‘Science experimenting’ in a leading AI state

NEW YORK CITY — New York has looked to stake its claim as a state leader on artificial intelligence, investing millions of dollars to boost its use throughout government.

The $400 million Empire AI Consortium, unveiled last year by Gov. Kathy Hochul, created an AI computing center upstate to research the technology and determine the best use cases. It also looked to engage heavily with the state’s academic institutions to accelerate research and train the next generation of AI workers.

The state’s 57 executive agencies and around 100 non-executive agencies are already experimenting with AI and seeing how it can help make government more efficient for residents and employees, including through generative AI.

On stage at the Google Public Sector, GenAI Live & Labs event this week, Dru Rai, the state’s chief information officer, said New York has dozens of use cases in various stages of development. And the state’s Office of Technology Services announced it would ramp up those efforts with an enterprise partnership with Google Public Sector to accelerate the state’s digital transformation.

Route Fifty caught up with Rai on the sidelines of the Google event to talk more about New York’s experimentation with AI, how its workers are getting on board and the importance of process improvement alongside technological advancement.

This interview has been edited for brevity and clarity. 

Route Fifty: You said on stage you’ve got several dozen potential use cases for AI. How do you arrive at them? How do you figure out what's good and what could work? The possibilities seem endless.

Dru Rai: Let's take the Department of Motor Vehicles and go back to the metrics. You can look at the call center, you can look at the direct experience and feedback New Yorkers provide to the DMV. It could be, how long does it take to issue a driver license? How many times are calls dropped? It could be any of those customer-facing metrics. Then you have internal metrics, which are probably not customer-facing. You start the customer metrics, and then you peel down to the internal metrics, like, why does my workflow have 15 steps? Why does it take me this amount of time to collate all the contracts?

You take all those business cases and work with the client agencies to prioritize them, because they're not all created equal. We generally start with the ones that have the highest impact on New Yorkers. That gives you a priority from agencies. The DMV can come with two, three, four, or five [use cases] the Department of Health can come with three, four or five, and so on and so forth. Next thing you know, you got probably 100-plus. A lot of those 100, there might be 20 of them that are the same. Now you find that the 100 turns into 25, so you take those 25, you consolidate it, you get the priority, and now you have a good starting point.

The next thing is you start [is] what I call “science experimenting.” Take a model, look at the data, train the data, look at the results and find it's 40% correct, and now it's 50% correct, and now 60% correct. Every time you do that, you're going to find a few things. You're going to find that the data quality is not good, so you need to fix the data quality.

One of the key things is transparency. Who is going to trust AI if you don't understand it? No company is going to say what their proprietary large language model is. At the same time, if I'm the customer service agent at DMV or welfare management and I decide to use the data to help somebody get benefits, I’d better be very transparent about what I'm doing. The best way to create transparency is the agents who are using this, they need to fully understand what they're pumping in and what they're getting out and what the accuracy is. That's why we wrote a policy where it clearly states that no software is going to have a direct impact on citizens. There’s going to be a human being. We're getting some success, we're also not getting some success, but it's just a matter of time where things just get better. At the same time, we need to protect the data and cybersecurity that's part and parcel of building private models. We want to make sure that we are not leaking privacy data.

Route Fifty: You have mainframe and legacy technology too. How do you manage these systems while you also have AI and such?

Rai: It is a very tricky balancing act. The people who know those old systems may not know the new, and we give them the opportunity to learn. The way we think about the balancing act is, for example, we are creating a “no wrong door” for our benefits system. When we wrote the old welfare management system used by the state, cities and counties 40 years ago, people wrote on a piece of paper, they went to an office, they handed it over, somebody takes it, they type it in, then they call you in two weeks and then you go and do more paperwork. Not that the COVID-19 pandemic was a great thing to happen, but the pandemic changed quite a bit. We thought, why can't we get all these things online? There are also disabled people who can't travel, so maybe we send a case worker. So whether a caseworker goes to them, or they fill out a paperwork and reach an office, or they sit in their home, any of these doors are open.

There is no point rewriting the old software into a new platform. What you really want to do is you want to look at what exactly is available in the software and what's not manual, and then you create a platform which takes care of both. The way we are managing is in a very agile way, because mainframes by nature are monolithic. The data and the code sits together. What you want to do is understand the functionality. You break those things into pieces, you look at the manual things you are doing, and then you rewrite the whole thing as a new process, so that the experience of the customer we are serving is far better than 40 years ago. That's a fine balance. Unfortunately, we have to keep it alive until we have this ready, and we are on our path to do that in the next few years.

Route Fifty: So the mainframe will eventually go away?

Rai: Yes. There's a ton of mainframe. The best thing about mainframe [is] it’s very stable, but the worst thing is nobody's trained on COBOL [the coding language needed to run it]. We have to call retired people and ask them to work 20 hours a week just to keep it up and running. They're all going to be slowly replaced with better systems, faster systems, more stable systems, more redundant systems and better user experience systems.

Route Fifty: I did want to touch on the training piece for your employees as well. How are you approaching that? You have a ton of employees, so this is very general, but is there excitement, skepticism, somewhere in the middle?

Rai: The first thing I want to say is, I want to thank the governor. She was able to see things like a visionary. Last year, she announced Empire AI, which is a big investment in the state of New York, a public and a private sector investment. She has instructed us to train our employees, so we're working on a plan which we’re going to publish. 

You start with general training, like what is generative AI? How do you prompt it? That's the first step. The second step would be taking the model and creating a private environment. So, for example, create an environment just for the Department of Labor. That data is inside the Department of Labor. We bring the model in the Department of Labor and train the model in that context only. First, you will have some technical people who can help the users direct which data is relevant and which data is not relevant. In many cases, they can help them tell the data is not accurate.

There will be some super users who really know their job, who also know their data and know the limitations of data. They also know the current process; they also know the current limitations of the system and process. They are the ultimate people we will be using because they are on the front-line, providing services to New Yorkers. We have to take that generic training and prompting all the way to them in a very contextual way so it's more relevant to them. That's why we don't believe in generic training, because the training for a DMV worker is not going to be the same as the Department of Health, it's not going to be the same as DOL. Personalizing it is the key. Generative AI needs to be personalized all the way to the agent.

Route Fifty: There’s been a lot of talk recently about government efficiency at the federal level. Some states are trying to experiment with the same principle, whether it's a Department of Government Efficiency or something else. What does an efficient government mean to you?

Rai: We are a democracy. We need to provide services to our citizens by law. The number one thing is we can't waste money; we just don't have money to waste, and we don't have time to waste. There are a lot of laws and regulations, which we have to follow, and that's built over decades. I'm part of the executive team where I can move faster. One of the things you learn in the private sector — not that public sector employees are any different — is that I don't have the shackles of the laws I need to follow. When you're in the private sector, you follow federal law, state law and the general business practices. But once you're in state government, you have the extra burden of holding all kinds of data. Our website needs to be accessible to a normal person and a person with challenges at the same time. Creating, prioritizing and moving fast and being agile.

That's one of the things in large government programs. You hear that we have to transform this, and we need tens of millions of dollars. That's all good, and we have changed our approach from a waterfall to agile. It's how do you break things down into small pieces? Some of those things may cost you in the long run a little more money, but the benefits are great. First, you get to see the progress. Second, you avoid the risk, because you either fail or you succeed when you do a large program. When you're doing it small, if you fail, you fail fast, and then you get back and recover.

We also transformed our team. We were a giant monolith. We have broken the team and flattened the structure. We have pushed our team straight to our agencies, so the decision making has gotten faster. Instead of going through six layers, now it's one or two layers. Each agency has a deputy commissioner of technology programs and a dedicated team, then we have shared services supporting that team to do the job faster. That was all one structure, now it's broken down.

We have a limited budget, so it’s constant prioritization, which can be frustrating. Now the federal government has changed, and those changes we are watching, we are trying our best, at the governor’s direction. We are business as usual. We need to serve New Yorkers the way we served them yesterday, and at the same time we need to watch out for the changes coming and the impacts they can have.

It’s about agility, flattening of the org, prioritization and being smart about where we spend money. I tell my team we have been doing this efficient way of delivery since I've come and we're going to do that, no matter what.

Route Fifty: Failing fast is all very well, then it happens publicly, and you risk the negative headlines. How do you balance trying something publicly, knowing that maybe it's going to fail, and that New Yorkers might be a little bit irritated with you for a while, but also make progress?

Rai: Nobody wants to fail for the sake of failing. When I talk about failure, there's catastrophic failure, and then there's a calculated failure. Calculated failure is never traumatic or catastrophic. It's more about tweaking the direction. Also, not having a mentality of perfection, because sometimes perfection can be a hindrance to progress. We need to do a better job every day, that's what New Yorkers expect. We don't need to say, “I'm going to create a perfect system once upon a time in the future.” You’d never get there, and nobody likes that.

Providing transparency and failing fast is more about breaking things down into pieces, taking calculated risk and tweaking the approach so we can get to the destination faster, and doing multi-processing as much as possible. You can do that if you break those things in pieces, minus some dependencies. I would say, keep everything transparent. There's no need to hide. We serve the government. Everything is open with us.

One of the key things which we have started implementing is human-centered design. I've asked my agency to go all the way to New Yorkers with a client agency. Don't be sitting behind a wall; my client agency will say those days are over. Digital experience now requires digital expertise all the way. If I'm sitting with my agency, they're going to tell me the requirements, but they may not be able to tell me the digital experience my constituents are looking for. Questioning that all the way to the citizen is new, and I'm really delighted that I have a team which can ask the right questions so I can improve the digital experience, because that's not been a traditional model. What we are saying is, let's engage the citizens as part of the process.