Zest AI’s Mike de Vere: ‘You don’t have to give up economics to be more fair’

When the pandemic hit, lending models didn’t catch up quickly enough. Machine learning, like the kind Zest AI has developed, moved quicker, helping lenders to figure out how quickly things were deteriorating and where. ML models also have been instrumental in identifying and avoiding biases in lending practices.

Zest AI CEO Mike de Vere joins us on the podcast to discuss his firm’s go to market strategy and business model and how the two kind of fit together nicely. Mike shares his views on how to bring technology to lending and specifically, de-biasing models to approve more minority borrowers. 

Subscribe: Apple Podcasts I SoundCloud I Spotify I Google Podcasts
The following excerpts were edited for clarity.

Background in data and insights

If I think about my past, I think I’ve had a consistent track record of building businesses where we’re harvesting insight from data, whether it’s at JD Power, way back in the startup days, all the way to Nielsen and leading their insight business in North America and Europe. And so it was a pretty natural fit for me at Zest AI as far as my background and experience go.

But I would tell you, I’ve learned so much over the last two years. I’ve learned a ton: the type of math that we’re using was actually created after both you and I were out of school. It’s really been an awesome opportunity for me to get up that learning curve.

Strategy and go to market

This is an enterprise software sell. There are many stakeholders that we’re going after within a financial institution. So though we do have client development professionals, it takes everyone from our CTO to myself to the head of our client practice. It’s really a team effort to go within a financial institution and to get them to make that transition. We’re asking them to change what has been in place for years for many of them. I mean, they’re using maths — logistic regression — from the 1950s.

We first have to convince them that machine learning is better math and will be better at predicting risk. That’s the first hurdle and then the second hurdle is to use Zest AI for that. So, it really takes a team. And what we’re finding is the sales cycle itself really comes together when you have legal, compliance, IT, credit risk, as well as a business leader, all involved in the decision process.

De-biasing lending

I would say that this is probably the area that I am personally most excited about. We have unique IP around de-biasing machine learning models. And it’s something where clients have actually come to us. Especially today, it’s all that more relevant. What’s unique about our approach is not only the cool math that we bring to bear to create a more inclusive model. But we don’t create this false choice for executives where you just have to give up economics to be more fair.

We actually have this approach called adversarial de-biasing where we’re able to use machine learning to do tradeoffs between accuracy and fairness and come down on that absolute efficient frontier, where you can literally have your cake and eat it too. You can have good economic returns as well as be much more inclusive at the same time. One of the largest mortgage providers in the US — roughly a third of the mortgages run through their platform — is leveraging our technology for this very reason.

How to partner with FIs

I think there’s two different approaches. We try to meet our customers where they’re at. And so smaller customers tend to need our help. And so we will go in and fish for them. Larger customers like to fish themselves and so our engagement with them tends to be teaching them about our software, sharing best practices that we’ve learned across our clients. And so it’s different engagement at different levels.

Our belief is that over time, machine learning IQ will continue to increase across the market. More and more customers will be able to access the Model Management System and actually start building ML models themselves, generating all the compliance and regulatory documents, models that are inclusive and fair, and then deploy it and monitor it. And that’s the journey that we’re on the next two years.

If you imagine on a larger account that wants to engage with u, and leverage our tools directly, they want to do the fishing themselves. Oftentimes, the team on that is much more highly technical, because they’re going to be pairing off against a team of 200 data scientists at a customer. The flip side is our client engagement team in the mid market with smaller customers will tend to be an ex-Bain or McKinsey consultant who not only has the quantitative chops, but is much more used to dealing with the customer and carrying them through a longer process.

Credit Unions

If our sales and our bookings are any indicator, many credit unions are reaching out and using this time, during the pandemic, to dig in and develop a better overall member experience. And you’re seeing this resilience, you’re seeing this creativity, you’re seeing this willingness to take risks.

Think about their process: they’d sit down with a member and collect 20 or 25 points of data. And then they would put it through a scorecard. It would go through three or four sets of hands before a decision would come back. With our approach, we’re able to build a model that has hundreds if not thousands of data points, that within less than a second is able to render a far more accurate decision back for that number and for that credit union.

Credit scoring

I hear a lot from banks and credit union. I think there’s a fairly consistent theme around them finding credit scores to be increasingly unreliable. And, frankly, to be honest, I think they’re also finding them to be less inclusive on race. There’s this disproportionate impact that affects people of color.

Banks and institutions that are stuck in the old way of doing things, when they see something like the pandemic, they’re rushing around, and it’s going to take them 12 to 18 months to get a new model into production using the old logistic regression approach.