Beyond the Efficiency Trap: Creating a shared vision for AI adoption in government

Ms. Garcia’s contributions to this work were made in her capacity as a Member of RethinkAI’s Civic AI Advisory Trust. This interview was conducted by Freddie Price (Senior Partnerships Manager, Apolitical).

What would it look like if governments treated AI as a chance to redesign services, not just speed them up? Across the United States, public servants are wrestling with how to adopt responsible AI at scale. But a lack of shared vision means approaches vary widely, and many governments are using AI to optimize existing processes instead of seizing the opportunity to reimagine how they deliver services.

In this conversation, Apolitical speaks with Mai-Ling Garcia about the recently published ALT (Adapt, Listen, Trust) framework. The framework provides a new approach to public AI adoption designed to help governments move beyond narrow efficiency gains and shift to people-centred AI. It guides civic leaders to adapt to AI-driven demand, listen to communities at scale, and build the kind of two-way accountability that rebuilds public trust.

Read Making AI Work for the Public: An ALT Perspective in full, written by co-authors Mai-Ling Garcia, Neil Kleiman and Eric Gordon.

Q: What specific problem or set of challenges were you and your co-authors trying to address when developing the ALT framework, and why did it feel urgent to intervene in how AI is being adopted in government?

The advent of generative AI tools has accelerated the investment and rush into further AI development. That acceleration has prompted a level of cognition that didn’t previously exist in governments. We’ve seen this dynamic before, with the rise of social media, or with the broader dissemination of the internet, which is historically where my work has come in.

In many conversations we were having internally and across the industry, it became clear that, unlike in the 2010s, visions for the role civic technologists should play are emerging unevenly across the field. (Some) governments are rushing toward AI (some aren’t). Providers and vendors are rushing in. And, perhaps because there isn’t a shared vision, much of the activity in the field is focused on controlling legislation to restrict AI, rather than a proactive effort to shape what this technology could do for society.

One of the biggest opportunities for AI is its ability to work with enormous datasets, which is exactly the kind of problem the public sector manages. Government isn’t running a business; it’s running an economy. AI is well-suited to that scale. And yet there’s hesitancy, a lack of vision and no field-level framework connecting the technology to real public outcomes. What we saw was mostly automation, status-quo reinforcement and restrictive legislation — the public conversation has skewed toward risk and automation, with less energy around transformation.

Q: In the paper, you are candid about some of the ways in which the civic tech movement has — despite good intentions — been unsuccessful, stating that the “marginally-better-with-technology approach has run its course.” As someone who has had a career within this movement, how have your experiences shaped the way you think about AI and how governments should be approaching AI adoption?

It’s one of my favorite lines in the paper because it captures something those of us in civic tech have all experienced. After nearly two decades in government, I’ve seen how easy it is to fall into an efficiency trap.

I started my career in Oakland trying to increase access to city services, largely by shifting to websites and web tools. Across the movement, we built a lot of technology, but the measures of success were pinned on the number of times people could complete a task, rather than whether we were delivering better public results — especially for the people who needed it most, who haven’t been well-represented in the field.

In the private sector, you’re often thinking about maximizing marketing funnels, maximizing conversions, and transactions. But in government, getting more parking tickets paid, for example, doesn’t necessarily improve the lives of citizens meaningfully.

The things that should matter most to governments are outcomes for people. But also technology, and the relationship between them. Technology is fundamentally changing how organizations can work, and should work, for the better of people. Sometimes we lose sight of the deep organizational changes required to fully realize that.

In my view, what we’re hearing about AI risks repeating the same pattern. We can’t afford to do that again.

Q: You mentioned restrictive legislation, which is something you highlight in the paper as a difference in how states and cities are approaching AI. Why are we seeing different approaches from states and cities to AI adoption?

States and cities function very differently. States often provide broader infrastructure: enterprise-wide scaffolding, sandboxes, walled gardens, mandatory training etc. They create the spaces for employees (and sometimes the cities within them) to experiment.

Cities, by contrast, are charged with solving urgent problems, one pilot at a time. They work on very specific projects: 311, wildfires, translation for direct services, permitting, and other issues happening in proximity to residents. Cities manage the most intimate parts of our lives — from basic infrastructure to education — so their AI applications tend to be highly specific and high-value.

They can absolutely learn from one another. States have the ability to build with stability and scale; cities excel in responding to acute needs. Both need the other’s strengths. There’s richness in understanding how to deploy AI both rapidly and in long-term, sustainable ways with strong guardrails.

Q: Who is getting AI adoption right today?

The governments that are getting it right are also getting it wrong — in public — because they’re experimenting. That’s exactly what we want.

Anchorage, Alaska tested and ultimately stopped the use of AI-assisted police reporting. That’s a success story. They tested, discovered it didn’t work, and talked about it openly.

New York City experimented with cameras on buses and ended up issuing large numbers of tickets erroneously. Again, that transparency and willingness to examine results is a strength.

There are also some really good examples of rural communities, e.g. in Georgia, who are clustering together to maximize their resources for things like wildfire detection.

Another example is Latin America, where cities are rapidly experimenting with AI to solve a range of problems, including mitigating increasing land temperatures or to better respond to and repair streets at scale. I’d say the U.S. has a lot to learn from places in Asia and Latin America when it comes to advancing quickly.

Q: I want to move on to the ALT framework that you have developed. Can you walk us through Adapt, Listen, and Trust?

ADAPT is about planning for the demand that AI unleashes if it works well. If AI improves processes, you may get more parking tickets, more diagnoses, as happened in Jalisco, Mexico. And so what that means is you have to plan in advance for how your organization will adapt to increased demand if AI works effectively, and how people’s roles will shift. That requires ongoing retraining and reorganization. AI won’t eliminate problems; it will create different, better problems, and new jobs to solve them.

LISTEN is about understanding what people truly need, without jargon. It’s not about collecting more voices; it’s about accurately interpreting the voices you already have.

TRUST is about building institutions that demonstrate two-way accountability. It’s more than transparency, it’s designing systems that are fair, responsive, and trustworthy. We shouldn’t be digitizing systems people already don’t trust; we should be changing how government works altogether.

Thinking through these three components for any AI or tech project leads to fundamentally different outcomes.

Q: We hear a lot from public servants that are both excited and anxious about what AI means for their jobs. The Adapt component of the ALT framework proposes that “governments must not design for job elimination; they should design for human redeployment and organization adaptation.” What does this look like in practice to you?

Ultimately, this means starting with service transformation and allowing workforce planning to follow from the redesigned service model — and doing so quickly, because those roles will inevitably evolve as the service changes. Practically, it means a couple of things. At a service level, it means redesigning an entire service, not just automating the existing one. So that's one: anticipating not just what jobs you need to do the service in its current state, but if you were to actually redesign an entire service, how would that shift the roles and the people and the players? That requires anticipating how roles and responsibilities shift when a service is rebuilt with AI.

At the organizational level, it means asking what infrastructure and structures you need to support a transformed service. Again, if AI increases public health diagnoses, for example, do you need more nurses? If AI reduces frontline diagnostic work, how do those roles change? The key is aligning organizational response with your values, structure, and legal constraints. These are the questions we should be asking now.

Q: What advice would you give regular public servants — not in leadership positions — who want to apply the ALT framework in their own context?

First, evaluate your outcomes against this framework. Public servants are excellent at reporting, observing, and collecting data, and they can participate directly in shaping the measures we’ll need going forward.

Second, if you need authority, ask for it. If you need resources, make the case loudly and often.

And if your role doesn’t allow it, build AI literacy through low-stakes personal experimentation. Understanding the technology personally will make you a better advocate, a better resource-gatherer, and a better evaluator of whether AI is actually effective.