Believe It Or Not, The Government Is Adopting AI to Make Your Life Easier

The public sector moves slowly by design. That might actually help it get AI right.

This conversation is sponsored by Kyndryl

The Department of Motor Vehicles might not be where most people expect to find cutting-edge technology. But in a recent interview, Brian Shell, a senior partner at Kyndryl, recounted a surprising array of AI use cases in place at state DMVs right now.

Today, DMVs and state agencies are using AIs to cut wait times, reduce paperwork, and rethink what it means to serve the public. Just imagine the license photo process, which can be arduous, made smoother by AI that gets the background exactly right. That’s rolling out in some locations today.

Rapidly advancing AI capabilities, Shell argued, have sparked to “a resurgence in the spirit of public service” and a desire to deliver for taxpayers. We get into what’s happening in this Q&A below, edited lightly for length and clarity.

Alex Kantrowitz: The public perception of government is that it’s slow to adopt technology. Is that true when it comes to generative AI?

Brian Shell: Well, maybe appropriately so to some extent. Governments are durable institutions, and so they should be thoughtful about their adoption of new technology, and we’re definitely seeing that here.

Governments want to have policies and regulations in place before they start adopting things wholesale. That said, there are a lot of people in government who are adopting generative AI. We’re seeing AI piloted and then moving past pilots in lots of states and agencies. Are they adopting it as fast as tech startups? No. But I wouldn’t say that’s a problem.

We’re about three and a half years past ChatGPT’s introduction. Does the fact that government is now actively engaging with this technology signal that no area of society can really leave it alone?

The pervasiveness of AI is very much clear at this point. It’s going to be a part of every area of the world that interacts with technology. It’s today’s version of ‘every business is an IT business.’ It used to be that you could run your business without IT but that’s not been the case for 30 years. This is in that same vein. It will become part of everybody’s technology landscape.

When I think of government, I think of deterministic activity, meaning concrete processes with concrete outcomes. Generative AI is inherently probabilistic. How does an institution set on defined results create the openness to integrate a probabilistic technology?

The way I see governments implementing AI right now is actually still more on the deterministic side of the house. I don’t see them using a lot of LLMs to generate new content. I see them using what they’re calling discrete, practical, or tactical use cases of AI to help solve specific problems.

Using AI to create new stuff — which generative AI is by its very nature — hasn’t really hit full scale yet, and I’m not sure how they’ll adopt it. It’s an area that doesn’t jibe well with the general setup of government. They don’t want to be guessing.

So then explain the use cases that are actually happening in government right now.

We’re seeing small, medium, and even large deployments that are more focused on specific business areas with well-understood problems. On the smaller end, one example is AI photo background removal at DMVs, which sounds simple, but it’s had a real impact on citizens. One DMV has taken their photo capture process from a three-to-five minute ordeal down to five to ten seconds at the desk. People used to have to walk to another area where a high-resolution camera and a big blue backdrop existed. Now there’s a camera at every desk: look over here, click, and the AI does the background removal, checks for glasses, head tilt, all of it. Ten seconds and they’ve captured the photo. That’s a small use case built into an already existing flow to make life easier for citizens.

On a larger scale, we’re seeing agencies use AI to modernize processes. As they work to replace legacy systems, they’re leveraging AI to help them move faster through enormous amounts of data.

The DMV isn’t exactly known for its efficiency. Walk me through how they went from seeing this technology to deciding, ‘yes, let’s implement it’ for something as seemingly sacred as taking the photo for your license.

This is the area I probably know best, because it’s where I work every day. One of the big themes for DMVs over the last five to ten years has been citizen experience. That’s become a pervasive focus across DMVs in the US. Some people might shake their head and say it hasn’t worked in their state, but overall it’s been a genuine priority: how do we improve the DMV experience for the people we serve?

And when you have that guiding principle, you can start identifying bottlenecks. Ask any frontline DMV worker about the pain points — they know exactly what they are. And then you can start solving some of those problems with AI.

Another good example is natural language search. The way government talks about things is not the way normal citizens talk about them. Government has its jargon, its official terminology — and citizens just want to know how to get their new license. So one DMV built a natural language translator: a citizen types a plain-language question, and the system maps it to the right place. That agency can still use its codes internally, but citizens can search the way they think and actually get to what they need. The accuracy has been really high.

And then there’s document processing. The old cool thing was optical character recognition, but that’s been totally leapfrogged by agentic models that can scan and analyze paper — which government still has a lot of. We’re working with agencies that are piloting and moving into production on things like title transfers, making them more digital, and even when people come in with paper, processing it very quickly using AI to check accuracy, verify it’s legitimate, confirm all the right fields are filled. The accuracy is really high, and the results are good.

How does the public actually experience the benefit of these AI programs rolling out within government?

It comes back to citizen experience versus cost savings and efficiency. One of the hallmarks of government is there’s always more work than the people can handle — and government workers get criticized for being slow and bureaucratic, for not keeping up with consumer expectations. But if you talk to frontline workers, they typically have more work than they can manage.

What citizens will feel first is faster turnaround times. ‘Don’t waste my time’ is a universal desire, and that’s going to be one of the most immediate and visible impacts.

One of the DMVs I work with actually tracks a metric called ‘citizen minutes saved.’ And we’re talking about saving tens of millions of citizen minutes per year across a single state. When you play that out — say, 3 million transactions a year, four minutes saved per citizen — that’s 12 million citizen minutes returned to people’s lives. For something they had no choice but to do.

I hear government officials say this a lot: people don’t have a choice to use our services. This isn’t like choosing between grocery stores. There’s only one place to get government services, and that comes with a real responsibility. I’ve heard a number of agencies say they don’t want to be thieves of people’s time — they want to serve. There’s a resurgence in the spirit of what public service actually means, and a willingness to use these tools to do it better.

In the private sector, ROI is fairly clear: am I getting my money back? What does the public sector look at instead?

Government isn’t a for-profit organization. They don’t have shareholders. They have a budget to manage and services to deliver, but there’s no profit line. They spend the money they receive to provide the services they’re responsible for. So success is measured differently — it’s driven by how well you’re delivering those services to your citizens. The metrics are just fundamentally different from how a private company measures success.

AI models have a tendency to hallucinate. In a government context, that would be really bad. How do you mitigate that problem?

Hallucinations tend to emerge when you start applying AI models to large-scale production scenarios — trying to analyze what needs to happen with a given input, for instance. That’s really where people have to get smart about how they structure their deployments.

We spend a lot of time on this. How do you build structures to catch and eliminate hallucination-type issues in the loop? And I actually think government, with its predisposition toward caution, will do quite well here. They’re going to set up really good parameters, policies, and regulations around their AI use — which will slow things down, yes, but will also produce reliable guardrails.

We’ve been working with a framework-based approach where you write policy directly into the code. The AI agents aren’t just operating completely freely — you’ve given them a framework to operate within. Think of it this way: if you don’t give them a framework, they’ll paint all off the canvas. But if you define the boundaries, they can be their AI selves within those constraints, and you get reliable output.

Government may actually be one of the better sectors for quality to trump speed. It’s a durable sector that isn’t trying to beat its competitors. It’s been around for a long time and isn’t going anywhere. That’s actually a feature, not a bug, when it comes to accuracy.

Talk about how you’re experiencing the improvement in AI models, and how that changes what you’re able to plan for or deliver.

It’s been pretty remarkable — hard to keep up with, honestly. And this is where I think governments really do benefit from partnering with companies that can spend the time at scale figuring out how these models are changing.

Early on, we were dealing with fairly narrow capabilities — photo background removal, things like that. But the models have broadened so significantly that we can now apply AI across almost every layer of a modernization project. It used to be that you could maybe do one or two things. Now the models can ingest a massive document — states have huge books of motor vehicle law, for example — and you can compare that to how the system actually works and extract real business insight from it. That’s not vibe coding. That’s not writing code at all. You’re pulling out extremely valuable, accurate business logic from these models because they can now synthesize, compare, and output in genuinely new ways. And that capability extends across data, encoding, the whole stack.

I’d say the new models have really broken down the idea of ‘you can use this in one discrete, siloed way.’ Now you have something much broader. You’re starting to see people build agents and then have to manage those agents. It’s blown up. And that’s enabling a lot of new capabilities that simply weren’t possible before.

Any final thoughts on where this is all headed?

I think there are a lot of good stories that will come out of the government’s use of AI over the next decade. It’s a space with potential for big improvements. The average door-to-door time at one DMV we work with — across a state of 7.2 million people — is 18 minutes. That includes your check-in, your transaction, your walk out the door. That wasn’t 100% AI-driven, but that’s where we’re trying to help governments go. And that’s a story worth telling.