Why AI isn’t showing up on your bottom line

0:00Azeem Azhar:Artificial intelligence is either a set of technologies that is really banal and ordinary staplers or photocopiers or even laser printers, or it is a general purpose technology, in which case it's going to have that kind of large scale systemic effect, not just within firms, but also across industries and within the economy at large. And the reason I set that up is because there's a big debate about where is the ROI and what should the ROI look like. If AI is a really trivial, ordinary, ordinary, non-general purpose technology, you would expect there to be ROI quite quickly. You just don't need to do that much.

0:41Azeem Azhar:If it is a general purpose technology, we would expect it to take a little bit of time to really start to show it's metal. I mean, that's what history tells us. It is the J-curve that Eric Brittany Olfson talks about. It's what Paul David famously described when he looked at electricity, and we drill into electricity quite a lot in our analysis. And the reason is that in order to absorb all the things that a general-purpose technology can do, you have to change a lot in your business.

1:21today i want to talk about ai and why companies are seeing the return on investments they are it is based on the essay and analysis that nathan and i put out early this week to try to make sense of what we are really seeing and i thought the story of electricity was really amazing and i

1:43Azeem Azhar:I think it's the best thing we have right now. The first companies to use electricity in the 1890s used it to illuminate the workplace, right? They just added more and more light bulbs to extend the working day. And famously, we talk about Serrano, who's an Italian coachmaker. He ends up making the first automobile for what became Fiat. And he used electricity very early on, actually 15 years before Henry Ford got started with light bulbs. And in many ways, where we are today with AI is in the light bulb stage. That is the co-pilot or the chatbot. But what actually happened in electricity was that the productivity emerged when people really rethought what happens when you think about a company in the age of electricity.

2:28So Henry Ford's business, Highland Park, that famous factory, wasn't more light bulbs. It wasn't a million light bulbs rather than 10. it was a complete rethinking of what it was to be a car manufacturer. It was no longer artisanal. It was no longer about stock. It was about flow. It was about standardization. It involved changing the supplier relationships. It involved changing the buyer dynamics. Lots had to change. It was not a million light bulbs. So when you think about AI and you're a company rolling out Microsoft co-pilot licenses, you're not going to really see the benefits of the general purpose technology by rolling out more licenses.

3:12Oh, instead of one license per person, it's going to be 10 licenses per person in the same way that you didn't get to Henry Ford with 10 light bulbs more per person. And I think that's a really hard thing to make sense of at this moment.

3:27Azeem Azhar:You know, Nathan and I spent a lot of time talking about the analogies. And one thing that didn't make it into our analysis was that I felt that this is a little bit like going from level two self-driving to level four self-driving, right? There is a complete phase change because going from something where human intervention is required to never being required is just a completely different way of conceiving of the system. And that I think is at the heart of the challenge that Main Street is going to face over the next three to five years. It's one thing to speed up a few processes with some co-pilots, mating transcripts, summaries of emails.

4:06The moment you start to get work groups effective within their groups, they run into this idea of congestion. You just produce more output than the next part of the process of the company can accommodate. I mean, imagine that you are a,

4:22Azeem Azhar:an equity analyst for sake of argument in a bank. And today, because of AI, you could now update your price targets much, much more frequently than ever before, right? You could have like a self-driving analytics network of agents taking in signals and sort of changing your expectations about where price could move very, very rapidly. But you'd still have to get that through, right? Your desk and compliance and publishing and everything else. So it doesn't really matter if you could make those adjustments. And it's not even clear that your customers on the other end, right, the sort of the buy side, the pension funds or whoever else is, could accommodate you providing real-time, minute-for-minute changes in your expectation of the price.

5:05Azeem Azhar:So there's this one tiny, tiny example, right, of how even if you get one part of the company working really quickly because of AI, and not just more quickly, but better potentially, the rest of it has to follow suit. So that switch, which we call moving from group drive to unit drive, we are just using William Devine's taxonomy when he looked at electrification of American industry at the turn of the 20th century, is going to be a really difficult one. A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.

5:45Azeem Azhar:What I think is happening right now is that the AI companies do understand that it's going to be hard. That's why they have all started to build up their professional services teams. They call them FDEs, Forward Deployed Engineers. I think what's behind that is that it is quite hard to go from the chatbot world into the workflow world. and it's even harder to go from that workflow world into the loop world that we describe in the essay. And that loop world is really about changing what the company cares about. It's not about changing its desire to make a good return on capital, its desire to be profitable and to increase those profits.

6:29But it's about saying that organizing principle is now about the decision making loop the autonomous sensing of the company and we argue and i feel quite kind of strongly

6:42Azeem Azhar:about this that you know if you just go out and do stage one and stage two and you get productivity and enhancements you drive down costs you create a temporary advantage because your competitors will do the same and that way is not a moat like a long-term durable position and in fact what would happen in that world is that company A will go off and do all its fancy workflows and the CEO will, in his quarterly earnings, talk about how they've increased the productivity of the sales force by X using AI. Well, within two quarters, their competitors will have done the same. And any advantage they've taken has been competed away.

7:24Azeem Azhar:And instead, what they're left with is an AI bill they didn't have before. So they may in fact be in a worse position than they were earlier. So that's not going to give you like a permanent advantage. What will give you that permanent advantage is anything that is a little bit more meta and dynamic to that, a bit second order, which is we're going to build an autonomous company, autonomous processes, which cycle through a loop where we've taken out the delay, which in exponential view, that delay that slows us down is known as azeem. Things just land on my desk and I don't answer my email. I don't even read my email.

8:00Azeem Azhar:And then Maria has to hunt me down to sort of get things moving again. But in every company, that's what happens, right? The decision-making slows down. So that's stage three. The company is orienting itself around the speed with which it can get through that loop. But that speed is not static, right? So it's not about, well, distance of the loop, circumference of the loop over time taken that stays the same. No, there's a derivative to that. It's the acceleration that you're looking for going through faster and faster and faster. And the reason is that you want to be adaptive to the external signals, to changes in consumer behavior, to changes in pricing, to a bottleneck in a key supplier and be able to adapt to that.

8:41So getting there requires a different mental frame for what CEOs and bosses need. and i think that's really difficult for large companies it's just really difficult frankly for

8:55Azeem Azhar:small companies to think like that for young companies you know it's not what you've hired for it's not what your internal standard operating procedures are for it's not why you've you've promoted people it is an entirely different way of of thinking now the idea of the loop of course exists in business agile lean um the ooda loop these things are all meant to uh encapsulate that But we're talking about turning this up to a different level. So if we come back to how this actually happens, I talked about the AI companies having forward deployed engineers and starting to think about professional services.

9:33Azeem Azhar:They're going to learn how to do this. But my bet is going to be those first engagements are going to be really, really difficult. They're going to be difficult because the forward-deployed engineer will not necessarily understand what it is to reconceptualize the whole of a business that they are being put into. So what then becomes interesting is these partnerships with professional services firms. And one of our readers is involved in this deployment company, which is OpenAI's JV with some private equity firms, to construct some sort of pool of capital and human talent to do these types of broader deployments as an indicator of just how difficult they will be.

10:17Azeem Azhar:Now, the obvious place to look for companies that can do this will be companies that are born after the chat GPT era. So you could look at maybe an Anthropic as an example of a company that is doing this well. If they are doing it well, and I confess that I know Anthropic a little bit, but I've not really been able to sort of go in and look under the hood. But if they are doing it well, it's certainly showing up in their revenue. The other companies that would be well positioned to operate like this will be companies where you have some degree of digital nativity, particularly from the CEO, who ideally may well also still be the founder.

10:55Azeem Azhar:and you have digital products, and you have digital touch points, and you have lots of data. So for that, you might be looking at, and we give an example in the pre-AI world of Sheen, Sheen, Shane, Sheen, anyway, that Chinese purveyor of massive amounts of pap and tat. But we also talk about Block, which is Jack Dorsey's company, which of course is a kind of fintechs and payments, so it's sort of well positioned, with the exception of any regulatory compliance overhead to do that quite well. I do think private equity firms will be looking at their portfolios and starting to think, which of my portfolio companies, so private equity, not venture capital.

11:38Azeem Azhar:So in private equity, you typically have that buyout where you have an operating business and private equity firm thinks they can do one of two things. They can change the sort of financial stack and juice a little bit more of a return out of it. Or they think they can come in with seasoned operators and make the thing more effective, more productive and start to grow sales. I mean, that's typically been the two approaches they've taken over the years. And I think right now, private equity companies, firms are looking at their portfolios. And these might be mid-market,$300 million,$500 million companies in the real world and starting to think this is a place we can apply that.

12:14The reason they might be thinking that is that you've already acquired the firm, you've already put in a management team, you own it, you're probably not as sympathetic towards aggressive change as the management team of a more traditional company that isn't owned by private equity. So you probably think that all of the pieces are in place to do that. And I've been puzzling over this because, of course, that's the obvious thing to think this is where this would happen. But it did occur to me that, you know, the nature of private equity is that you do a deal on a very, very tight set of hypotheses.

12:56I mean, the diligence these guys do is really, really incredible.

12:59Azeem Azhar:I mean, kind of blown away in awe of how good it is. And then the operating plan that comes out of it is really, really tight. And the game of private equity is not to swing for the fences. This is not a game of venture capitalists touting that they lost money on 99 out of 100 deals, but made 10 ,000 times their money on the one deal. Private equity is really about threading the eye of a needle time and time again. Eight portfolio deals, eight wins, maybe seven wins and one that returns its money. And so for everything in PE companies' portfolios today, those have been invested in a kind of pre-Chat GPT world, certainly a pre opus 4.5 or we're up to opus 4.8 today world and so any change you make to say well we can get to azim's and nathan's unit drive phase three uh you know autonomous tight looped company because we own it and we've put the management team in place is a diversion from the plan that you sold to your investment committee from the way in which you've always been successful or have been over the last 20 or 30 years.

14:03Azeem Azhar:So that makes me feel a little bit skeptical about whether that particular channel will yield lots of success in companies transforming. Where do I end up in all of this? I think that it's going to be much harder than people expect to get up to that kind of unit three level. I don't know what people's expectations were. Mine were always in the six to seven year range for a reasonable number of companies to be able to do that. They sort of still remain in that range. By the way, not six to seven years from today, six to seven years from 2023. So over the next three years. But I don't think that says much about how rapidly AI will be taken up in organizations because there's a hell of a lot that you can do in phase one and phase two.

14:48Azeem Azhar:And even if it doesn't have a jolt of TFP to the veins of the American economy, it's still companies buying a lot of AI services as they move up through those levels of maturity that we describe in the essay. So let's just sort of walk back through what I have sort of talked about, which is, you know, the real benefit comes when companies are able to absorb a general purpose technology. And that absorption is not a simple linear process from co-pilots through to wherever you you get to. It wasn't with electricity. It wasn't with earlier general purpose technologies. So whatever productivity benefits we start to see, I suspect will be a shadow of where we will get to over that six to seven year period.

15:43And I think sometimes people object.

15:46Azeem Azhar:Actually, this is an important point. I want to make, I forgot to make it earlier. Sometimes people make the observation that, well, it's because the technology isn't very mature. It's unreliable. It's hallucinating or confabulating. You give it the same process three times and it gives you three different answers. And it can be sycophantic and all the rest, right? So it's kind of a technical problem with the quality of the technology that is getting in the way. That's not right, right? That's not right at all in any meaningful way, because we can play the the Gedanken experiment, the thought experiment of what if this technology was perfect, whatever that means, in the terms that the person who's saying the reason this isn't working is because the technology is weak, right?

16:29Azeem Azhar:What if it didn't hallucinate? What if it could do reliably three hours worth of work at a time? And it was, you know, cognitive and its sort of abilities to handle a broad range of tasks, but still reliable. So just kind of imagine that, that is, you know, opus 5.6 or call it whatever you want to call it. Even if we had that technology, the transition from phase one to phase two to phase three that we described in that essay is really hard. It's still really hard to get from phase two to phase three. The thing that's blocking this is not the reliability or the efficiency of the technology. Let's face it, the electricity that Henry Ford was getting in 1908, 1989 is just not as reliable as the electricity that the Chinese industry gets today.

17:17The challenge is actually the reconceptualization of what it is a firm does.

17:23Azeem Azhar:It's not actually a transformation, right? A transformation is like, we've been using goldmine, ACT goldmine for CRM, and we're moving to Salesforce and that changes who can access CRM data. I mean, that's a transformation. We're talking here about a reconceptualization. So I would hold that this isn't a question of the technology not being good enough now, and that the unlock will somehow come when the technology is better and more reliable across a set of unspecified criteria. If you go and talk to Main Street, as I do very regularly, like dozens of execs every quarter, they're having problems with absorption rather than problems with reliability.

18:06I mean, the reliability is something to overcome, but ultimately, you know, then they're struggling to absorb the technology in the state that it's in. So, you know, with that said, the demand for phase one and phase two, as we call it, is clearly visible.

18:23Azeem Azhar:And we are going to see absolutely, you know, inordinate and staggering growth in revenues and for the AI companies, we've already seen it. And in fact, we were just doing a review of our forecasts with the team earlier this week. And we thought we were pretty bullish about 2026 in terms of how much companies were going to want to use AI. And we were really quite far off where it is. So the demand remains very, very significant, even if we haven't snuck out of the other end of the J-Cuff. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening.

19:06Azeem Azhar:That's all for now, and I'll catch you next time.