Welcome back to the London Futurist podcast. The most highly anticipated development in AI this year is probably the expected arrival of AI agents, also referred to as agentic AI. We're told that AI agents have the potential to reshape how individuals and organizations interact with technology. Our guest to help us explore this is Tom Davenport.
Distinguished Professor in Information Technology and Management at Babson College and a globally recognized thought leader in the areas of analytics, data science and artificial intelligence. Tom has written, co-authored, or edited about 20 books, including Competing on Analytics and the AI Advantage.
He's worked extensively with leading organizations and has a unique perspective on the transformative impact of AI across industries. He's recently co-authored an article in the MIT Sloan Management Review. Five Trends in AI and Data Science for 2025, which included a section on AI agents, which is why we've invited him to talk about the subject. So Tom, welcome to the London Futurist Podcast. Happy to be here, gentlemen. Thanks for having me.
It's a pleasure to meet you, Tom. Same here. Tom, before we get into AI agents, for the benefit of the tiny minority of listeners who are not already familiar with your work... Would you please introduce yourself? Give us a brief review of your career and work. I think you started as an academic, moved into consulting for some time, and then back into academia. Is that right? That is more or less true.
There were more iterations. I did start as an academic, a quantitative, computational sociologist. I got more and more interested in business. And after teaching for a while as a sociologist, I went to work for a consulting firm, but I kept going back to being an academic only this time. business schools rather than arts and sciences appointments. In many cases, including now, I did academic things and also worked with consulting firms.
So now I'm a senior advisor to Deloitte, which I've done for the last 15 years or so. Despite the job, I basically did the same thing, which is to write and speak and teach and so on. all about information technology and how it's used in organizations and how it affects people. Can you remember what it was that first got you interested in AI? In the 1980s, I worked for a consulting firm and we had a multi-company sponsored research program.
I did this with a guy named Mike Hammer, who is best known for his work in re-engineering. He'd been an MIT professor and we did a project on expert systems, prospects for early development, I think. And my consulting firm had created a little offshoot that did expert system work in the financial planning and what's now called wealth management areas. So I got fairly familiar with that.
I was working in Kendall Square in Cambridge and Symbolics and List Machine Inc. were both there. So AI was in the air in a big way, but then all of a sudden it was no longer in the air. We had this AI winter. And then I got interested in it again when I did analytics and somebody informed me one day that predictive analytics and machine learning were pretty much the same thing. And I said, wow, that can't be true. But then I confirmed that it was true.
So I moved from analytics to big data to AI in a fairly smooth transition since they're all pretty similar. So you've been both sides of the fence, consulting and academia, a lot of the time kind of mixing it up. And you've been interested and involved in AI for a very long time, much longer than most, certainly before the Big Bang in 2012. So you've got a long heritage in it. Let's start the exploration of AI agents with some definitions.
What is an AI agent and what distinguishes AI agents from other forms of AI? Well, as with most terms, God did not see fit to provide us with a definition of AI agents. So that means humans are free to define them however they want. But I think the consensus definition is probably that they are programs with some intelligence that can operate autonomously or semi-autonomously.
And typically, but not always, perform an individual task. And as a result, to accomplish anything, they have to be teamed up with other agents to really get much done. and sort of an ecosystem. Beyond that, I think it's pretty much up to the perspective of the individual. I also think it's important to say that these things are not entirely new.
Well, that's my question, because when I heard about AI agents, I thought, we've already got them, haven't we? We have bots running around social networks, which are often connected to language models, and they're able to come up with... quite plausible sounding conversations, often targeted to a particular goal. It might be to persuade people to vote for particular candidates, to hate particular candidates.
Or in some commercial cases, it might be persuading people to sign up for an account on OnlyFans. So what's different with the agents we're likely to see in 2025? I would say it's really a matter of degree rather than kind. There'll be many more of them. I was yesterday with a group trying to implement artificial general intelligence at Amazon, they reminded me that Echo slash Alexa actually was like an agent in that it did things.
It would turn on lights. It would turn on your heating if you wanted it to at certain times. It certainly would give you answers to questions like, what's the weather going to be today? So we've had these agents for a while, and then people talked about swarms of agents in this whole complexity movement a couple of decades ago now, accomplishing things with agents.
So I don't think there is anything different except that we have generative AI now, which will not be the only source of intelligence and agents, but I think it'll be an important one. And we've started to have the ability to not just know things, but also to do things with these agents. That point about Alexa is very interesting. I hadn't thought of that. And it's true.
Alexa does act as an agent. David and I, I think, have a bit of a disagreement about this because it does seem to me that AI agents with large language models are different. Currently, large language models really just provide information. They may provide a sounding board. You can have a conversation with them. You can fall in love with them. But it's all about information. They don't currently do stuff that affects the real world. And that's, I think, the difference with agents.
Alexa acting as an agent is really just acting as an on-off switch. It's not doing much more than that. What large language model agents will do is they'll seek out some information, combine that information, make some decisions. And I think that may be the key thing.
make some decisions which the human isn't involved in making, and then carry out an action on the basis of that, which of course carries the seeds of a great deal of concern. Will they get it right? And if they don't get it right, might they do something quite damaging?
So it does seem to me that there's a distinction in kind. I mean, as you say, God didn't give us a definition, so we're all free to make up our own. But I don't see LLMs at the moment making decisions and taking actions which will affect the real world. That's what I think will happen this year. Well, I also wrote something recently about analytical AI versus generative AI in Harvard Business Review.
with a co-author named Peter High. Analytical AI, what some people I think unfortunately refer to as kind of legacy AI or traditional AI or whatever. Good old-fashioned AI. Good old-fashioned AI. Analytical AI made decisions all the time. Should I issue a credit card to Callum? What offer should I send to David to most tempt him to buy something? What price should I charge? We've had AI making decisions, but you're right, it wasn't generally generative AI.
I do think that it's likely that we'll have both analytical and generative AI in agent ecosystems. Some of it will be relatively new. generative AI making decisions and taking actions, as well as just informing us. And some of it will be quite familiar. One example is that Marc Andresen has allegedly created an AI that was free to buy and sell shares, or at least crypto tokens, and seemingly made a significant amount of money out of it.
said to be the first AI crypto millionaire. That sounds like taking actions in the real world based on some kind of intelligence. That was already 2024. Well, yes, but we have to remember that people like The late Jim Simons of Renaissance Technologies made a lot more money than Marc Andreessen did by using AI to make decisions about investments, not cryptocurrency generally.
And again, not with generative AI, but with good old machine learning. While we're on the subject of definitions, you mentioned AGI back there. And since we are the London Futurist podcast, let's just do a little bit of diversion into that.
What's your definition of AGI and when do you think we might have one? I think that we're a long way away from it. If you define it as doing every cognitive task that humans can do only better... which I think is probably, again, the consensus definition of AGI. I think we're a ways away from it. And I agree with the maybe emerging view that it will take more than just generative AI to make that possible.
That would differ from people like Sam Altman in that regard. But there are a fair number of people saying generative AI systems don't really understand anything about the world. It's all statistical in nature. Correlation is not causality, so it will make mistakes and it will fail to seize upon what even young children eventually figure out fairly early in their development.
So you bring up the statistical analysis of being the root of how gen AI works. There's an interesting debate between people who say that generative AI is a really predictive text on steroids. Predictive text is when you analyse a corpus of text and you discover certain things about it. So, for instance, if you had the phrase, the cat sat on there, you've got a high probability that the next word is going to be Matt. And that's how predictive text is where you've had first.
over a decade in mobile phones and so on. That's how that works. My understanding is that Gen AI works in quite a different way. And in fact, Jeff Hinton in a recent video was very much a pain to say that large language models are not predictive text on stewards. They allocate numbers, I think they're called vectors, to tokens, and tokens are parts of words. And they also allocate numbers to the relationships between these tokens. For instance, the relationship between king and queen.
is very similar to the relationship between man and woman. The vector that goes from one to the other might be the same in both cases. The numbers that they allocate are very, very large and sometimes apparently multidimensional, whatever that means. To simplify it, the relationship between king and queen might be 3,000, the number 3,000. And then the same relationship might apply between men and women. Now that would suggest that to some degree they have a model of language.
They don't understand it in the way we do. They're not conscious. And there's a lot that we can do that they can't do. But to a degree, they have a language model. And perhaps that's where the word comes from, the large language models. And it's very, very different from predictive texting. Where do you fall in that debate?
Well, I would say that it is still predictive testing. It's just better predictive texting than what we've had in the past because it is informed by context. And I think that was one of the big breakthroughs of the Google researchers and the attention. is all you need paper in 2017, but turning things into vectors and embeddings and numbers instead of words and understanding the relationships among them.
It doesn't change the fact that we're still using deep learning models, incredibly complex and hugely trained models to predict the next thing that ought to come in a sequence of. words or proteins or amino acids or what have you. We still see every day that it can go awry at times because it doesn't really understand context very well. And I don't think that's going to change anytime soon. And I don't see how in a statistical model of the world that it can change.
Maybe we need to bring back some of the logic that was used in expert systems. Some people are suggesting that that technology didn't scale very well, so I'm not sure exactly how that's going to work. But I don't think we're going to get there with just statistics, even though... It's quite amazing what we've accomplished thus far with it. Talking about mistakes in AI systems, we've just had in the news that Apple's AI has had to have a feature withdrawn.
because it was summarising news stories and often getting the summaries wrong in a misleading way, so much that the BBC complained to Apple and saying, please don't do this. And Apple have said, it's true, we have to withdraw this feature. If even a company like Apple can't get this right, what hope is there for mere mortal companies? And they wait a long time, too, to issue Apple intelligence. They thought they had it all right. So if Apple are messing up...
What advice do you have for other companies? Is it stay well clear and wait for three years and then gently enter the field once things have been worked out? Or are there things that companies can do that are more sensible? In terms of how we deal with traditional generative AI, the informational generative AI, I think we certainly still need a human in the loop.
I was interviewing someone from London not too long ago. I think it's called A.N.O. Shearman. He was the head of AI at Allen & Overy, which merged with Shearman & Sterling. We've had him on the podcast. Oh, yeah, he's great. And he said, you know, if you don't have a human in the loop with generative AI, you're going to get bad law out. Even if it's only looking for hallucinations or confabulations or whatever you want to call them.
but also generative AI. Believe me, I've graded a lot of things with generative AI, get my students to do a lot of their work with it. It can be quite boring if it's not augmented by human capabilities. So there are a lot of things that humans can still do. Then when we get to agents, I think it becomes even more important because if we're asking these agents to take action rather than just give us some content, it becomes really quite scary.
We either need to be able to have a human review the transaction before it's done or be able to reverse it somehow if it's not appropriate and correct. I mean, once we start talking about doing investing or credit card transactions or planning a trip and not just planning it but actually executing on it, it's quite scary.
I don't want to overstate what large language models can do, and I don't want to give the impression that I think we're on the threshold of AGI. However, you could argue that to the degree that they are predictive texting machines, so are we, as I speak. I don't really know what I'm going to be saying 10 words from now. And my words are strung together. And I'm pretty sure that the eighth and ninth words are heavily influenced by the sixth and seventh word.
So I think there's quite a degree of predictive texting going on in human speech and human thought as well. And when I get GPT-402 to write a proposal, write a letter or whatever, it doesn't just do a sort of stream of consciousness, it plans. It goes away and finds information and comes back and synthesizes it and does some planning in the same way that I do. So I think we need to be careful not to dismiss too much what they do. And then also on the confabulation side or hallucinations.
Humans do the same thing. Human memories of the past are not stored in little pigeonholes and we send a little homunculus off to go and retrieve them and then unscroll them in front of our eyes. We actually make them up as we remember. So you think about a road you went down 15 years ago, which you have to remember for some reason, maybe you had an accident or something. You populate that road with buildings that you think might be there. That's how your memory works. It is made up.
which is exactly what the machines do. As I say, they're a long way short of where we are. And I think you're probably right that we won't just be able to scale up large language models to get to AGI. We're probably going to need to reintroduce symbolic AI.
People are calling that neurosymbolic AI, or maybe there's some other paradigm breakthroughs that are needed. But I'm always repeatedly impressed by what the large language models can do. And I think we need to be a bit careful about being too sneery. I'm not suggesting you were, but being too sneery about what they can achieve. Well, I agree with you. It's very impressive. And you're right, certainly, that humans are figuring out as we go along as well.
It's one of the reasons why if you're going to give an important speech, you write it down in advance and review it to make sure that as you were predicting the next word, you did a good job of it. We're all used to the idea that human biases affect our analytical AI models. But in many cases, we make better decisions when we have AI, even when it's trained on human examples in the past, just because there's more rigor to it, I'd say.
But the fact that mistakes are made in both cases means that we still need human oversight. And I think one of the issues with agents is the primary appeal for many organizations with agents is they think we won't need to involve humans as much. And I'm not sure yet whether that's true, at least for important business decisions and processes. Yeah. This reminds me of the old days when we used to argue, will people buy things online?
People thought you'll only buy things online if it's a very low cost. You wouldn't use it for anything serious like buying a car or anything like that. And then the question is, well, you might buy things on your PC, but you wouldn't buy things on a small screen mobile phone. But people get more and more comfortable with it and end up engaging in much more complicated, much more expensive transactions. I guess it may be the same with agents. At first, people will wonder.
What kind of tickets is this agent really buying for me? I asked it to find me a nice holiday. Where's it suggesting? Oh, that looks quite nice. Yeah, I'm glad it bought these tickets. But then there's the risk that it will, in one time in a hundred, make a really bad decision for you. And because you weren't paying attention, you end up going to Antarctica wearing your swimwear. Or worse. I think you're right. We humans pay...
more attention to machines making errors than we humans making the errors. One death in an autonomous vehicle gets far more attention than the many deaths that we humans cause in our vehicles. That may slow down the adoption a bit. But I did some interviews about agentic AI, along with a survey for a vendor, UiPath. And that's exactly what some of the business AI leaders said.
They'll start out with less critical types of transactions. It won't be used for insurance underwriting. It won't be used for large-scale business transactions. But maybe it helps you record your... vacation preferences with the HR department or something relatively simple like that. We can talk about adoption and how far we'll be happy to use these things. A lot of people saying that AI agents will become mainstream this year.
What does mainstream mean in this context? Is there an agreed definition of 25% of companies need to be using them routinely or 50% of consumers need to be using them routinely? And what do AI agents need to be able to do to be able to become mainstream? I think it's going to be hard to determine whether they're mainstream or not, because as we were saying, we had them before in some form, and there's not a good definition.
But over the last few days, I've seen some vendor plans for agentic AI. It's quite impressive what you'll be able to do. Large scale processes over time orchestrated by higher level agents, I guess you'd say, or maybe in some cases it's things like robotic process automation doing the orchestration and small agents. big agents and supervised agents and semi-supervised agents and fully autonomous agents. That's all going to be available this year.
The question is how rapidly can organizations make use of it effectively? And I think that comes down to, do you have carefully defined workflows leading up to large-scale business processes? Can you make a decision about what humans are going to do versus what the agents are going to do? How do you define escalation paths to humans if there's something really important that needs to be looked at?
And all that's going to take some time. So I don't see really large-scale adoption in 2025. I see large-scale availability and the beginning of... implementation in organizations. But with generative AI and with analytical AI, we saw a lot of experimentation before you started to get production deployments. We'll be right back after a quick break.
If you spend way too much time online, first of all, same. And second, you need promo guy talk pills. I'm Rawal, and with Alex, we break down everything happening in tech, digital marketing, and the internet's latest chaos. No boring corporate talk, just real discussions with real... So go on, press play. You know you want to.
Do you think there will be the kind of change that might cause great companies to stumble in the way that when there's a new technology, initially it doesn't seem commercially significant or practical, but then companies who do it well? Making the jump to this new technology can displace the previous greats, who are inadequately prepared for their new possibilities.
Like the main manufacturers of mini-computers were largely overtaken by the new companies that sprang up around their disruptive new product category of PCs. How were the mighty fallen? unlike the later transition from closed-sourced operating systems to open-source software, which also resulted in casualties. Is that the kind of transition possible with agents, or is that hyping it too much?
Some people seem to think so. Satya Nadella said it's going to replace most types of software. Jensen Wang at NVIDIA said it's the new form of software. I doubt that Microsoft is going to be turning off Excel anytime soon. And there's still a lot of people using Windows 95 in the world today. So there's quite a lag time, I think, before these things start to happen.
I think it would behoove any organization to start thinking about how to apply this. Don't do it in a way that risks your business, but certainly experiment with it and see if there's some processes you can incorporate it into. that would provide some benefits in terms of productivity and speed and so on. So are there scenarios in which the transition might happen more quickly?
because agents are combined in some clever way with some other process change or some other technology. After all, most of the breakthroughs in use of technology usually rely on two or three things coming together rather than just one. That's a good question. Some of these existing software vendors, like the robotic process automation vendors like UiPath,
are seeing that in combination with agents. I'm guessing that the SAPs of the world will do that for transactional systems. Salesforce is already doing that with agents for CRM systems. Will we have yet another combination? I don't know. It's pretty clear to me that there are going to be great marketplaces of agents at some point, and they'll be inter-organizational. That makes it more possible that you can accomplish great things.
But it also increases the risk, I think, that if we're needing five agents to perform a particular task and some of them go outside of your organization, you're not really in control of them. If something goes wrong, who did it and what do you do about it? It's going to be very complex. I think most large companies will be conservative about it. It may be that...
As you suggest, David, the startups are the ones that develop completely new ways of doing things. And sometimes those really take off and sometimes they don't. Do you think it's likely there'll be a market for validation services? So as a vendor, you put your agent into this environment and it finds out how the agent behaves in different circumstances. And you get a kind of certification. This agent is unlikely to go rogue. Do you think that's going to be an industry that will pop up?
There are already startups in that space, actually. And then some of the... Mainstream software vendors are saying that they will have some rankings of the agents that they make available to see how people have found them in day to day performance. Not only in terms of do they do what they're advertised to do, but how fast are they and how expensive are they in terms of the computational load that they have to bear. This reminds me of the adoption of smartphone apps.
because before apps could be widely adopted, there needed to be some verification system, some guarantees that the apps wouldn't do terrible things to the networks, which the network operators were worried about, or indeed do too much damage to individuals' phones. And layered on top of that was the user feedback system so that people who are thinking of downloading apps could look and say, oh, this has only got one star reviews. I'd better steer clear of that.
So I think this will be important, especially as you've suggested. The real innovation here might be in large networks or marketplaces of agents in which there might suddenly be more interesting.
services built by joining together smaller individually less capable agents yes i think you raise a very good point but who's the apple or the Google, in this case, that's going to do the equivalent of certifying that the apps perform well and don't harm your children and so on, and take 15% of the sales as a result.
It's not clear at all who those people are going to be. Maybe it'll turn out to be Apple and Google again, just because of their size. But I don't see anybody right now who's capable of doing that on a large scale. Should governments be getting involved? Well, not the U.S. government, that's for sure. And although I think the EU has been much more progressive and fast moving on AI legislation, it's hard to imagine that.
even in the great bureaucracy of Brussels, that they could set up an agent evaluation structure and process fast enough to really make that feasible. Yeah, I'm not at all sure this is a government regulatory activity. Probably either major vendors set up registration teams or, as you say, startups. And actually, I was asking a slightly leading question because a startup that I'm a co-founder of called Consum is developing.
agent validation as a service line. It's fairly new. Some of our people are at WPP, the large communications group. WPP works with many of the big firms around the world, and these firms are developing agents, and they need help in developing the agents and they need help in making sure that they don't do what they're not supposed to do. So our people are already kind of working on that within the WPP world. And I think it's going to be quite a big new industry.
I do think in that area of marketing and advertising, it's not generally tragic if somebody gets an ad that's not perfectly suited to their needs. We get thousands of those every day. So if agents are perfect to start with in advertising and marketing, it won't really be that much of a problem. If one person gets a highly inappropriate ad, maybe not. But if billions of people do, then you can trash a brand.
Pretty quickly. Sure. Yeah. As a lot of brands have found. So it can be quite important. Good point. I think things might go faster than we think. I completely agree with you that I am rather terrified of the idea of even GPT-4, which I like a lot. buying me an aircraft ticket because i think it will make me change in the middle of the night in dubai or somewhere which is a process i abhor and so i'm going to check it really carefully before i let it press the button however if people
do apply agents successfully in other areas which are less mission critical to the people involved. And it works over and over again. It works successfully. I think we might find that actually we move to trusting them. more quickly than we think. And as David mentioned online purchasing, I can remember when I thought I would never buy anything online because it just seemed too scary. It seemed too intangible, too unaccountable, too untransparent. And now I do it all the time.
And they turn fairly quickly. And it'll be interesting to see whether hackers or whatever you want to call computational evildoers start trying to inject agents into these ecosystems to say. I'd like to take a little cut along the way of whatever that transaction happens to be. I suspect the smart ones are planning it already. I'm sure they're already working on it. I don't like that they're not.
My last question, Tom, is imagine we are back in the beginning of 2026 and we're saying, gosh, we failed to spot a trend. We thought it was going to be AI agents, but in fact, something else was very significant in the world of AI. Have you any thoughts about what that other trend might be? I have been focused for the last couple of years on how do we get real productivity out of generative AI tools. And how do we get them into production deployment?
And at the beginning of last year, I did a survey with AWS and only, I think, 6% of companies had any production deployment. And now it's maybe up into the kind of 20% range on average. But that's still pretty low. Somebody estimated that in order for companies to recoup all the money that's being spent on NVIDIA stock and H100 chips and so on.
we'd have to save $650 billion through generative AI. So if there were a breakthrough in productivity from generative AI, other than agents, I think agents hold some possibility there. That, I think, would be something that's really useful to the world and would help us spread this technology more broadly. Okay, so agents just need to make between $600 and $700 billion a year.
to be worthwhile. There's no pressure, no pressure on the agency. Tom, thank you very much indeed for joining us on the London Futurist Podcast. My pleasure. Thanks for your thoughtful questions. Thank you very much.