.png){kind=logo}
Roundtable CEO and co-founder Madeleine Smith took the stage at Code for America to share what a year of public servants actually using AI has revealed — and where it's headed next.
At Code for America, Roundtable CEO and co-founder Madeleine Smith shared what the team has learned from a year of supporting thousands of government workers using AI for real work. The findings were surprising.
Government work is hard. The question is where AI actually helps.
Public servants navigate fragmented systems, overflowing inboxes, and shifting policy — all while constituents demand faster answers and agencies are asked to do more with less. The honest question isn't"what can AI do?" It's "where can AI actually make an impact on the people doing this work every day?"
To answer that, we looked at how our users — public servants across more than 4,000 organizations working in public health, housing, emergency response, and beyond — were actually using AI inside the platform.
The pattern was clear:
Government workers, in other words, aren't asking AI to reinvent their jobs. They're asking it to help with “the work to do the work,” everything that needs to happen before a service can be delivered or initiative implemented.
There are three deeper lessons we gleaned from how public servants actually use AI.
Lesson 1: Find the zone of opportunity. AI adoption comes down to a simple calculus — find the low-risk, high-reward workflows and deploy there. Some workflows tolerate mistakes. Others don't. If the IRS could process returns faster with AI but 2% of Americans (7M people) overpaid as a result, that's not a tradeoff anyone would accept. The job is matching AI to the workflows where it can succeed without introducing unforeseen risk.
Lesson 2: When AI fits existing work, adoption flows. When AI lives inside the work people are already doing, training is almost an afterthought. Questions about context and status. Searching for a conversation buried in an email chain. Drafting a summary or memo. None of it is flashy. All of it eats hours every week — and it's exactly the kind of administrative weight that makes talented public servants feel like they're spending their days on overhead instead of impact.
Lesson 3: Sensitivities (rightly) slow adoption. There is an AI danger zone — sensitive data, decisions with real-world potential downsides, public-facing outputs. Trust in government AI isn't a technical question; it's a contextual one. It has to be earned workflow by workflow, agency by agency. That requires transparency about what the AI is doing and clear guardrails about where it doesn't belong, informed by the public servants closest to the work.
The path forward is not a chatbot. It’s AI that disappears into the workflow — automations running in the background that extend the hand of the person already doing the work.
In practice that can look like a housing coordinator who searches for eligibility guidance. AI immediately surfaces the relevant policy and the three most recent cases where it was applied. It drafts a note to a shelter with available beds. No new tab. No prompting. The next step appears because it’s expediting workflows public servants already execute (albeit onerously and manually).
Government workers currently spend 20% of their time — one full day a week — searching for information. Embedded AI can give that day back. Not through a big digital transformation, but through hundreds of small moments of time saved that compound into real capacity.
We’ve learned government AI isn't about the models. It's about three things: fitting the workflow, earning trust through transparency, and treating public servants as the experts they are. We know that if agencies get those things right, AI won’t add to the complexity of government work.
It will finally keep up with the public servants on the front lines, empowering them to do the work only they can do — and that's good for communities across the country.
Subscribe to our monthly newsletter below.