Leveraging Incentives and Data for Social Good: Q&A with Eric Liu

Just a few years after graduating from UC Berkeley in 2013, Eric Liu carved out a special niche in the use of data science as a tool for social change, reforming the government contracting process and carrying out projects that help agencies do a better job of serving people in need. As cofounder and executive director of the nonprofit group Bayes Impact, Liu built a team of engineers, product managers, and researchers that revamped IT operations for Medicare and the French Department of Labor (among other agencies) to make them more responsive to the public. In the process, he helped make contracting more value-oriented by accepting only fixed-rate projects that cost taxpayers far less than traditional pay-per-hour arrangements.

Liu, a 2020 MBA candidate at the Stanford Graduate School of Business, later joined Atrium, a technology-based start-up law firm that is challenging the billable hour business model that dominates the legal profession. Liu, who came out of the venture capital world, spoke with us about data science’s potential as a force for social good and his vision of using government contracts to create socially useful open-source software available to all.

What got you interested in using technology for social good?

My background was in venture capital, investing in data science companies, particularly in finance and advertising. I was looking at improving ad clicks by one percent and things like that. I had this feeling that using data science in the social sector would have a lot of impact. I didn’t see many companies that were using data science and AI to address pressing social challenges, so I did a 180, left my job, and in 2014 cofounded a nonprofit called Bayes Impact.

What does Bayes Impact do?

Bayes Impact has IT contracts with large government agencies to try to improve these huge systems using data science and algorithms. Our goal is to be much better and more efficient than typical government contractors and use our excess funds to launch free, cool software that helps underserved populations.

How did Bayes first get traction in the marketplace?

We started with backing from Y Combinator, the Gates Foundation, and other donors and got our first big contracts in 2015 with Medicare and a couple of other agencies. That really helped kick things off. Toward the end of that year, we launched an office in France — where my cofounder is from — and started working on labor market issues.

Why did you choose to work in the government contracting space?

A lot of what happens in government contracting creates huge amounts of inefficiency. Because of the pay-per-hour model, there are huge incentives to just stack more people on, extend contracts, and never deliver. The contractor can drive up costs without adding value. So what we did as a nonprofit [was negotiate] for projects that were flat-rate and value-based. Those are a minority in the government contracting world, but it was a fundamental thesis behind our organization that by imposing this market incentive on ourselves we produce better results. With a flat rate, there’s a cap on how much we can take out personally, and that incentivizes us to actually deliver.

How do these value-added contracts actually work?

First, there’s an estimate of the amount of work, and people bid for it. Then the contracting agency puts in a dollar amount and sets up a schedule for deliverables. There’s an ability to iterate over time — that is, use a set of milestones that can be adjusted.

What’s the social element in your approach to government contracting?

We use our contracts as ways to build relationships with governments and get a deep understanding of the data they have. We typically negotiate to have a lot of the things we build be open source, so we’re building free things on top of whatever we’re doing. That allows us to fulfill our unique mission and also deal with the perverse incentives in the government technology space.

Can you give us an example of an open source tool you’ve built on top of government data?

In France, unemployment is always a big problem. We were working with the French Department of Labor on a project, and we were able to build what is essentially an AI job counselor that recommends careers and training options. It’s a massively scaled software product based on a unique dataset never released before. We made sure it was completely free so it could reach everybody. Google.org funded us when we first came out with the application. Within the first four or five months, we had 150,000 people registered to use it.

What have you learned about the government IT sector?

The U.S. Digital Service is a federal organization that started in the wake of the Healthcare.gov debacle. They work with a lot of different agencies to improve software, some of which they build directly. What they found to be most effective was to change existing practices, educate CIOs about purchasing behavior, and help write contracts. We’ve seen innovative procurement models applied for Medicare and Medicaid to improve service delivery. It’s probably at the state level where you have the most decrepit systems, the worst overpayment, and often no delivery. There are projects that have had three or four iterations and just never delivered.

Why is it so hard for governments to get IT contracting right?

Part of the challenge is that government contracting models are based on building a physical object with defined characteristics, like a missile. So there’s a misalignment between the need in IT projects for rapid, incremental innovation and constant change, and a procurement system that was designed for products. When you’re building something for a lot of different users, you may not know how they will behave. For example, for a social service worker using the child welfare data system, the interface you develop may not be useful … but with software, you can iterate and test quickly. There’s no marginal cost for delivery. In contrast, if you’re building a tank, your iteration cost is extremely high.

What did you do for Medicare?

We had a project for Medicare that typically would have cost $100 million and taken five years. We delivered it for $1 million and in a year. The contract was to help Medicare improve claims processing by focusing on the quality of care patients receive, a payment model known as value-based health care reimbursement. Under this model, reimbursements are based on actual health outcomes, not fees for services.

This was a huge undertaking. Medicare processes 300 billion claims every year. Our fundamental thesis was that you can have scaled decision-making and incentivize it correctly. The value-based care system we developed is by nature statistical — what’s happening in populations over time, what’s happening to diabetic patients, what’s happening to people with cardiac issues. This orientation allows the system to effectively measure value.

What else has Bayes worked on?

We developed a digital police use-of-force data collection and reporting application called URSUS. It’s effectively the first in the country to get granular, individual, incident-level, use-of-force interactions. Every police department in California is using our tool, which is cool. It’s open source, which makes it easy for law enforcement agencies to publish use-of-force data according to a common standard.

What is your role at Bayes now?

I was a founder and executive director, then executive director of the U.S. office. I left a full-time job with Bayes in 2017, but stayed on the board of directors.

You then joined Atrium, a start-up law firm. What did you do there?

I came on as the second business hire responsible for sales and marketing when there were 15 people. Later, I launched Atrium’s Fundraise Concierge business, which guides startups through the fundraising process. Atrium is both a law firm and a technology company. Instead of billable hours, it works entirely on a flat-rate basis. Lawyers are evaluated based on the value of their work, not only by the number of hours they put in. It has a lot in common with Bayes. If you build a more efficient law firm with better incentives, it will be much better for the client and produce a lot more value. Atrium’s structure includes an equity vehicle that changes incentives to focus on long-term outcomes. In a partnership model, partners take money out at the end of the year, but suppose we want to invest in data systems that will produce value five years from now. Partners may not want to make that investment. What Atrium has done is build incentives for these longer-term outcomes, allowing it to create technology that makes a law firm more efficient.

What are the next opportunities for applying better incentives and leverage with technology?

I think there are many opportunities moving forward for these tools. A key example would be to improve student loans through the use of income share agreements as well as other methods. There are also opportunities to explore using value-based models more extensively within health care. We need a push from the public sphere to improve interoperability of data. This needs to come from the government, because it is costly to establish standards and often against what individual companies want to do. Also, interoperability between different social services would have vast benefits.

— —

Learn more about the Golub Capital Social Impact Lab at Stanford Graduate School of Business.

Follow us @GSBsiLab.

Learn more about Eric Liu.

With writing help from Sam Zuckerman and Peter Cioth.