Transforming Cancer Diagnoses and Treatment with Bristol Myers Squibb CTO Greg Meyers

Evan Reiser: Hi there, and welcome to Enterprise AI Innovators, a show where top technology executives share how AI is transforming the enterprise. In each episode, guests uncover the real-world applications of AI, from improving products and optimizing operations to redefining the customer experience. I'm Evan Reiser, the founder and CEO of Abnormal AI.

Saam Motamedi: And I'm Saam Motamedi, a general partner at Greylock Partners.

Evan: Today we're talking with Greg Meyers, Chief Digital and Technology Officer at Bristol Myers Squibb.

Bristol Myers Squibb is one of the top ten pharmaceutical companies, with about 30,000 employees and roughly 45 billion dollars in annual revenue, making specialty medicines in oncology, immunology, and neurology. Greg runs an organization of several thousand people, which gives him a rare view into how AI is reshaping drug development from the lab to the doctor's office.

A few things stuck with me from this conversation:

First, BMS is the only pharma company with four commercially scalable AI products in the hands of doctors. In a year-long lung cancer study across 14 clinical sites, AI-read CT scans doubled the number of newly diagnosed patients and caught 116 new cases. When the average lung cancer patient gets about five years after diagnosis, early diagnosis changes lives.

Second, BMS runs an "AI accelerator": six to eight engineers and twelve weeks to answer one question: can AI do X or not? Greg deliberately keeps the teams tiny so they're too small to fail in any public or embarrassing way, which means they can easily throw out the work and start over. At a big company, permission to fail quietly is the thing that actually unlocks scale.

And finally, Greg's advice for activating AI in a tradition-bound industry is to stop asking how a process works and start asking why it works that way. Most of the constraints people treat as fixed are just old assumptions that AI quietly made obsolete. He calls it the "clean sheet": let go of why there has to be a spreadsheet, and ask what the spreadsheet was ever really for.

Evan: Greg, first of all, thank you so much for joining us today. Maybe to kick us off, could you give us a little bit of background about your career, maybe how you got to where you are today?

Greg Meyers: Sure. Thanks for having me. Well, I started off in college not really knowing what I wanted to do; whether I wanted to be going into business or going into computer science. And I was sort of really lucky that my career has been this intersection of those two things the whole time. When I came out of college, it wasn’t obvious that business and computers would intermix as well as they have, so I’ve been really lucky to do that. I’ve spent about half my career in life sciences, which is where I am now as the CTO at Bristol Myers Squibb, and then spent about half my career in other industries at Motorola, and in a large agricultural industry, so just different industries. And I’m really glad to be back in life sciences. I think there’s probably never been a better time to be in technology and in life sciences, so it’s been really cool.

Evan: Do you mind giving a sense of the scale of operations, a little more about your role, what you’re leading at the company?

Greg: Yeah. Well, Bristol Myers Squibb is one of the top ten pharmaceutical companies. So, we make products that are primarily in oncology, immunology, and neurology. These are considered specialty medicines, so these are people with serious diseases. And that’s sort of what the company is. We’re about 30,000 employees, about $45, $46 billion in revenue per year. And I run an organization of several thousand people, all trying to figure out how to apply technology to maximize the opportunity to help BMS really deliver more medicines to more patients faster.

Saam: What are some of the most exciting use cases of AI that are currently being deployed at BMS? And if there are several that you think might surprise our listeners, I’d be curious.

Greg: Yeah, I think probably the most surprising place where AI is making a difference is as you get closer to how doctors are treating patients. Sure, everybody has heard about the opportunity for AI in diagnostics, but we’re actually pretty involved there. In fact, we’re the only pharmaceutical company that has four commercially scalable AI products on the market being used by doctors. We had one example in oncology. This is for non-small cell lung cancer, which is a pretty common form of cancer that’s actually quite hard to detect early because patients often don’t produce symptoms.

So, the goal here is to try to find cancer tumors before they become a problem. The way that’s typically done today is a CT scan would be taken, and a radiologist or radiation oncologist would take a look at those scans and try to see if they can find, using their eyes, where those tumors are.

And in one study that we’ve done with one of our partners using AI to read those CT scans, this is about 14 different clinical sites, over about a year-long study that we did, it led to a two times increase in the number of newly diagnosed patients with non-small cell lung cancer. It flagged about 2,600 patients in that one study alone and identified 116 new lung cancer patients.

You know why that’s important? The average lung cancer patient, by the time they’re diagnosed, is typically given about five years to live. So, if you can actually detect this thing early, you’re making a big difference for those patients and their ability to seek treatment sooner.

Evan: You guys are already a leader in using AI in all these different ways. I’m kind of curious, when you look over the next two or five years, what are other things you think are going to more fundamentally transform the industry?

Greg: There are 40 trillion cells in the human body. Every one of those cells has about a trillion molecules in it. So, if you add that up, that’s seven octillion atoms that make up the human body. And there are cases where one cell or one atom out of place could potentially cause, cure, or prevent a disease. So, there is a huge computational space where just the surface has barely been scratched. This idea that you can mine this information without having human beings do this very painstaking part of the process, really what we’re excited about is the ability to uncover effectively dark data inside biology.

And a lot of times, if you look at things like even trying to get something to be successful in what we call the preclinical stage. So, if you find something that you think might be workable, if you look at the whole pharmaceutical industry, if you can improve the success even in the early stages of R&D for a drug, because very, very few of them actually make it to the end. They typically will fail somewhere in the clinical trial process. Like 80% of drugs fail before they even get to the stage where we would apply for the FDA.

If you can get just a 2.5% increase in success rate by just how you design the molecule earlier on in the process, that would actually generate probably 30 new medicines or new drug approvals over the next ten years, in the United States alone.

Saam: The stakes are also a lot higher, right? You’re operating in a highly regulated environment where mistakes have real consequences. So, how do you square that with AI iteration and AI velocity?

Greg: I think you have to pay a lot of attention to the quality and quantity of data that goes into these tools. You’ve got to spend a lot of time doing context engineering, so that the AI understands all the different data sources it has and what they all mean. You have to teach it how to think the same way that you would teach a person. I describe this often: if you were to hire a new employee, you would have to sit down and explain to them how we do our work, how we think through things, and what this data means. You have to do the same thing with AI. So, there’s a lot of work in doing that. I think people almost underestimate how hard it is to get AI to the level of quality, particularly, as you point out, in really high-stakes decisions, that it’s good.

But I can say that we are really reaching the inflection point. I would say we started this journey maybe three years ago, and three years ago we really couldn’t trust a lot of what it was producing. Now it’s reaching the point where, if you do a good job with the things that I mentioned, it can do about as good a job, if not a better job, than humans at a lot of these really intricate analyses that are really important. So, I think just as time has gone on and the models have gotten better, and also we have learned the hard way how to engineer these things the right way, we continue to build a very high confidence that they can play an important role, and that you can manage things like hallucinations.

Evan: Are there lessons from what you guys have built over the last ten, twenty years that are informing your usage of some of the newer models?

Greg: Well, it’s funny that you said that. We have not historically thought of ourselves as a data company. It’s totally obvious, as you say, why you would think we would think that. But we’re actually typically a science company. If you think about what most scientists do, they’re trying to run an experiment, right? You’re usually trying to fail to reject the null hypothesis, if you remember back from college. And whatever data gets expressed or exhausted off of that experiment is a byproduct. It’s not the point, historically.

So, what we’re finding is we do have to think of ourselves as a data company, because to models, negative results matter as much as positive results. I like to describe it as a game of connect the dots. The more dots that you have on the page, the better the drawing is going to appear. But historically, if you’ve been very experimentally focused, you’re going to focus on the things that succeeded and you’re going to ignore the results that failed, because you don’t need to know them to move an experiment forward.

So, it really has caused us to have to rethink how we store data, how we annotate it, how we curate it, because those models need all of the data. And in many ways it has been a real change, a cultural change in terms of how we think about what we do, where the data is not simply the exhaust from an experiment, but it becomes a gold mine of data that you can then retrospectively use later to help inform other experiments down the line.

Saam: I think any company can become AI-native. I’m curious, what’s your philosophy on this? How have you transformed, or are in the process of transforming, BMS to be AI-native at the core of the company, versus just something you do as an add-on?

Greg: What becomes really challenging, and part of when you think about how you make AI foundational, is you have to be able to ladder up and look across the processes of the company and ask yourself: okay, which problems require fundamental rethinking? And then being clear about what stuff should continue to go through the continuous improvement route, and what do you really need to start completely over again. I think the other thing is that when you’re taking on something completely new, it’s very intimidating because, as you pointed out, we’re in a highly regulated industry. There’s a lot of risk.

The cost of failure is very high, so you can’t do that in a Big Bang way. You sort of have to have an eye for it. And almost every conversation I run into starts off something like, “hey, can I do X?” My answer is, “I don’t know. We have to go see.” And the ability to take this concept and get it down to a really clear, testable hypothesis, and actually get a very small group of people to prove or disprove whether AI can do X or not, at least in a controlled setting, gives you the ability to, in a short period of time, get to the level of conviction that you build internally. Also, a level of self-confidence that you can take on a really big, completely new way of approaching a problem. And I think that is sort of the secret to scale: you create an environment where you can fail quickly and quietly, or you succeed quickly and quietly. And that basically is a path to proving that your hypothesis was right.

And that is where you start building enrollment and self-confidence that you actually can do something fundamentally different.

Evan: Let’s hear your thoughts on the cultural leadership required to create that environment where people are experimenting, because it’s kind of the nature of your business, experimentation, but also making sure it’s being done in the right way. How do you balance innovation and rigor? Any cultural rituals or guidelines you share with the team to help activate that mindset?

Greg: Yeah, we created something called the AI accelerator, which is an environment that’s agile-based. What it does is it gives people a lot of room. These are like six to eight people really trying to pursue the answer to the question of: can it do X or can it not do X? You give them a common set of tools. You give them the best engineers we have in the company, and then you give them a time-compressed period. We typically give them two-week sprint increments. We give them six sprints, so twelve weeks. And over those twelve weeks, every two weeks they come in and present what they’ve achieved.

So, a couple of things happen. One is, culturally, because it’s a small team, you give the group a lot of latitude to fail. A lot of times at big companies, when you create these big projects that are very highly visible and expensive, you create an incentive system where it’s just not okay to admit that you’ve failed in some way.

But when you have a handful of people working for six or seven weeks, you can totally throw away the work and start over again. So, I think making them too small to fail in a public or embarrassing way was really important. And I also think this exogenous pressure, that every two weeks you have to demonstrate that you are starting to triangulate in on the core hypothesis, because groups will tend to get really distracted by wanting to build UIs, and what are we going to call it, and what does it integrate to… you get really lost in the sort of stuff that’s not really important. If you really get them to focus on the kernel of, “can I do this or not?,” you then create room where you’re not getting distracted by the window dressing or the theatrics about something, and you really focus on the core hypothesis.

So, that accelerator has really helped us keep the teams really tight and really crisp, allowing them to show progress quickly. And if they don’t show progress, there’s no embarrassment in the answer being no, AI can’t do this well, and we move on to the next thing. If one of your teams comes to you with an idea and they haven’t built it yet, they haven’t entered the accelerator.

Saam: Do you have any interesting mental models on how to evaluate if an AI idea makes sense, or if it’s worth investing in?

Greg: We learned our own lessons. I think the first year, this is going back maybe two and a half, three years ago, we didn’t really know. So, we kind of let a thousand flowers bloom, and that didn’t really work well. And the reason it didn’t work well: remember how I said earlier that at a big company someone has got responsibility for some subset of the process? Everyone was trying to build a tool that helps them do their part of the process, and nothing actually was working together. So, what we decided to do was to take more of a top-down approach.

And so I sat down with my peers, who are the people that run the company, and we just sat down and said, “okay, in research and development, and commercial, and HR, and finance, what are the big bets that you want to make? What does your gut tell you are the places that AI can make the biggest difference for the company?” And then that actually creates a natural filter by which all the ideas come through.

And then we let them go to work with their own teams. So, the research and development team did this. They sat down internally and said, “okay, if I’m in charge of drug development, what matters most to me?” Well, what matters most to me is: can I accelerate my clinical trials? Can I lower the cost of a trial? Can I actually increase the probability that this trial will be successful? So, if you have those three anchoring points, what becomes emergent are the obvious things that you need to go do. So, we kind of blended this top-down and bottoms-up approach that I think created a natural filter of what the surface area of opportunities was for us with AI. And then we had this methodology that those groups could plug into to actually demonstrate that their concepts worked.

Evan: How do you activate that? A lot of the leaders we talk to struggle, where the obvious applications are doing the thing a little bit better, a little bit faster, but it’s very hard to step back, especially across a business process that spans multiple teams or multiple functions. Any advice about how you get more of this transformation versus kind of the digitization of the thing?

Greg: I think part of what I’m seeing is that we’re going to be able to reach over and ask, instead of asking someone in the business how something should work, asking them why does it work this way? And really trying to understand why the process is what it is today. What you’ll find when you really ask that question is that there are certain reasons for it that are baked-in constraints, that when you really look at it from an AI perspective are no longer constraints, where you can revisit whether they’re actually constraints. So, you almost have to completely explode the process. And that’s why we use the term “clean sheet.” It’s like, okay, let go of all of our biases about why there needs to be a spreadsheet to do this, and just ask: what is the spreadsheet doing and why does it exist in the first place?

And then you’re really starting to get to this space of what might be possible. I think what’s really hard about this… I describe it like a Venn diagram. On one side of the Venn diagram is: what does technology make possible? And that’s a constantly changing picture. You can’t expect people who are accountants, or people in HR, or people that do clinical trials for a living, to know that. They don’t know the difference between what Opus 4.6 and Opus 4.8 does. And that’s sort of our job, right? On the other hand, as an IT professional, you can’t just get away with, “I don’t really know how the process works.” So, it’s where these two circles meet, almost like a Venn diagram, that becomes the sweet spot where you’re really looking at what technology makes possible, and you start questioning the foundation of why something is done the way that it is.

This is where you start finding these sort of discontinuous intuitions that you really want to start testing about whether certain things are possible now that maybe weren’t before. And the other thing that’s hard about this is you have to be humble about what you think is true, because something that maybe you tried with an LLM six months ago and it didn’t work well, can work well now, because you’re like five versions ahead of where you were six months ago. So, that’s also hard, that you can’t be too certain about what will or won’t work. It’s a constantly changing picture.

Evan: So, for some of your peers that are in older industries, where they have a great way to do this that’s worked for 100 years, and now it’s faster, how do you activate that imagination about what’s now possible?

Greg: Einstein said this: that if he had an hour to solve a problem, he’d spend 55 minutes thinking about the problem and five minutes solving it. IT people will consume their entire time thinking about solutions, and business people spend their entire time thinking about problems. And I think we’ve got to... You can’t lead with AI as a hammer and, “now let’s go look for some nails.” To really influence other business executives or other parts of the organization, you’ve really got to be paying attention to what their problems are and what the opportunities are, not looking for an excuse to deploy AI or some other technology. That’s just a recipe for a solution looking for a problem.

And that’s, to some extent, a reputation that we as a profession have built over the years. So, people really want you to meet them where they are. When I’m sitting down talking, for example, to our head of research, he doesn’t really want me talking about all the startups that are doing drug discovery that he should be considering. He wants me to meet him where he is, with the challenges that he has, like increasing pressure to deliver more molecules, and a growing library of compounds to choose from, and how you use computers to sort through that. So, I really spend most of my time thinking carefully about the same problems that my peers have, but I have a unique way to look at them because I can connect the dots, because I’m lucky enough to see across the organization. And I also have experience in other industries where I can see if this is maybe something that’s worth trying, or we should consider. But it’s not coming at it from a solution mindset. It’s coming at it from a problem or opportunity mindset.

Evan: At the end of the show, we like to do a five-minute lightning round where we give you basically five questions. They’re impossible to answer in the one-tweet format, but that’s what we’re looking for. So, maybe just to start: in this AI era, and we’ve talked a lot about the change, how do you think companies should change the way they measure the success of the CDO or CTO?

Greg: I think, as a CDO or CTO, at the end of the day, we’re enablers of the core business strategy. I’ve always been a bit allergic to an IT strategy, because I don’t think it is a thing in and of itself. There are certain ways you’ve got to run the organization, but the IT strategy should be the business strategy. And so, if the business is trying to achieve whatever sales growth or cost reduction or whatever it is that it’s trying to accomplish, I think it’s the CIO or CDO’s job to basically create the flywheel that allows those results to come better and faster than they would have otherwise, by using technologies and access points. So, I try not to overthink that. I look at what my boss cares about trying to accomplish, and I try to set my priorities to match. How can I help change the outcomes that the company is trying to achieve, just using it as sort of a leverage point to achieve those things?

Evan: A lot of people I talk to try to get input on, “hey, what are the kind of quick wins?” If you’re trying to kick off the major transformation and put some points on the board, if you were to give three projects that would be the must-do, what would they be?

Greg: Right. If you haven’t done it already… Well, I’ll tell you how we got started, because I think that was the quick win. The first thing we did was we gave all the frontier models to all the employees. So, we created a tool. We allowed the end users to pick a model, and we didn’t spend a lot of time worrying about costs and restricting them, because the first thing is you just have to get employees using the tools. You’re not going to get anywhere if the employees in your company don’t know how to use AI, other than maybe to summarize an email or to draft a performance review. So, you’ve got to expose those tools to people. And it’s not hard to do that. There are many that are out there. 

And then you have to invest in the training. There are people in their personal life that have figured out that these tools are great to do travel planning or things like that, but you’d be surprised that when they look at the blank screen and have to ask a machine a question about their jobs, it’s not always obvious what they should be doing. So, you do have to invest in training people on how to use the tools. And then it becomes this self-reinforcing process. As people start to experiment and use the tools, they start to find more and more uses for them. And I think that really becomes the flywheel that starts to expand the types of ideas that you can choose from to apply into the company.

So, I think the quick win is just making sure that your employees have access to the tools, even if they’re just the out-of-the-box LLMs, and that you’re training them on how to use them. That becomes the playground by which a lot of the really great ideas come next.

Evan: What do you think will be true about AI’s future impact in the world that maybe many of your peers would consider to be science fiction?

Greg: Technology, particularly things that seem really big, always has a bigger impact in the long run than you expect, but it takes a whole lot longer than you think it will to actually achieve that. So, my base case is that AI is going to have the biggest impact of our lifetime, but it’s going to take a lot longer than people think it’s going to take.

Saam: So, maybe one final question: what is your primary advice for a new or aspiring IT leader stepping into their first major leadership role?

Greg: Yeah, I mean, if you’re a technologist stepping into an executive role, I guess the first thing I would say is that psychology and sociology trump technology every time. You have to become a master of psychology. You really need to understand the incentive systems of your company, the incentive systems of the other executives you work with. And I think Charlie Munger said, “if you show me the incentives, I’ll show you the outcome.” That’s 100% true. Even back to your last point around timing, the reason why I’m more bearish than, say, Sam Altman or Dario about how fast AI is going to change the world isn’t because of the technology, it’s because people’s willingness and ability to adopt it is the thing that I think always takes longer than expected.

So, a lot of what a technology executive’s job is, is having the right conversations for possibility and then really building the enrollment that changes reality. Because at the end of the day, part of what you have to do is to convince people that they can do things that have never seemed possible before. That’s really hard. And you have to do that by understanding the incentive systems in which they operate, but also some of the psychological biases they have, the concerns, the fears, the worries that they have. And I think that’s really important.

Evan: Thank you so much for joining us today. Really excited to have seen you, and looking forward to chatting again soon.

Greg: It’s my pleasure. Thank you very much.

Evan: That was Greg Meyers, Chief Digital and Technology Officer at Bristol Myers Squibb.

Saam: Thanks for listening to Enterprise AI Innovators. I'm Saam Motamedi, a general partner at Greylock Partners.

Evan: And I'm Evan Reiser, the founder and CEO of Abnormal AI. Please be sure to subscribe so you never miss an episode. Learn more about enterprise AI transformation at enterprise software (dot) blog.

This show is produced by Abnormal Studios. We'll see you next time.