek to Build Morals on a Foundation by Jeffrey L. L. Pressman and Aaron Wildavsky. You’ll love it! I promise.

• Recoding America, if I may!

• Coming in December: Why We Can’t Have Nice Things by Nicholas Bagley. Pre-order it!!

Newsletters and podcasts:

• Statecraft, from IFP (Santi’s interviews are arguably the best practitioner education available)

• Factory Settings, also from IFP — fantastic lessons learned on the implementation of the CHIPS program

• Eating Policy, if I may!

• Greentape from Thomas Hochman

• Dave Guarino’s Plausible Legibility and LinkedIn posts on the benefits delivery and AI

• Bloomberg’s Odd Lots episodes on procurement and government capacity

• Niskanen has a new blog called Trust the Process, worth watching

That brings us to our final question: any last words of advice or inspiration for people thinking about pivoting into these spaces, or working on the issues we’ve talked about?

Jennifer Pahlka: My go-to piece of advice is: it’s going to be hard, and don’t give up. And I guess I would repeat my advice from earlier. I’m in my mid-50s, and I have seen my generation want to change government and have some modest success, and I think it’s fair to say it hasn’t been enough. It’s not meeting the need. Younger generations are going to have to come into this with an impatience. It’s got to be respectful, it’s got to be curious, but it’s going to have to be a little bit impatient, and push the change that is now pretty overdue.

So hone your skills of change management as much as you possibly can. I still think that meeting people where they are is really, really important in government, and that means understanding how they came to the views they hold and understanding the incentives they have. I have a line in the book to this effect: when something doesn’t seem to make sense, it’s not usually that people are missing something. It’s that they are operating according to incentives that you can’t see. And that’s always the case. Understanding those incentives is one way you meet people where they are.

But we can’t just meet people where they are. We have to have somewhere we’re taking them. So articulate what that third horizon is for you: how you want government to meet people’s needs, what you think that world looks like. That’s the place you’re trying to take them. You don’t take them there in one day. But know that that is where you’re trying to take them, so that you don’t meet them and then end up staying where they are.

We can’t just meet people where they are. We have to have somewhere we’re taking them.

Big thanks to Jen for taking the time. If you enjoyed this and want to learn more:

• Recoding America the book and Recoding America the organization

• Jen’s Substack, Eating Policy

• Our civic tech guide on emergingtechpolicy.org