A Data-Driven Reality Check on AI in Business (Interview with Tom Davenport, Babson College)

0:00Hi, everyone. I'm really excited about today's episode. I've got an interview for you. My guest today has spent decades studying how organizations actually use data and technology. And you may have heard of some of his work, even if you don't totally realize it. He's left fingerprints, if you like, all over the language that we use to talk about data science today. If you've ever heard the phrase, data scientists have the sexiest job of the 21st century, that was him. It came from a 2012 Harvard Business Review article he co-wrote with DJ Patil. He also coined an idea that you might have heard of too, competing on analytics.

0:39Tom Davenport is the President's Distinguished Professor of Information Technology and Management at Babson College. He's a fellow of the MIT Initiative on the Digital Economy and a longtime senior advisor to Deloitte's analytics and AI practice. He holds a PhD from Harvard, and he's written or edited more than two dozen books. These include some of the first books ever written on business analytics. He's written about knowledge management. He's written a lot about enterprise AI. And he's a prolific writer in the article space as well. He's published over 300 articles, which makes him one of the most frequently published authors in the Harvard Business Review's history.

1:17But when I reached out to Tom, I wanted to have him specifically because he's got a really unique position. He isn't only a researcher and he isn't only a consultant. He's both of those things. And in particular, he's right in the intersection point, the overlap area in the Venn diagram between the two. So he goes and talks to executives all day as some of the biggest companies in the world. He's watching what they're really doing with AI. And then he comes back and distills what he's seen and tells the rest of us the truth of it. As you listen to him, you'll come to appreciate, I think, that he's very clear-eyed, he's very data-driven, and he's a bit skeptical in a way that cuts through an awful lot of the hype around AI.

1:59So with that, enjoy this interview with Tom Davenport.

2:07Hi, everyone. For this week's Linear Digressions, you have the pleasure, as do I, of listening this week to Tom Davenport, an old buddy of mine and maybe someone that you've read some of his work, whether you realize it or not. Tom's got some great perspective on AI, but also a lot of other technical paradigm shifts that have happened in business. And he's got, I think, one of the best calibrated, most realistic reactions out there to what is actually happening with AI and businesses. So I'm delighted to have him on the podcast this week. Tom, thank you for joining us. Thanks, Katie. Happy to be here.

2:47Good to see you again. You're listening to Linear Digressions. So jumping off from that intro, as I mentioned, you have been working for quite some time looking at how businesses are approaching and integrating paradigm change after paradigm change, starting with analytics, various IT revolutions. When you and I first met each other, data science was, I think there's someone I know who called it the sexiest job of the 21st century. We'll come back to that. Who was that anyway? Who was that? That was Tom, guys, by the way. Pointed that phrase that you might have heard. So now we're sitting here talking about AI.

3:27And I want to start by calibrating with you. Everything feels like a big deal at the time. So does this feel like the next phase of this periodic wave of new technical approaches come in and revolutionize business? Or does this one feel qualitatively different to you? I think a little bit of both in that it's still, as I tell my students, it's still trying to draw lines and curves that minimize the distance between themselves and errors. And we've been doing that for a long time. And even the most complex generative model still works that way. But the predictions now are linguistic and that I think has made a huge difference in terms of the number of people who can understand something about it.

4:25It makes it much easier to work with these models. You don't have to know anything about technology or statistics really to work with it, which is different from data science and it It just feels, I don't know, smarter to most people. I'm not sure it really is smarter, since it's all just predictions, but it's easy to lose track of the fact that it's just predictions of the next word or the next image component or whatever, and people are dazzled by it, which not very many lay people were dazzled by a traditional, what I call analytical AI, machine learning model. They didn't understand it. They didn't know what to do with it, but people know what to do with the output of generative AI.

5:16And so put another way, when you go and talk to especially leaders inside of these organizations, like they're, of course, very engaged with each of the new waves that they need to bring into their companies to get them into the next phase. Are they engaging with AI in a way that's different or deeper than you might have seen them engaging with? I don't know, maybe an initiative around digital transformation or data consolidation or one of these other things that's very important, but a little bit maybe more technical or dry? Some of them certainly are. And I think people like Ethan Mollick are smart to say, you really need to use this a lot to fully understand it and so on.

5:59But I think they're very busy people and they don't really have time to spend a lot of time keeping up with what the latest stuff is. And so if you see these surveys of how many, what percentage of CEOs and board members in particular, I think board members who are really old radically overstate their understanding of this technology. And so I think organizations can get into trouble. And sometimes I feel like I'm on a one-person crusade about this, but I don't want people to believe that AI was invented on November 30th, 2022. And so I continually tell them, oh, there's a really important set of AI tools that you still need to be using that are still important to your success with AI that you don't even consider in many cases.

6:57And for senior executives, traditional machine learning or analytical AI is really good at helping you make better decisions because it's predicting the future from structured data in the past. And that's a really useful capability. And people just forgot all about it in many cases. Do they want to hear that when you say that though? You say, yeah, we solved this problem back in the decision tree era. Like why do you need an LLM to tell you whether, I don't know, this is a fraud case or not, or somebody's mad in the conversation? I try to say you need all of these tools. And I think there's some really powerful outcomes that can result from combining generative AI with other forms of AI, but they just don't really, it just doesn't really register.

7:50They want to know, should we be using Anthropic or ChetGPT or maybe Grok or God forbid? And I don't think that's the kind of problem they should be worried about. Well, and so I wanted to ask you this kind of at the end of our conversation, but I'm going to pull it forward to now. I gave you a little bit of a ribbing a minute ago about being the progenitor of this sexiest job of the 21st century label. I'm not sure if the test of time has been kind of that one. I think maybe a foundation model researcher might have beat out data scientist. But in all seriousness, this is something that I think about as a data scientist.

8:27And as I advise other folks who are also in those shoes, we're going through a version of what many other fields are right now, trying to think about where the human relevance remains in a world where there's magical algorithms that seem to be able to do everything and even more week after week. So I think I hear you arguing for the continuing relevance of more classical data science skills, AI and ML approaches. What's your general thesis right now on what a data scientist worried about their career should be concerned about and what they they should double down on as some of the core skills that an LLM just isn't replacing right now.

9:11Yeah. And by the way, I would say that the foundation model researchers are data scientists. Good answer. Good answer. At least that component of that aspect of the job, that variation of the job is still quite sexy. This has been an evolution that generative AI has accelerated. It's gotten easier and easier for non-technical people to do data science work, aspects of it to interact with complex algorithms. And we've had automated machine learning capabilities for roughly a decade now or so that started to make that much easier. And now we have generative AI where you can, with a two or three line prompt, You can get several pages of machine learning analysis now.

10:08So I think what's always been happening is that the sort of garden variety data science tasks about getting people access to algorithms, statistical algorithms, has become easier and less, I think, necessary. because people can do it to some degree on their own. For day-to-day work that doesn't involve life or death situations, I wrote a book a few years ago now on what was then called citizen data science. Now, we read a little bit before the vibe coding term or vibe data science came out, but I remember somebody at Microsoft who was in charge of this said, hey, if people want to create their own machine learning models for marketing.

11:00It's not as if anybody's going to die if they don't do a great job of it. And then tons of people were actually doing that. I just think it's gotten easier and easier. It wasn't, when I first started out in the mid-1970s with statistical analysis, it wasn't that hard then. It wasn't point and click, but it wasn't, you didn't have to write Python code or even those days it probably would have been Fortran code. I just think it's gotten easier and easier. And so the requirements to be a professional data scientist have changed. You don't just have to give people access to algorithms. I think in many cases, you have to be responsible for other things.

11:45You and I co-authored this article. I always found it quite amusing that we had a hard time getting it past the peer reviewers' deployment as a critical data science capability. And many of the peer reviewers said, deployment? Why would you worry a data scientist's brain with that crazy idea? And now I think more and more organizations have realized, I don't know if I ever mentioned this to you, my favorite phrase is no deployment, no employment for data scientists. Yuck, yuck, yuck. You have to work out relationships with people and stakeholders and worry yourself with change management related issues and all those sorts of things that historically data scientists were not trained to do and didn't worry about much.

12:36You gestured towards something in your answer that I want to pick up on about how with generative AI tools, there's a lower bar for entry for folks who are not data scientists to produce these data scientist type outputs. And that's true in many fields, right? Like you can vibe code even if you're not an engineer. You can write a PRD if you're not a product manager. But the folks who have those jobs will often look at the output that they're a specialist in and say, this is not that great. People call it AI slop, work slop. I think I was reading up a little bit some of your recent work right before this interview, and you introduced the concept of process slop, which was a new one, but I actually love.

13:18I guess my question here is, as you're seeing this actually get picked up in organizations, yes, it lowers the bar. There's also proliferating AI slop. What did you mean by process slop? Is that the thing that we're all going to be calling it in six months or a year as a particularly pernicious form of this? And is there a tipping point at which having to review your colleagues' low quality AI output garbage mitigates any of the efficiency gains of having the button that made the output in the first place? Yeah, I think that might be officially work slop, and it only becomes process slop when it starts to cut across an entire business process, often interorganizational.

14:03So we were looking at process slop as something like, I don't know, HR recruiting, where you, if you're looking for someone to fill a job, you might have AI create a job definition for you. and you send that out into the world and a potential employee sees the thing or maybe it's told by AI that it exists and they have AI customize their resume and a cover letter and so on and send it in and then AI reads what they send in and evaluates it and creates a rejection letter in most cases to say we don't need you and the net result is people have lost faith in that process. And it's happening in a variety of domains.

14:55College admissions, the same sort of thing is happening. Scientific research, more and more things are being generated by AI. People are reviewing articles with AI, having AI write reviews. So it's just, I think, lowering trust in a lot of different areas. And I think at some point we'll see this wasn't quite the productivity boom that we thought it was because we have to clean it up, whether it's just your co-workers or the entire business process. Have you seen anyone successfully pull out of that downward spiral? I think we see the problems accelerating. What I'm wondering is if there's a solution that we know of yet.

15:41I think there's still much more belief in the productivity gains from it than the downside. And we don't really measure in the vast majority of organizations, we don't really measure carefully the productivity that you get from, at least at the individual level, AI, from generative AIs. I don't think yet, although I think there is a backlash that is starting to emerge to AI, and ironically, it's coming from young people who use it all the time. In many cases, it's hard to say, yeah, this isn't really going to make me the kind of writer I want to be or the artist or whatever, and I'm going to stop using it or boo the commencement speaker who mentions it or something along those lines.

16:30By the way, if anyone wants to hire me as a commencement speaker, I will not sing the praises of AI. You are, yes, I would characterize your general stance on it right now as realist. I think you're in an interesting spot, and I did want to ask you about this explicitly. I see some of the numbers out there of how leaders of companies are investing in AI and 90 % of them say that I think this is going to bring a lot of value to my company as we implement this and I'm pouring money into it. It seems like it has very strong adoption enthusiasm within that class. You talked to many more of these folks one-on-one, so I'm curious if there's complexity that hides underneath that number.

17:14But I also can contrast it, though, with this emerging research that there's a lot of backlash, exactly as you said. Majorities of sort of regular people don't like AI, certainly don't like it everywhere, or they may be resistant to aspects of it. Some people reject it entirely. So we have those two pictures that are in tension. And I think that you have some perspective. You talk to plenty of people that are probably on either side of those divide. Is it just different perspectives looking at the same thing? Is there something that that picture isn't picking up? Like, how do we reconcile the polarity that is emerging around this topic?

17:55Yeah, I think it's in part how close you are to it. And you were asking before about executives. And I think many surveys say that they have a much higher opinion of the potential, at least, productivity gains from AI than people who work with it on a day-to-day basis. Although they are starting to say, I'm having a hard time measuring these productivity gains. gains. And I, in many cases, converted to a, sure, you can have individual use of sort of what I call spray and pray, AI co-pilot for everybody or whatever. But if you really want to get economic, measurable economic value, you need to have enterprise level use cases for it that would change some aspect of your business significantly if you succeed with it.

18:46And I think we're slowly coming around to that, for example, many people probably saw, many of your listeners probably saw that MIT NANDA study, 95 % of generative AI use cases have zero value. And I think it was really not a very good study, but the zero value was basically zero measurable value because 95 % of the use cases are individual productivity oriented and nobody's really measuring individual productivity. It would be very difficult, I think, to measure it across a large organization. So basically, nobody does it. And so for folks who might be working at startups, or they've been in smaller companies, they haven't been inside of an enterprise that might have an enterprise use case, can you maybe tell a story or two about examples that you've seen that you think maybe illustrate some really good cases of enterprise AI.

19:48And what did they have to do? What actually went into identifying and then executing against something that really delivers value instead of just good vibes? Yeah, it's hard work. The nature of deployment in a large organization, if I'm looking at software development, for example, and I think that's certainly one of the areas in which people are getting value, you have to have, in most cases, I think you have to have somebody review the code. And I think many companies are in a situation where they're generating more code than they have humans to review it, which is somewhat problematic. It means that people, if they're not going to be generating code as much in their time, then they should be spending more time understanding the requirements for the solution and working with stakeholders of it and so on.

20:42And they may or may not want to do that. So you might have a mismatch of skills and personality there. You have to work on those kinds of things. I think in many cases, you have to shift toward a more of a product management mindset that says we're not successful when we finish the program. We're successful when people are using it and we should be monitoring its usage and so on. And so that's a cultural shift for a lot of organizations. So just in that one area, I think in many cases, it reduces the productivity improvement from 50 % to 20 % or something like that. It's still pretty good, but it's not quite as revolutionary.

21:26And I think the same thing would apply for a customer service application, which is probably the other most common use case. or maybe it's FDA submission for a pharmaceutical company. That's clearly a mission-critical thing. If you do it a bit faster, that would be great, but you've got to be really careful what goes into it. The same with a legal brief and a law firm, and obviously a lot of companies have not, a lot of law firms have not done that well, and they get in big trouble with the judges and so on as a result. So consulting firms are getting into big trouble with this sort of thing. So people are not being careful enough, I think.

22:07And this is getting at another challenge in some of these AI pilots that I think you've called out and called out attention in a very interesting way. I'm going to recap it for you what I read, but I'd love to have you kind of expand on this, which is the human in the loop we've found is still pretty mission critical for not having the AI run off the rails. but at the same time if the human is always in the loop then they're kind of not necessarily getting efficiency gain you're not sitting on the beach while the ai is doing your work flawlessly you are watching over its shoulder which for the record many people hate to have as a job not that satisfying of a role i think to play yes exactly and so where are the gains coming from if i'm spending all of my time watching the ai do the job rather than doing the job myself i'm still sitting there for eight hours and the work is still more or less getting done.

23:03So I think that's a putting those two pieces together and realizing these don't go well together, efficiency gain and high quality output, at least not anymore. Number one, I think that was like a really great point that you crystallized. And do you see that understanding also starting to gel for some of the organizations that you work with, or they're really thinking about the human in the loop and being like, oh, once we do the calculation, we can't take the human out of the loop right now, but doesn't really do us that much good to leave it in. How are they thinking about these organizations?

23:32So those executives who are very excited, how do they think about that? Yeah, I think it is starting to dawn on people. That's why we're starting to see some issues around the valuations of AI companies, maybe a question about, do we need a data center on every corner in America? I don't think and I don't want the whole structure to collapse, but I do think that people are seeing the leaks in the structure. The air is starting to escape a little bit, and we need to think more carefully about value. We need to think more carefully about cost, and more and more organizations are trying to counter the token-maxing idea.

24:21crazy in some ways that they weren't paying attention to cost, but I don't know, maybe Anthropic and AI just had really slow billing processes. They didn't know how much they were spending. But in any case, I think we're starting to see a realization that to really get value from the technology, you have to be much more careful and intentional about it and to look at a smaller number of enterprise level use cases. Sure, people want to generate PowerPoints faster or emails faster or slightly better spreadsheets or whatever. Great, but you're not going to get a whole lot of measurable value. One of the companies that I work with, DBS Bank in Singapore, very progressive with AI, and they say all that individual productivity stuff, we think of it as an employee satisfaction and retention tool, but that's about it.

25:18And the other thing that I was wondering too, I don't know if you have any predictions you want to make about this, but yeah, on that personal productivity thing, I have both tried to sometimes measure what I think it's doing for my own productivity. I fail miserably because I don't track how long it takes, how many emails I send out in a day. Nobody does that. But at the same time, if someone were to come along and be like, yeah, we're going to turn off your AI, I would have been like, absolutely not. Right. And so I wonder, my best guess is that we're just going to muddle along like that for the foreseeable future.

25:51Everyone has vibes based. I like it for my own personal productivity, AI usage. And we're going to continue to probably not have a lot of really strong, measurable numbers to back that up. And that's the world that we live in. But I do wonder if you would push back on that at all. Like that at a certain point, somebody is going to say we're spending how much for AI and nobody knows what it's measuring and start to ratchet that back or instrument our workflows in ways that we haven't done already to try to quantify some of that stuff. We're probably looking down the road a little bit with anything like this that might play out.

26:25But I just I do wonder if the mode that we're in right now is one that's sustainable in the long term. We could be done. It could be measured. we could set up semi-controlled experiments where this group uses it, this group doesn't use it, or one group uses it more aggressively, one uses it less, and one doesn't use it at all. And you measure how fast is the output, what's the quality of the output, which is always a challenging thing to measure. But that takes discipline and a culture of measurement. And I wrote an article, I don't know, a couple of years ago, a guy named John Sviokla, in Harvard Business Review about what kinds of capabilities it takes to really be successful with generative AI.

27:12And one was a sort of a culture of disciplined experimentation. So sometimes when I'm presenting to large audiences, I'll say, raise your hand if your company has a culture of disciplined experimentation and not a single person in the room, in the U.S. anyway, sometimes in Europe you get some, will raise their hands. And you mentioned workflows. We don't even know what our workflows are in most organizations and nobody owns them. We have owners of business functions, but we don't have owners of processes. And we don't even like the term process in many companies anymore because it sounds too bureaucratic.

27:47Workflows, for some reason, is still OK because, I don't know, it doesn't sound as big or bureaucratic in the process. But it's the same thing, really. And I think in particular, if we're going to use agentic AI to take over these workflows, we need to understand them much better. We need to have owners of them. We need to measure them, et cetera. Some parts of the world are better at that than the U.S. We're more, I don't know, we like heroic activity. We don't like disciplined activity. my last vein of questions you invited this a little bit earlier by making the mention of ethan malik and his thesis around just use the stuff there's no substitute for that for understanding what ai can do i always like to ask people about their own ai usage so my first question for you on that topic is have you finished your great american novel about the robot football team oh wow i forgot and if did you have an ai ghost right i've spent some time I spent some time in your sub stack, Tom.

28:48I spent some time in your sub stack for this. I did run it through a couple of different foundation models and asked what it thought. And for background real quick, because this is very insight, you had a book that you were writing, I think maybe during the pandemic, you picked it up. You've been like trying to write it for quite a while, wanted to see if AI could take a swing at this. And I think we're underwhelmed, shall we say, maybe by the results. I think that's the CliffsNotes. Yeah, this is a novel about robot American football. And I've written, I think, 28 business books, but I had not written a novel before.

29:24And when I showed it to people pre-generative AI, I finished it before 2022 or finished a draft. They all said, eh, this needs a lot of work. And I wasn't really inclined to put in that work. And so I thought maybe AI can help me put in the work. And so I fed it to both Claude and Chai GPT, and both of them agreed that it needed work. And it had some pretty good suggestions. Ironically, even though I finished it in, I think it was maybe even pre-pandemic 2020 or something, They said the technology hadn't really changed all that much in the robotic football space, but it needed all sorts of character development and so on.