← Data-Smart City Pod
Data-Smart City Pod
Building an AI-Ready City Government
00:28:38
City leaders are eager to deploy AI, but the real opportunity lies in preparation: building the right organizational structures, expertise, and culture first. Host Stephen Goldsmith speaks with Teddy Svoronos, senior lecturer in public policy at the Harvard Kennedy School, about how to structure your city government for Agentic AI, why small, empowered teams work better than broad rollouts, and what mental models and skills leaders actually need to manage this new relationship with AI tools.
Loading summary
Transcript41 lines
- [00:03]
A
From datasmart city solutions the bloomberg center for cities, this is the datasmart citypod. Welcome back. This is Steve Goldsmith from the Bloomberg center for Cities with another episode of our podcast. Today we have a guest that we're quite excited about, Teddy Svornos, who is a senior lecturer in public policy at Harvard Kennedy School. And Teddy teaches courses on statistical methods to improve public policy. What caught our attention lately is his work on how he teaches artificial intelligence, the applications of agentic AI tools to solve problems. Teddy, welcome.
- [00:43]
B
Thanks so much, Steve. It's great to be here.
- [00:45]
A
So we reached out to you, anticipating you were probably some PhD in computer science or math and found you have a health policy background. Tell us, how did you get from health to merely working on statistics?
- [01:02]
B
That's a good question. And I mean, some would see it as an upgrade, some would see it as a downgrade. I think the thing that's interesting to me is I decided to do a PhD in health policy because I cared very much about specifically global health and making policies that improve people's lives in discernible ways. And I found pretty quickly that I lacked the tools to really understand if any of the work that I was doing actually was having an impact on anyone. And I found a lot of my colleagues were in a similar place. And so I thought I should, you know, explore in more depth what it means to evaluate these kinds of programs. And sort of unbeknownst to me, while I was in my PhD, I completely fell in love with teaching as a thing that I wanted to do, love to do, cared about doing. And so I pivoted pretty hard to just wanting to focus on helping people, be they students or mid career professionals or whoever, understand the kind of underlying concepts that matter so much for doing the stuff in the real world. And so that led me to teaching statistics more generally. And then my work on technology and stuff like that sort of pushed me into generative AI as well. And so here I am today.
- [02:11]
A
It's an interesting background. I've been hanging around local government for about 30 years. Mayor, Deputy mayor, our work now continues to look at the use of technology to innovate. So I want to spend a fair chunk of our time on AI. And but before I get there, this course that you and some of our Bloomberg colleagues did, using data and evidence as a city leader, if you want to set out as a mayor, county executive, senior leader, to create a culture of using data to solve urban problems, maybe moving from a set of routines to a set of more precise preemptive actions. How do you go about creating that culture of data? What have you been teaching?
- [02:56]
B
I think one of the most important things on getting a city or an organization or whatever to take a more kind of data forward approach is to get away from the notion that data and statistics is reserved for analysts or people with lots and lots of expertise who know how to crunch the numbers and run all these crazy models. Because in the end, even if you look at what it means to make really good evidence based policy, now it's really about asking questions and asking the right questions at the right time. So before you implement a program, asking what do we know about programs like this? What has failed in the past and why has it failed? Before you start the program saying, what would it look like if this was wildly successful? How would we know it was wildly successful? These are questions that I think anyone who cares about policy asks, ask themselves, ask their colleagues. And figuring out how to sort of mainstream that in your organization, to me is like this single most important thing you can do to create a culture in which caring about data and caring about measurable progress is valued instead of just paid lip service to.
- [04:02]
A
So let's say I'm the mayor and I want to unleash this power throughout my government. What should I do? Not specifically, but as a leader, how do I set up that culture to be implemented throughout the city?
- [04:17]
B
I think as a leader a really important thing to signal is certainly to ask questions for sure. But honestly, I think a big thing of being really evidence based in the work that you do is to acknowledge the possibility and sort of make space for the possibility that the thing you're going to do is going to fail, that it's not going to help anyone, that it's going to be a huge expenditure of money and nothing is going to be able to show for it, which is a absolute worst case scenario for anyone in any kind of public office, certainly in a leadership position. But if you don't allow for that, then any data based thing is just confirming your expectations, not actually probing whether or not what you're doing is actually having an impact. And so I think a really important thing that leaders can do is to say we're going to do absolutely everything we can to make this program a success. And there are things outside of our control that may make it fail, or we may have been wrong about the underlying theory of change that we were trying to implement. And if we don't take that seriously, then we'll never grow as an organization and we'll never make our programs any more effective than they already are. So I think that really is that willingness to accept that failure is actually a very important step.
- [05:27]
A
I was a young elected district attorney in Indianapolis, Indiana, and I engaged Indiana University to evaluate the programs. I started diversion programs, programs for young offenders. And I had this big news conference and I announced, we're going to do this. I was so proud of myself. Then over the next two years, I found that nothing I did made any difference. Right. If I diverted them, if I counseled them, I did this, I did that, I couldn't. Nothing was making a difference. So I appreciate your point about being ready for the negative. Let me take off on something you mentioned earlier and maybe get some free advice from you. So we have a project where we're trying to modernize performance management in cities. What used to be called stat programs still is a little bit. And one of the issues here is not the utilizing of AI to answer a question, it's how do you use it to ask better questions? So let's say specifically that my neighborhood has more potholes than the next neighborhood. Let me figure out what's causation, right? So how do you think or teach about the use of data and AI for purposes of improving the quality of the question?
- [06:44]
B
So I think separating it the way you just did is a really important step. I think the way that generative AI models work, and this goes way back to the early days of 2023 to today. All the models that we have are very interested in solutions. They're very interested in coming up with, here are 12 things you can do to make it potentially better. Right. I think, as we know in policymaking, that's often a mistake. To start with a solution or to jump to solutions is a really big mistake. And so part of it is actually interacting with AI in a way in which you're explicitly stating the only goal of this conversation is to characterizing problem as effectively as I possibly can. Let me know what data would be most helpful for me to understand this. What are the blind spots that I'm bringing to this conversation in which I am sort of assuming things are true? Having it sort of as this interlocutor can be a very helpful way to push you to kind of refine your question more and more. It's often useful to say, you know, I'm not a technical person, but it's important to me that this be statistically rigorous or that any causal claims I make are real causal claims. Just like adding that little thing into the prompt that you give The AI can be quite helpful in that regard. But one of the things that make these tools so good is that they make a lot of these ideas very accessible regardless of the degree of training that you had. And so you can actually have this like v very well informed conversationalist that doesn't know your context but knows a ton about statistics and impact evaluation and explicitly make that conversation a goal oriented thing around characterizing that problem.
- [08:15]
A
So you wrote this interesting piece called Agentic Everything. The title is quite provocative, but the piece was as well. And you talk about this fundamental shift moving from kind of chat to something categorically different. So how does Agentic Everything affect your answer to the last question? How has agentic helped me ask better questions?
- [08:38]
B
So I think my view of what makes agentic AI agentic, and you know, there are different definitions of this and I think there's a tendency now to call basically everything agentic in some way. The three things that make agentic AI what it is is first of all the ability to plan so make a task list for itself and to actually have this kind of set of things that it's going to do to accomplish a goal which requires very strong mod to the ability to do more than just give you text. So even the ability to search the Internet or to read documents is what we call tool use in agentic AI. Being able to look at your local stats database to figure out what, you know, indicators exist and stuff like that, all of that is sort of part of this tool use thing. And then the third and the one that I think is the most interesting is its ability to iterate. So the thing that really mattered to me as I started using Agentic AI in the past couple months I've been on sabbatical and it's been a very well timed sabbatical because I'm just using the heck out products is it will create something and then before it even shows it to me, it will evaluate its work and make a new version, and then make a new version and on and on and on until it is satisfied by some set of criteria that it is worth it for me to see. And that ability to sort of self evaluate and then improve changes a lot of the kinds of interactions that we have. And so in this context I think you could go from saying help me think through this problem to I want to understand as well as possible the nature of the problem. Right now I want you to make a pretty comprehensive report or dashboard or website or whatever that brings together all the information that you have access to Ask me if there's additional information that would be helpful to you that I can provide. And then the thing about these tools is you just kind of leave it alone and leave, like 20 minutes later you come back and it actually has a thing for you. And that lets you start from a much, much more deep perspective on what's actually happening in the problem that you wouldn't necessari get from just a back and forth. There's a risk to this, of course, as well, which is that it'll make something that's so nice and clean and shiny that you're like, oh, this is the problem. Like you figured it out, this is exactly the problem. And part of what these tools are really good at is making things look the right way, the way you'd want to see them. And you, as a policymaker, need to bring your own expertise to realize the fact that how something looks and what's actually happening under the hood are quite different. And you know about what's happening under the hood, right? You know all of these things about your local context, you know about where the data is coming from, you know what the current policy scenario is. And then. So being able to push back on those things is quite important because there's an extent to which these tools, you know, you're kind of letting go of the steering wheel a little bit more than you used to. And that could be great. But as you can imagine, it could also lead you down a path you might not want to go to.
- [11:20]
A
Well, I have lots of questions about what you just said, and I'm supposed to be asking you questions, but I have to tell you a story first, so.
- [11:25]
B
Great.
- [11:26]
A
A few weeks ago, like, on a Sunday night, I was talking with a fellow I've known for a long time, Santi Garces, who's a CIO in Boston. And we, with one of the tools, asked three one, one to help us predict rat outbreaks in Boston. Like, where would you find rats? And Santi and I asked some questions, and like I said, I've been local government for 30 years, so I have a lot of questions. And. And then at the end of the first phase, the tool responded, well, you haven't asked me about. I'm changing the word slightly, but you haven't asked me about food inspections. I should really look at food inspections as well as a predictor of rats. So this iterative process you talked about in an agentic form, I was really quite impressed and surprised that it told me what I forgot to ask about, not just gave me the answer to the question.
- [12:21]
B
Right. There's a story that I really like to tell that Charles Babbage, who made the sort of the first analytic computer, was totally mechanical. And he lamented the fact that people would come to him and say, hey, this computer of yours, if you give it the wrong questions, will it give you the right answers? And he was just so shocked that someone could ask that question. And I will say there are some folks in the AI space who are basically saying we might actually be getting closer to that in that you're asking a question and the machine is also trying to figure out more what you really mean.
- [12:53]
A
Right.
- [12:53]
B
And like go back and forth and help nudge you in particular ways. So it might literally be the first time that we have a machine where you can give it the wrong information and you can get a better answer out of it. But that is very contingent on it being right about a bunch of things. Right. And it knowing your context and knowing the way that you're nudging it. And so it's a very interesting thing that the thing that I've been writing about so much while I've been on sabbatical is this new sort of relationship to this thing that really is a co worker. It's like a collaborator on things. And the extent to which you delegate stuff to it and the extent to which you go with its plausible sounding suggestions, which may be fantastic and ones you hadn't thought of, or they may be a complete rabbit hole and you having the control to decide which ones they are and kind of go in that direction is I think, a new sort of cognitive skill that we need to get better at.
- [13:44]
A
I had this fellow work for me when I was mayor of Indianapolis. He was one of my most senior people and he had an mba. And one day he turned around as he's walking out of my office and said, I never do what you ask me to do. I only do what I think you intended for me to do. Right. So now that we have agentic, I can have people tell me what it is. I meant on the last comment that you mentioned and consistent with your agentic, everything you've written about, AI agents, as they get better, require stronger mental models on the part of the user. So what's, let's go back to putting that question with the first one. So you're, you're a leader in a city government, you're enthusiastic about agentic AI, but what skills would you then try to develop on the part of your, your leadership so that you're getting the right answer, what's the right partnership between the developing mental skills and the model skills?
- [14:44]
B
Great. So I think the first thing is hopefully something that people still believe, which I have come to believe even more strongly than before, which is that the domain specific knowledge is still extraordinarily important, that the AI doesn't obviate the need to really understand, like what it is you're asking, what the context you're in that you're asking about and so forth. Because not having that is such an easy way to just get completely misled by AI. So actually knowing the underlying stuff that you need to know is a very important thing. And developing that skill, I don't think it comes easy. I think it's through friction, through experience, through failure, through all that stuff is still super important. But then comes the part where you're actually interacting with the AI. You have this skill set and whatever and you want to actually use it. And the term that I like to describe this is called cognitive debt. And it comes from software developers have a thing called technical debt, which is when you're trying to ship a piece of software, you do all these little shortcuts and you kind of leave all these bugs in there because you know you have to get to them later, but you have to just do it. And so you slowly accrue technical debt with every new feature release until you have to pay it off. I think with agentic AI in particular, cognitive debt is this distance between what the AI is doing and what you think the AI is doing. Right? So you are naturally delegating stuff to these tools that is actually their feature. What makes them useful is you can delegate to them, but every time you do, you are kind of like taking out a little bit of a loan in terms of what's actually going on and what you're actually having the thing do, which over time accrues. And any economist will tell you debt isn't bad in and of itself, it's only bad if you have so much of it that you can't pay it all back.
- [16:26]
A
Right?
- [16:27]
B
And so you need to start to manage how much you're offloading, what degree of understanding you're offloading, and then knowing when you have to come pay it back. Because right now, for example, I do a lot of statistical analysis, I write very little code now. Most of the code is being written by AI and I am not doing line by line code review, I'm doing like chunk by chunk code review. I've delegated the line by line stuff to this tool And I now have that cognitive debt. Anything you're doing in the city government, anything you're doing to try to make people's lives better, every time you delegate to even a subordinate, you're doing that to an extent. But with AI, you do it with so many different small things that you have to really be thinking about what you're offloading to it and what you're keeping, and treat that not as a bug, but as a feature like that is literally what these things are for. You just have to get good at managing it.
- [17:15]
A
How does all that relate to prompt engineering?
- [17:19]
B
You know, it's kind of funny. So I teach a generative AI course. I teach it to HKS students. I teach it to exec ed participants. We have a free version of it online, though it's two years old. So it's ancient in terms of the content that we teach. And we used to focus really heavily on prompt engineering. Right. The idea of crafting your prompt exactly right. So that that's the input that the AI has. I think that that type of, like, very concrete skill is actually less important than it was before because these tools now are very, very good at writing their own prompts. Here's a very simple example. I asked it to help me make a presentation. When I asked it to help me make a presentation, it actually created sub agents for every slide that I wanted to make. And a sub agent is an independent AI that is just focused on making this one thing. And they were all happening in parallel. What had happened is that the AI had given an individual prompt to six other AIs to make those slides. So my prompt wasn't actually what anyone was using now. It's just like sort of the starting point. Right. And so we're actually a little bit more abstracted from prompting now. And it's more about trying to see the shape of what your interaction with it is doing, what assumptions are kind of underlying in it, and to what extent you can get the AI to document the decisions that it's making so that you can come back to it if you need to. It's a crazy world, Steve.
- [18:37]
A
I'm trying to keep up with you. I thought I was up to date an hour ago, and now I've decided I'm behind the times again.
- [18:44]
B
Steve, I highly recommend going on sabbatical if you can do it. It's a great way to learn this stuff.
- [18:49]
A
Thinking would be quite the interesting approach. I have to consider that concept. If you think about the Transportation Secretary or the Deputy Secretary of Public Works for Boston, whatever the title would be. And we thought about developing agents that would help them do their jobs better. Is that also antiquated in an agentic world, or does an agentic world make the agent better?
- [19:17]
B
So I don't think it's out of date. Definitely not. I think if anything, the notion of having these things that can kind of act autonomously still holds. I'm personally, and I was listening to the episode that you had with Mitch Weiss earlier, and I was sort of thinking a little bit about, you know, how he was framing stuff. And I think his way of framing things is just trying stuff is a really useful thing to do. But my only pushback is that right now my view of how the agentic ecosystem works is it's much more optimized toward an individual or a team using stuff more effectively than any kind of like public facing tool. Because these agents right now, you know, part of what you get from an agent is giving it more capabilities, more access to data, more things that it can do. And as a result, there's more ways in which it can go wrong. Right. So there are lots of ways in which these can spiral out of control. Where you know, there are examples of chatbots in like the New York City, for example, small business chatbot that like was giving people completely fake information. Imagine if instead of that, the chatbot could actually apply for, you know, small business licenses based on these faulty laws. Like, that would be really quite bad. And so right now I think of it as ways to optimize how government workers do their work, right. To do things more quickly and in larger scale, but to do so in a way that they can sort of control and monitor as opposed to one that's more faced towards citizenry. Because right now I think the guardrails are kind of falling behind the capabilities.
- [20:47]
A
So I wrote this article for Fast about a month ago that I think you just said was totally wrong. Oh no, just listening carefully. I've had a couple jobs where permitting reported to me, right? So you like take New York City. There's no restaurant permit in New York City. There's like a dozen permits that if you make it through without exhaustion, you get a restaurant permit. It's kind of your reward, right, to open a restaurant so you could think about merging those systems. Or in this article we suggested that an agentic front door would allow the system to operate as if it was customer resident facing even before you'd integrated it. Right. So that would be an agentic front door to a applicant. But I think you just said Be careful because you're maybe giving the applicant the wrong answers or simplistic answers or the answers they want to hear.
- [21:41]
B
I don't necessarily think. I think what you described does sound like a very interesting and cool use case. I do urge some caution in terms of how that kind of works. And it sounds like what you're just describing. There are still sort of humans that are reviewing stuff and kind of getting a sense of things. You're just trying to automate out the more mundane tasks that I think agents are quite good at. But the more I see people talk, like use agents in ways in which they have very, very little guardrails and maybe a very expansive remit in terms of what they're asked to do, they can really get pretty crazy. Like in January and February, this open claw thing became a big thing where everyone had their own little agent that was like emailing everyone and posting to each other's social media and doing all kinds of crazy stuff. Even though agents are incredibly capable, I think my natural inclination is to look for individual things that you can optimize with narrow tasks. Even though there's so much to be gotten from like this much more general approach. But maybe I'm just too risk averse for this kind of thing.
- [22:43]
A
I don't know. Well, I don't know either. So what else should I worry about in terms of running a city and AI, Energetic AI, does this make equity better or worse? Does it make community engagement better or worse?
- [22:58]
B
So for both of those things, I think it is a very classic Kennedy School response. It fully depends on the way in which you use it and how you implement it. In terms of what to be worried about. I think that the biggest thing that I would ask people to think about with these tools is just realize that there is kind of like a one to one trade off between utility and giving up control. Like the more data you give these things and the more you let them do stuff, the better they are and the more data you've given them and the less control you have. Right. Like that literally is the thing that is good about them, but that then also has risks associated with it. And there are lots of ways to get around that. There are lots of evaluations and benchmarks and things you can do to sort of monitor the behavior of the tools that you use. But just think very carefully about how much can I get from these things. And I think, you know, Mitch is suggesting that you should just go out and try it. I think is good because truly once you use these things, the way that they are sort of fully meant to be used. It is. I spend maybe the first two, three weeks of my sabbatical just like floored. Like, I just couldn't believe how much I could do in ways that I never thought possible. And I literally studied generative AI and have done so for many years. And it was so knew that it was amazing. So I think seeing what is possible and then seeing what do you have to give up to be able to use that functionality in a big way, managing that tension is really quite important. And this is sort of an extension of that. Don't lose sight of the expertise that you're bringing to these tasks. So even though these tools can now fully do literally the full pipeline of statistical analysis that I do is now fully automatable by these tools, if I didn't have the training that I had, I can think of many times in the actual pipeline where I would have gotten really, really misleading results. Not because the bot was wrong or like hallucinating. It was just making different decisions than I would make. And those decisions can really matter when you're actually like in the weeds of a particular analysis. So don't lose sight of what your kind of value add is to the conversation and don't give up so much that you then lose the agency that you have is I think, the other kind of big, big thing that I would caution people with. But it's an incredibly, like, truly an incredibly exciting time to be working on this stuff. And the more that you can kind of like glimpse what they're capable of, I think the more you'll see just how much it's going to transform a lot of stuff.
- [25:21]
A
Many people talk about AI with human in the loop. We talk about bureaucrat in the loop and the world we come from. So what would you tell a city leader trying to unleash the best of agentic AI in their city? What. What steps should they take? How would you create the environment for appropriate and ethical use of AI in a way that improves the quality of services?
- [25:47]
B
I think a good way to start. This is different from, I think, a lot of other attempts to get AI more used in local contexts where you just kind of like want to give whatever chatgpt for government to everybody in the organization with agentic tools. There's, at least in their current form, there's enough customization and enough uniqueness and enough risks that it's worth kind of thinking of putting together a small team of very driven folks in your organization and give them access, empower them to really, really stress test and figure out what is possible in the organization and let them spend the time to play in that space and see the step change that agentic AI represents and then come with suggestions and ideas and so forth. But I think that more concentrated, team based way of doing it is prob a better fit than these more general AI tools where everyone can write their emails better and things like that. This feels different because it really does require like very thoughtful and interesting architecture and the right people with the right kind of drive. And I don't mean like technical people, I just mean the people with the right kind of drive can really be the drivers of this kind of change within the organization.
- [26:55]
A
If anybody is excited and motivated by your responses like I am, if they're a mayor somewhere, can they sign up for an online course? How do they find out more about the things you're talking about?
- [27:08]
B
Absolutely. So in Harvard Exec Ed we have both a week long in person and a six week online version of a generative AI course where you get all the way from literally what are these models at all? Like, how do they work to are there going to be jobs in my city in a year? The entire kind of spectrum of things and pushing you in that way I think is really worth checking out. We spend a lot of time updating it every single run because things change very quickly, quickly. And so it really, I think is a cool opportunity to connect with other folks who are working on this stuff and kind of getting a sense of the cutting edge in a way that is still rigorous but also accessible to people who don't spend all their time working on AI.
- [27:47]
A
Sounds like I should sign up and then I can try these questions again.
- [27:50]
B
You're very welcome, Steve.
- [27:52]
A
This is Steve Goldsmith from the Harvard Kennedy School. We've been talking with our colleague at Kennedy School, Teddy Sivronis, on his work in agentic AI. Teddy, thanks so much for your insights and your leadership. We look forward to learning from you.
- [28:07]
B
Thanks so much for having me, Steve. If you like this podcast, please Visit us at datasmartcities.org and find us on itunes, Spotify or wherever you get your podcasts. This podcast was hosted by Stephen Goldsmith and produced by me, Betsy Gardner. Thanks.
- [28:26]
A
Thanks for listening.