Mihir Gandhi, a former GM for Lyft and current vice president of customer and operations for eightfold, chats with our own Siya Raj Purohit about building models to solve for driver productivity and payment incentives—and about the future of driverless cars.
The full video is below, but here are some of the highlights.
Mihir: I was a general manager for Lyft for the last two years here for Northern California. It was an interesting time to be there. Certainly, we saw a lot of growth, a lot of change. It was a remarkable experience, enjoyed that quite a bit.
Siya: And while you were at Lyft, you developed three models that actually impacted driver productivity and payments. Could you talk a little bit about those?
Mihir: Not only is there competition on the demand side for passengers, but there’s also competition on the supply side. Everybody here knows that most drivers drive for both Lyft and Uber, or multiple ride-share companies at the time. At the time, SideCar was still around. And we tried to figure out, you know, what does it cost per ride, what is the average ride length, what are common ride routes, etc. But there’s multiple externalities that impact all of those data. Things like rainstorms, construction, the Giants making the playoffs. Major events like Oracle World or so on and so forth.
Siya: The Pride Parade, which shuts down SF.
Mihir: Don’t even start me. It was one of my, one of the reasons I’m happy I’m not at Lyft is that… those major events just shut down Market Street because as everyone here knows it debilitates the city from a transportation perspective.
And so, as we thought more about where is the demand coming from and how do we service the demand in those key moments of time, the strategy was really around how do we get people to not switch.
We’re unlikely to be the first—at the time—the first app downloaded. Most people have probably already downloaded Uber. They’re like, OK, there’s some reason to try downloading Lyft. Now, it could be a coupon dropped on them, it could be they saw a billboard, it could be their friends, it could be a combination of those things, and of course, culture and all of those things sort of play into it. But how do we make their first experience exceptionally good? And the first experience has to be not just their first experience, but then everyone else’s experience too. It doesn’t matter if your first experience was fantastic if your second one was terrible, right. So now it’s not just how do you optimize for the first person’s ride, or a new customer’s first and second rides. If the third ride’s terrible then like, you kind of lost a little bit of credibility there.
So the question is not how do you optimize for a segment of the population of your customer base, it’s how do you elevate your overall service levels? And so we did a lot of work that now retrospectively looks unbelievably obvious, which is a longitudinal, geo-temporal analysis of rides and fulfillment rates. What that means is: 8 a.m. on Monday morning in the Marina is a much more valuable moment than 3 a.m. on a Wednesday morning in Pacifica, for example. Many more requests, people who are much more frequent riders, etc. And so as we think about parsing this out, you can go to an nth level of detail, 8 to 8:15, 8 to 8:10, 8:10 to 8:20, what is the right level that you should be going to help induce that. And the secondary question is how valuable is it to retain a rider, and what does a rider’s ridership actually look like?
Now, I used to live in Russian Hill and prior to joining Lyft, I would ride probably three times a week, I would commute three times a week to my office, which was in the Mission at the time. But, you have no idea of what your wallet share is at that point, right. And many companies that you will join don’t know their wallet share. Even Coke doesn’t know their share of wallet. Actually, Coke defines it differently; they call it “share of stomach,” which I think is really interesting. So we took this geo-temporal view of where is demand generally coming from, where is supply generally positioned, and how do we reposition supply around demand, and what are the incentives we should put in place. And so, has anyone here driven for Lyft or Uber?
Audience member: I’ve been in a lot. [Laughs.]
Mihir: You’ve been in a lot. So, I’m not an active driver anymore, but I was an active driver at the time, and it’s interesting how incentive-oriented—and I’ve of course talked to thousands of drivers—how incentive-oriented folks are, and this should be no surprise. Behavior is driven by incentives, right? Now, are these long-term incentives? Short-term incentives? Super short-term incentives? Are they a week long? Are they daily? Are they hourly? Are they per ride? And so now you have a series of variables that you can start testing. And this was the fun part. It was not just identifying, OK, supply and demand, here is where it’s happening, here is how it’s evolving over time. When rain happens and we see demand kind of shift… But then, what incentives are incentivizing what types of drivers?
There are multiple types of drivers. There are the weekends-only, nights-only, full-time drivers, evening, all-day, multiple segments of drivers, right. Some drivers earn a full-time living. It’s a very small percentage of ride-share drivers. Most drivers are what we call part-time drivers, where the income they earn from ride-share is supplemental. And so if their income is supplemental, how do you have them optimize when and where they’re driving? That’s a very challenging thing. Many times drivers will just turn on their app and wait for a ping and they’re sitting in their driveway in Pacifica, Daly City, or Russian Hill. So how do you convince them, hey, you should actually be going to the Marina, positioning yourself to then be able to take these rides, and how do you think about incentivizing them to go back there afterward, right? What do those incentives look like over a longer period of time, knowing that they could very well turn the app off as soon as that ride’s done, and then not be on until the weekend or until the next week or later?
So again, like, there’s this really rich set of data around how do you incentivize driver behavior, and I’m going to say that in a way that sounds somewhat crass, but the reality is that is what it is. You’re trying to incentivize a certain behavior and reward that behavior with different types of per-ride, per-set-of-rides—like, over a period of time, let’s say rush hour, if you complete this many rides during 7 a.m. to 9 a.m. you will earn this bonus, and we will tell you where to go because we know where the demand is. Or a week: if you complete this many rides in a week, and if you want to complete this many rides in a week, you can very quickly do the math and driving between 1 a.m. and 7 a.m. is not going to get you that volume. If you drive between 7 and 9 and then 3 and 6, right—and so helping drivers figure that out, and optimizing themselves, was an interesting problem.
Siya: So now, diving into the hard questions: Lyft and Uber do have self-driving car teams. Do you think that a fully autonomous future is close?
Mihir: No, I don’t. Fully autonomous is really difficult for a host of reasons. And when I say close, I mean like within the next five years or 10 years.
Siya: What do you think is the more appropriate time frame?
Mihir: Twenty to 30. But I do think in our lifetimes, fully autonomous will exist. And by fully autonomous, I also mean, every city, every, you know, region. How many people here are from the U.S.? How many people are immigrants? In your home countries, you guys can think of the towns—I’m from India, I was born in India, like my dad’s town and my mom’s town, there is zero chance there is going to be an autonomous vehicle getting to my dad’s village in Maharashtra any time soon. So fully autonomous to me means that. I think it’s a very high bar.
But I think that it is in the future. I think part of your question is what does that mean for on-demand or gig economy workers or people who start to rely on some of these jobs as part of their income? And the reality is there’s enough complexity at least for the foreseeable future, in things like ride-share, that I don’t foresee this impacting the need for drivers. In fact, I think ride-share is still less than like 2% of all miles traveled by a vehicle in even the major centers of the U.S.
And so, yes, it’s still a very nascent industry in a lot of ways. But the complexity of what’s happening on our roads is only increasing. And what I mean by that is: man, I was even driving down Townsend earlier today, and I haven’t been on Townsend Street in a long time… and like, they’ve blocked off an entire lane for a bike lane, and then they put a parking lane in there. And so there are these changes around how do we have a multi-modal city that allows for bikes, scooters, cars.
And so, as long as these changes are happening and we’re still, like, trying to figure out what the multi-modal transportation mechanism of the future is—and I think flying cars are a ways away, probably further than fully autonomous—there will always be a need for drivers and for a human element of making these decisions of x, y, and z. Like, should I turn left, should I turn right, should I swerve to avoid the car running or the dog running into the road? And those people would need to be compensated for their time.
And so the hope of Lyft and Uber fundamentally changing their PNLs by reducing their driver cost are no time in the immediate future, in my opinion.
Want in on AI? Check out the Machine Learning Engineering Career Track—it’s the first course of its kind to include a job guarantee.