type: issue
id: “component-00018”
display_name: “ProsaicTimes — 2026-11-08”
headline: “Will AI help us think more — or less?”
The PE fund CEO thought for a second and then asked “But where will the next set of wizards come from?”
We were discussing the impact of AI on knowledge work. I had observed that professionals would spend more time building ontologies and tending graphs than pushing boxes around a slide or copying numbers from one sheet to another.
The CEO reminisced that he developed his instincts about the financial mechanics of a deal by building hundreds of LBO models himself in the 1990s — so he could now sit in the investment committee and say “Wait a second. Something seems off with your capex projections in the out years. Explain that to me.”
I reminded him that not too long before that bankers built models with graph paper and calculators rather than spreadsheet software. He countered that moving from spreadsheets to GenAI-built models represented a bigger disruption than moving from graph paper to desktop software. He had a point, and it is stronger than my analogy allowed. Spreadsheets removed the arithmetic but left the model building, and the model building is where his instincts formed. AI may eliminate that work. This is a genuine difference from the move from graph paper to spreadsheets.
I said that we’ve been wrestling with the cognitive impact of new information technologies for millennia. The Iliad and the Odyssey pre-dated the re-introduction of written communication to the Greek world by centuries. You imagine crusty bards complaining the youngsters would never really learn Homer if they could just refer to a passage in a scroll.
The CEO liked the example. I suspect we were both right. AI will eliminate some of the work through which professionals traditionally developed expertise. Whether it creates better ways to develop that expertise remains to be seen. Each individual will need to decide whether he or she will use AI to think more or think less. Each institution will need to decide whether it will use AI to produce more work or to develop more capable professionals. The first path accepts AI slop. The second builds the tools and incentives that enable and encourage employees to achieve things not before possible, and it is the more interesting strategic choice. It is also where the next set of wizards will come from.
My colleague Ryan Davies and I discuss this topic and many others in this week’s Prosaic Times podcast.
Gen X, Vic 20s and Pac-Man
In this section:
Early computing memories: a Vic 20 word processor, Pac-Man ghosts evolved by genetic algorithms.
Twenty-one years at McKinsey on top of an applied math and computer science background.
The conversation is set up around finance and the economics of technology.
James Kaplan: Hello, it’s James Kaplan with Ryan Davies and another Prosaic Times podcast. Welcome, Ryan.
Ryan Davies: Thank you, James. I am so excited to be here with you.
James: And I’m excited we surmounted our minor audiovisual difficulties in getting this set up.
Ryan: In the age of AI, sometimes it’s the little stuff that trips us up.
James: Exactly so. So Ryan, we start out with why don’t you give people just a brief introduction or a brief background?
Ryan: Okay, my name’s Ryan. I’ve been 21 years now at McKinsey. Before McKinsey, I studied applied math and computer science. And probably the closest I’ve ever been to the frontier of AI was in high school. I grew up in New Mexico, and there was something called the supercomputer challenge, where kids got to code something for the supercomputers at Los Alamos to run. And my sophomore year, we used genetic algorithms to evolve Pac-Man ghosts, and we competed against some more deterministic heuristics that we also coded and the genetic algorithms just beat all the other ones and we won. And that was a really awesome experience and again probably the closest I’ve ever been to the AI frontier in the mid to late 90s, quite a while ago now.
James: That is terrific. Well my first computing experience was programming a Commodore Vic 20, which is not so glamorous as a supercomputer.
Ryan: Yeah, but you’re very credible.
Old school cred.
James: That’s a function of being Gen X. I will say I was talking with a senior executive and we realized we were the same age and we’re talking how we first started with technology. I mentioned I wrote a word processor in nineteen eighty-two on my VIC twenty so I wouldn’t get bad grades in my papers for my lousy handwriting.
Ryan: Ha ha.
James: And this guy ran outside into the file cabinet outside of his office and came back in with the Vic twenty he used or the Commodore 64 he used as a teenager back in the nineteen eighties. So that is how emotional an attachment to those of us who are Gen X can have to our early computers. I do not still have either of the physical Commodore machines.
Ryan: Excellent. Yeah.
If I interpret Twitter correctly, I am a quote geriatric millennial.
James: Our mutual colleague Charlie Lewis describes himself as a Xennial, I believe. We Gen Xers are always skeptical because we don’t think the Xennials or the geriatric millennials are sardonic enough to hang with Gen X.
Ryan: Mm. I am very earnest.
Tokenomics and who burns tokens
In this section:
In the three-to-five-year window, building AI may cost more than running it.
An enthusiastic engineer with an unlimited budget can consume far more tokens than anyone has modeled.
Mapping a code base into a graph spares the tokens spent searching it again every turn.
James: So one of the reasons I’d asked you to come on was to I know you do a lot of work in corporate finance and spend a lot of time with CFOs, so we should talk a little bit about that and you know how to think about it’s a particular passion of mine to think about technology w from a rigorously economic standpoint. So we’d love to hear a little bit about your perspective on that and we can range more broadly because I know there’s a whole bunch of things related to AI and technology far beyond the CFO function that I know you’re all also interested in. So tell us a little bit about how you think about that.
Ryan: Yeah, happy to. And I think by the way, personally very interested in AI and longtime personal and professional interest in capital markets and finance. There’s many ways in which those two come together, all of which are really interesting. One of the pet projects that I haven’t really gotten off the ground, but I think there’s actually a lot more research we could potentially do in a bunch of different areas of this. We also, by the way, have a fantastic team digging into tokenomics and exactly this question. So Tanguy and Wasim.
And a bunch of other folks. We have a tech ROI effort that has done some good work on this. And frankly, we’re still much nearer the beginning than the end, I think, of having settled all the theory. And then there’s more work to get from great theory to great practice.
James: Yeah, I admit to being a bit of a tokenomic skeptic.
Ryan: Say more about what you mean by that.
James: Okay, so there’s a couple of ways companies might consume tokens. They can consume it via chat interface, but that’s not that many tokens. Very few people can type that quickly. They can consume it via applications that engineers have built. they can consume it via the process of building those applications, and they can consume it via citizen development. Now, in the three to five year time frame, if you think of those middle two things the number of software engineers you have is a natural barrier on the number of tokens you can consume. So I’m hypothesizing, I haven’t modeled this out entirely, that in the short term, the cost of building AI for many enterprises will exceed the cost of running AI.
And finally the big unknown is to what extent companies will succeed with citizen development. Then you widen the number of people who are both using coding tools and therefore consuming tokens and building things that might consume tokens. Does that make sense at all?
Ryan: I get the framework, and the driver in the driver tree that my instinct is to pressure test first is the maximum tokens that a software engineer could consume. Because with cursor and
James: That is the X factor, I agree.
Ryan: cursor and Claude Code and more and more loops and layers of abstraction, it wouldn’t surprise me if we have orders of magnitude to go before. And I do think there’s a ceiling on just cognitive capacity.
And I think there’s less of a ceiling than there should be right now on the quality of the code and software that’s being developed. And hopefully for the sake of everyone, that starts to bind a little more as well. But I think an enthusiastic software engineer with an unlimited budget can probably spend a lot of money with access to the right models and harnesses.
And I do think there’s a ceiling on just cognitive capacity. And I think there’s less of a ceiling than there should be right now on the quality of the code and software that’s being developed.
James: Yeah, and just as a note, I found out of the forty four thousand people in McKinsey last month I hit the top ten. Not the top ten percent, but the top ten in terms of token consumption.
Ryan: Good for you.
I thought that was mostly BAs in the top ten, so you’ve really earned some cred to get up there.
James: Well that’s a thing. Even though I’m old, I try and hang with the young.
Ryan: Yeah. I don’t have to watch my McKinsey token consumption that much because I’m not so much of a big user, but I have a personal, Claude Pro or Claude Max or whatever subscription. And so it is interesting to me when I’m, vibe coding something, a personal agent or I was recently using Fable for some personal legal advice, which may or may not be advisable. That’s not legal advice or life advice, but I wasn’t getting into big numbers.
James: Yeah, it’s interesting to me, and I think there’s gonna be a lot of discussion about architectural choices you make that will impact token consumption. For example, one thing I did with one of my repos is I mapped my entire code base into a graph. So each function is a node,
Ryan: Mm-hmm. Cool.
James: each function call is an edge.
And I think I haven’t instrumented this yet, but from what I read, that mitigates token consumption. Because you’re not doing this investigation of the code base with every turn.
Ryan: You’re not like searching every time.
Deterministic code, open weights and JEV
In this section:
Agentic software engineering that produces deterministic code can draw on thirty years of quality assurance. Agentic software at runtime cannot.
Open-weight models may be good enough for most deployed applications, but frontier-lab optimization could erase the cost edge.
Turning free text into structured data is the killer use case for a cheap classifier like JEV.
James: Exactly, exactly. Now that said it is hilarious to think that it’s sometimes hard to get your code base to act the way you want. You watch and you see it start grepping half the world, you’re like, no, don’t grep. Walk the graph, walk the edges. But the other thing I’m thinking about a little bit is where do we want to use agentic applications where you’re in effect consuming tokens at runtime. And when do we want to use agentic software engineering to build deterministic applications? Which has two advantages. One of which is if you’re consuming tokens once, that’s more token efficient. And then perhaps more importantly, once you write the code, we can use thirty years of software quality assurance practice to make sure we’re comfortable that the code does what we want it to do. Quality assurance for a non deterministic application is a little bit hairier a thing.
Ryan: Yeah. I totally agree with that point. I think one, I agree, there’s a lot of the idea of building deterministic or largely deterministic software with very very contained, non-deterministic, LLM calling pieces is the way to think about things. That’s putting aside, for example, I expect we’ll find, again, orders of magnitude efficiencies, one through using open weight models where useful. By the way, one of my pet theses that is difficult for me to fully express financially is that secure end-to-end American sovereign provision of great open weight models is gonna be a massive business. I also think sorry, you say something on that.
James: Yeah, well let’s strike that department our colleague Sven Blumberg, who I have a lot of regard for, his hypothesis, and he’s still getting his head around this a little bit also, is that the Frontier Labs are doing some massive optimization that we don’t see. So it’s not clear to him that open weight, when you consider the TCO, is less expensive than a proprietary model. [1]
Now, I haven’t seen the analysis on that. I don’t think he’s done it yet, but it’s an interesting hypothesis for me to examine. Because originally even a few weeks ago I was hypothesizing, yeah, I was thinking, gee, we all have this massive, nobody’s using their laptop capacity. Why aren’t we putting local open weight models on every software engineer’s laptops and just sopping up spare capacity.
Ryan: Yeah. It’s an interesting hypothesis. It’s counterintuitive to me, which doesn’t mean that it’s wrong.
My intuition is that there’s, for a significant majority of tasks, actually, when it comes to inferencing in an established application, a great open weight model will be good enough. By the way, I’m not even going to pretend that I fully understand JEV, but there’s probably some other things around routing to different models and model types to further create efficiency.
James: JEV is fascinating to me.
Ryan: And so I want to hear your thoughts on JEV, but just the closing thoughts.
It feels to me like, for the vast majority of deployed applications and deployed agents, the number of times you’re calling Fable or Astra or Argo or something like that in the running of that application should be very small or zero.
James: Yeah, I think that’s a really good point. That the frontier models matter more in software engineering than software execution. The other thing, I haven’t had a chance to play with JEV yet, but here’s why I’m so intrigued. One of the killer use cases to me for Gen AI is the conversion of free text to structured data.
And I would argue, for example, since the beginning of time, McKinsey associates have created scads of value by sitting down with hundreds of pages of free text and coding that in such a way that they could turn into structured data and we could do some sort of analysis with it. Something so simple as was an insurance claim paid in adherence to policy or not?
Right. And then we come back and say, hey, we sat down with a sample of a thousand insurance claims, and guess what? Twenty of them are paid inappropriately. You in your career, I’m sure, have seen a thousand analyses like that. Here’s a hundred transactions that went sideways, here’s a thousand transactions that where you had some sort of breakage and here’s why they went sideways. And so the ability to use something like JEV to quickly and cheaply act as a classifier to me is incredibly exciting.
Ryan: Yeah, I have the same intuition without any of the technical understanding to underpin it properly.
James: Well remember you have a background in computer science. You’re well ahead of many people. And you were building genetic algorithms in the nineteen nineties.
Ryan: I was at MIT, our partner university two weeks ago, and one of the things that I took comfort from was that the MIT professors were excellent, by the way, really, really good. And it gave me some comfort that even they, who are able to spend much more time than I do trying to keep up with the frontier, struggle to keep up with the frontier.
Leaders cannot delegate AI fluency
In this section:
Fluency comes only from building with the tools.
Senior technology executives who gave up the keyboard can build again.
A family-calendar agent shows the payoff comes as much from clearing bottlenecks as from writing code.
James: Yeah, the only way and it sounds like you’re doing one of the most important things, unless you’re spending real time with one of the coding agents, I don’t even know how you stay conversant.
Ryan: I don’t think it’s possible. One of the points that one of the MIT professors made, which completely resonates with me, is that AI leaders cannot delegate AI fluency today if you want to be an effective leader in domains for which AI is relevant, which arguably is becoming pretty much all of the domains that exist. And the only way to develop fluency is to be messing around with this stuff. And so yeah, I’ve found I try to do a lot of different things with it professionally. Frankly, the coding I’m doing almost entirely it’s for my own, like on my own personal account. So I made a home agent called Claudine who helps maintain our family calendar. And I set up, Claudine’s got a WhatsApp account and it’s set up on Twilio and it’s running on Railway and we can text Claudine if we wanna update, the calendar or something like that. By the way, the LLM for that, I think I built it with Opus 4.8, extremely helpful not only for the code generation, but for the debottlenecking: hey, the Twilio verification is hung up somewhere, what should I do? It would have taken me five times as long.
One of the points that one of the MIT professors made, which completely resonates with me, is that AI leaders cannot delegate AI fluency today if you want to be an effective leader in domains for which AI is relevant, which arguably is becoming pretty much all of the domains that exist. And the only way to develop fluency is to be messing around with this stuff.
James: You never would have gotten there. You don’t say it would have taken you five times as long, you would have given up.
Ryan: It would have taken me five times long to write the code, but I would have stopped on some of these administrative bottlenecks that the AI was very helpful to solve.
James: Yeah, this is an interesting dynamic I sometimes see with senior technology executive CIOs, CTOs, and what have you, which is folks with these roles in big companies, they gave up the keyboard a long time ago. Partially because if you’re the CTO of a big bank, there’s hundreds of different technologies. You can’t stay current on all of them.
Right, and you’re necessarily managing via a level of abstraction. Things are changing very quickly. And I can’t tell you the number of senior technology executives said, well gee, now I can build again. Because it allows them to move more quickly and it allows them to debug things and figure things out even if they haven’t been reading sort of staying current on all the forums and newsgroups and what have you so not only, I would suggest, can senior leaders, to your point, not outsource AI fluency, but now have much more capability to be fluent in AI than would have been the case or anything else even a couple of years ago. Does that make sense?
Ryan: Yeah. I agree with that. And I think there are some prominent counterexamples. It’s like Tobi Lütke from Shopify is posting all the time about all this stuff he’s building, and it’s awesome, and I think highly, highly credentializing. And my encouragement to all the folks out there with computer science backgrounds, or even not, is, build with this stuff, one, because it’s gonna make you better, and two, it’s outrageously fun. It is outrageously fun to bring things to life in this way.
James: It’s the world’s best Lego set. Let me test the hypothesis.
Ryan: Yeah. So just the next frontier for Claudine, my wife spends real time every Sunday night transcribing from a list of recipes for a week to the grocery list for the week. And that’s
James: Yeah. Oh God.
Ryan: obviously something AI could do quickly. So when I find the chunk of time, and it’s not gonna be this weekend because I’m away at an AI conference, but it’s when I find a chunk of time, that’s my next thing to build is the use case for Claudine to do that grocery list transcription.
James: Yeah, I have to admit, I’m fanatical I’ve put everything about my professional life into a knowledge graph which I built myself. Our home life is still very pad and paper. And there’s some degree of which it’s I enjoy the paper tangibility of it. Also with the kids in college things are much simpler than they were ten years ago.
Ryan: Yeah, my extracurricular and driving schedule on the home front sounds like it’s probably more complicated than yours is at the moment.
Will knowledge workers become engineers?
In this section:
Knowledge work may look like DevOps and capital-markets strats: people who build the tools and know the domain.
Some people will use AI to think more and some to think less, and the second group loses.
Cognitive time under tension and refusing to delegate understanding keep the mind sharp.
James: Yeah, yeah, now it’s like, oh gee, yeah, Thanksgiving’s coming. They’ll be home.. It’s much less logistically challenging than had been the case in the past. Let me test a hypothesis with you. I wonder if knowledge workers, business analysts, finance analysts, what have you, are going to become more like software engineers.
And the two metaphors I sometimes use are it’s like DevOps for knowledge work. It used to be when people provisioned a server, they would sit down and do the configuration directly. Now they’ll write a script, or use a script to do the configuration. The other is like strats in capital markets where you the people who are both very good software engineers and also understand capital markets or understand various trading strategies and what have you. And so do you think we’re at the edge of a transformation in knowledge work and what knowledge work looks like and feels like. Taking finance as an example.
Ryan: Yeah. It’s a really interesting question.
It brings me back, by the way, to when I was interviewing out of undergrad for McKinsey. I was also interviewing with Goldman Sachs. And to Goldman’s credit, they spent a lot of time trying to figure out what role to put me in. And they ended up giving me a couple of offers to be what’s called a sales strat, which is I guess what you are at Goldman if you can code and do math at a high level, but also are good at talking to clients or something like that.
James: Yeah. The forward deployed engineer is a new name, it’s not a new thing.
Ryan: Yeah, maybe sales strats were the forward deployed engineers of the Goldman FICC division of the day. I like what you described. And I think the best knowledge workers probably will go in that direction. They’re capable of building, deploying, tweaking these tools like at the edge, at the coal face, if you will. On the other hand, I’m not sure everyone’s gonna do the work — the personal work — to develop the capability to get there. And there was an article I read, I can’t remember where, but my synthesis of the article is that there are some people who will use AI to think more, and some people who will use AI to think less. And people using AI to
James: I think that’s exactly right.
Ryan: think less is a problem and will not work out well for them professionally.
James: I think that is precisely correct. Because you do have people say, gee, AI can write the executive summary for me. No, no, no, you write the executive summary. AI helps you structure the appendix. Or AI helps you gather the information.
Ryan: Yeah. One other concept I liked a lot, there was a podcast, Derek Thompson interviewed a guy named David Deming, who’s now the dean of Harvard College. Interestingly, Ohio State undergrad has become dean of Harvard College, which is impressive. Good for him.
And there were a couple things I really liked about the interview, but one of the concepts that really stuck with me, which was actually from Derek, was this idea of cognitive time under tension. In the same way that physically we build muscle by time under tension, cognitively we build cognitive ability by thinking and thinking hard and pushing our brains. And so I think that concept is really useful because if we’re, using AI to explain.
And think and pressure test, which is certainly how I feel. I think it’s gonna, our brains will stay strong and we’ll be magnified by it in a way that’s really impactful. If we use it to offload too much, we risk, cognitive atrophy. And again, I don’t think that ends well for people who go down that path.
James: I think that’s a really interesting point. And one of the dirty little secrets of knowledge work, whether it’s software engineering or anybody anything else, there has traditionally been a lot of mindless scut work where you could turn your brain off. It’s like copying and pasting and I always hated back in the 90s sort of coding UIs I always felt was sort of brain dead but necessary, or what have you. I’m sure many will disagree with me. And it’s interesting to me the degree to which by abstracting away some of the scut work.
Working with AI forces you to be on your game for a higher percentage of the day.
Ryan: Yeah, you hear stories of burnout in software engineers, for example, where it’s like, they’re doing so much more. But again, there’s real cognitive load to it, and our energy, our physical energy and cognitive energy is finite. So yeah, and I certainly relate to that. I feel like I go to bed tired as often as I ever had. And it reminds me, our colleague and CFO Yuval Atsmon published something in his newsletter on LinkedIn a week or a couple weeks ago and the title was Don’t Delegate Understanding. And it was a really thoughtful piece about you know how to use AI and not use AI. And I think this idea of by all means use it, but don’t delegate understanding. By the way, it’s a good policy in knowledge work in general and has been for long before Gen AI was a thing. But it’s a really good framing and it takes real effort and energy to do it well all the time.
And I think this idea of by all means use it, but don’t delegate understanding. By the way, it’s a good policy in knowledge work in general and has been for long before Gen AI was a thing.
James: Yeah, one thing I want to do some research on is the distinction between, system one, intuitive versus intentional system two, intentionally analytical thinking when you’re using AI. Because I often find myself using a lot of system one thinking, a lot of instinct, which is not that different from what you do when you’re working with early tenure associates, you develop some instinctive pattern recognition about the stuff they tend to get wrong in the spreadsheet. And you don’t quite know exactly why you’re saying, go tell me about this cell. But you do, and it turns out that that’s something that’s messy. And I find myself using a lot of that, which does also require a lot of concentration.
Ryan: Yeah. I had a system two example last night where we had a we’re preparing for an important workshop for a client next week. We have a draft deck, we have a draft draft agenda and all that kind of stuff. And the client helpfully like wrote us what they had in mind for the agenda.
Same spirit, but different in some ways.
And we spent ten or fifteen minutes on the team problem solving talking about it. And then I said, you know what? I actually just need some quiet time with this. Let’s end the problem solving. I need to sit and really think about how we incorporate the other input while staying true to what we think is really important in this session and make that one fused thing. And that was an offline old school manual. System 2.
Redlines, contracts and legal work
In this section:
An automated redline between a client email and a revised draft catches what a tired reviewer misses.
Comparing contracts conceptually, not just word for word, makes legal work the canary for knowledge work.
A formal contract language would expose vendor terms and open second-order economic effects.
James: Yeah.
I love that example. The one thing I’ll add is one of the things I love about incorporating Gen AI into a workflow is the ability to take two things and sort of compare them. For example, there was a document that I sent over to a client. Client said, gee, this is great. I circulated with the people. Here’s these 12 pieces of feedback. Could you incorporate this? Had a associate partner update it. And I looked through it and like, yeah, this looks pretty good. But the ability to do a virtual red line or an automated red line between the email I had gotten from the client and the updated draft to do the punch list of whether everything had been addressed.
Yeah, I would have done that by hand, but I’m sure that the machine would have been more conscientious than I would have been by myself.
Ryan: Yeah. I had a funny personal example of that recently. Does Chrome want to relaunch now? No. Dismiss.
James: Chrome might want you, but you don’t want Chrome to.
Ryan: Yes. This is in the context of a prospective real estate transaction in my personal capacity, but we’re going back, there’s some stuff that’s a little complicated. So we had drafted a contract and the other party had had their lawyers draft a different contract sort of from scratch from a different template. And I was using Fable on this and its ability to almost instantly pick apart the comparison, not just at a word for word level, but a conceptual level was amazing. And for example, it identified something about the template that the counterparty was using that was actually favorable to us, not them, which I wouldn’t have caught, maybe, even if I had stared at it for 24 hours, but it just picked it up like that. By the way, I think legal work is gonna be a fascinating canary in the coal mine for knowledge work in general because these models are amazing at some of that stuff
James: Well, and a contract is a form of code. Although I will admit, it astonishes me that nobody came up fifty years ago with like formal language for certain types of legal contracts, but I suppose what’s that?
Ryan: Like lean for legal contracts? That’s a great idea.
James: Yeah, XML, XML, yeah, exactly. Yeah. We have formal language for mortgage contracts. That’s why banks can trade mortgages because there’s a standardized, parameterized way of describing mortgages. Now I’m sure there’s a were a million in stakeholders that did not want formal language for various types of contracts, but yeah. One of the second order things that I think nobody’s thought through you think about it, contracting is incredibly intransparent. Most companies, even if they sign a contract, it goes in a virtual file cabinet, they never look at it ever again. Because it’s just not worth it.
But now you can interrogate your contracts with your ten thousand vendors to figure out what the discrepancies in payment terms are or whether you’re fully taking advantage of the payment terms you have access to. There’s going to be, I think, some interesting second order microeconomic effects to the fact that we will now have much more transparency into and transparency over time.
Does that make sense?
Ryan: Yeah.
It’s super interesting. From the math angle, it’s been really interesting what folks have done in recent years with Lean and Axiom Math and Frontier Labs are getting in the mix. I hadn’t thought about this idea of an analogous, truly formalized language for the law, but it’s a really appealing idea. And for any entrepreneurial-minded folks out there, I’m not convinced
James: Except to lawyers.
Ryan: that Harvey and Legora are gonna, survive in the end of day if the core intelligence is coming from the frontier models, but maybe there’s a different legal entrepreneurial angle there.
CFOs and the ROI hand wave
In this section:
CFO fluency on AI varies widely, and perhaps half of CFOs admit they can’t see the return.
Tesla, SpaceX and Amazon show that option value, not current operations, can justify an investment.
Treating an AI case like a venture bet may be more honest than scrubbing it like a tactical purchase.
James: Exactly. I wanna talk to you about the AI conference you’re going to in a couple of minutes, but before we do that, where are most CFOs on AI investments, on how they’re thinking about how to measure value from AI and, you know what? Less about how AI is gonna change the finance function, because everyone knows that’ll get more automated, but how will it change the economics of large institutions? Does that make any sense?
Ryan: It’s a great question. A couple things. First of all, I recorded a little video snippet in February where I said that I thought ROI on AI is gonna be the most important question in business in twenty six, and I feel pretty good about that standing the test of time. Second is you see in CFOs like other executives.
A very wide, perhaps shockingly wide dispersion of AI fluency. And from people who are doing impressive things to people who, and I certainly won’t name you, feel like it’s like, have you been living under a rock for the last three years? Like, I don’t want to be dispersed. Yeah.
James: There’s a lot of rocks in corporate America for people to live under.
Ryan: Third an experience that speaks to this that really stuck with me. So we have a global CFO forum every year in London in June. It’s a fantastic event. Everyone out there should nominate your CFOs. It’s really, really world-class only McKinsey event. One of the things that surprised me most there, because this question is top of mind for everyone, how are you thinking about ROI on AI?
I was expecting all of the CFOs, or substantially all of the CFOs, to be really hard edged, yeah, we’re looking really hard at this and really scrubbing the numbers and sweating the business cases and all this kind of stuff.
And what I heard instead was actually, maybe half the CFOs were in that camp, but many, many CFOs, maybe half, were much more in the hand wave. We actually don’t know exactly where the return is gonna come from. We kind of know it’s the right thing. We had the CFO of an almost hundred billion dollar enterprise, super rigorous guy, former banker, getting interviewed on stage, and he was like.
Yeah, I don’t actually have line of sight to the ROI. I just intuitively, know it’s the right thing to do. And that that surprised me. But it’s interesting. I don’t actually know that it’s wrong. Yeah. More honest and not necessarily
James: Is it possible that’s the more honest answer? Is it possible that given
Ryan: wrong.
James: the degree of uncertainty that trying to scrub the AI business case the way you would scrub the business case for sort of a more tactical technology investment. Maybe that’s the wrong thing. Maybe you have to think about this a little bit like a venture capitalist.
Ryan: Yeah, a capital markets analogy that comes to my mind is if you try and explain, I’ll pick an extreme example or two. If you try and explain the value of Tesla or the value of SpaceX through the current operating businesses, you just can’t get there. Those companies are valued the way they are because they have these real businesses. And Elon might do other awesome stuff on that platform.
If you try and explain the value of Tesla or the value of SpaceX through the current operating businesses, you just can’t get there. Those companies are valued the way they are because they have these real businesses. And Elon might do other awesome stuff on that platform.
James: When, in one institution, you have, what, seventy percent of the world’s space launch. I’m sure it’ll be in the percentage.
Ryan: But the thing is you can’t get to one and a half, two trillion dollar just on the space launch business. You just can’t get there. And so yeah,
James: Of course, I understand that, but I’m saying who the hell knows what might happen.
Ryan: yeah, but so the point is a financially rigorous take on it could be there’s some near-term ROI, but there’s, the value of us being at the cutting edge of this will create new frontier opportunities that we don’t even understand yet, but we have conviction it’s important for us to be on that frontier. I frankly feel that way about us as a firm as well.
James: Yeah. think about it from an Amazon standpoint, from my understanding outside in, a large percentage of Amazon’s cash comes from AWS, another big chunk comes from advertising. In 2003, who would have predicted AWS or Amazon’s advertising business. Nobody had foresight into that. They leveraged the platform, they used the option value of the platform and leveraged it in different ways.
Ryan: Yep, the platform, the talent, the leadership.
The Curve, P doom and growth
In this section:
The Curve gathers a multidisciplinary group around positive outcomes from transformational AI.
P doom tracks geopolitics, and the larger worry is P degradation: thinking less, and sycophantic models.
Growth could add a hundred basis points a year, and recursive self-improvement is doubtful.
James: Exactly. so tell me about this conference you’re going to. That sounds super interesting.
Ryan: Yeah, I’m very excited and intrigued. The conference is called The Curve.
It’s in Berkeley, California this weekend. [2] It’s put on by a nonprofit called the Golden Gate Institute. And the intent of the conference is to bring together a diverse multidisciplinary group of thinkers to discuss how we can, as humanity, achieve positive outcomes from transformational AI. And this is the hosts have conviction that no discipline has all of the answers, and so it’s important to bring people from across many different disciplines to discuss, which is a mindset that I like. A good friend of mine who runs a nonprofit related to career skilling called Career Village went last year and said it was amazing. He highly, highly recommended. He nominated me to apply, which I appreciated, and the organizers felt like I would be a net addition to the conference which I also appreciate and so I’m gonna go and there’s a bunch of formal sessions and a bunch of just mingling around with different people and there’s some there are some reasonably big names there and I’m sure a bunch of other interesting people who I haven’t heard of but look forward to meeting.
James: Terrific. We’re coming to the top of the hour. Anything else we should talk about? I feel like we go for another hour.
Ryan: Yeah. Let me think of some hard questions for you that you can play back to me. Let’s start at the foundation. One, what’s your P doom and why? Two, what do you think AI does to economic growth and why? And three, how real do you think this sort of RSI recursive self-improvement is gonna play out to be in the coming years?
James: I think P Doom is very low.
P Doom to me is correlated entirely with your beliefs around geopolitics. Which is if you have major nation states using AI to attack each other, that gets really scary. If you don’t that feels a lot better to me. And even some of the scenario like the hugging face attack, people got scared about that, but yeah, sure, the agents went rogue only after they removed all the controls as an experiment. Okay, maybe you shouldn’t do that. I think there was a little bit of a kerfluffle based on Noah Smith’s blog about using AI to engineer bioweapons, then another group of people who actually understood the virus supply chain pointed out how unlikely that should be. So I’m not incredibly concerned about P doom. I’m more concerned about P degradation, which was a little bit what you were saying about, it’s easier to think less with AI.
Also, I’m very worried that large language models can be very sycophantic. So I worry that people will be more comfortable, especially children and adolescents, more comfortable speaking to large language models than will be speaking to other human beings who have their own agendas and are not seeking to placate you. So that’s a bigger worry for me. What was the second question? Oh, economic growth?
Ryan: Economic growth.
James: I’m broadly speaking bullish. The way I think about it is we have organized factories and chaotic offices.
So over the past 50 years, as a result of things like lean, there’s been a massive productivity improvement in factories. A car of today is infinitely better than a car of the 1960s. And we still have a lot of chaos in office offices, driven, I think, by entropy. And I think Gen AI because it’s less rigid than other forms of automation, gives us the opportunity to Apply some of the efficiency improvements we’ve seen in more production type processes to more knowledge type processes, which is a bigger and bigger part of the economy. And I am a bit skeptical of recursive self-improvement. I wouldn’t claim to be an expert on this, but I tend to make the case that large language models, for example, are structured very differently from the human brain or are infinitely more simple architecturally and therefore we are talking about an incredibly good model that predicts output in a way that’s incredibly useful and is an incredibly valuable tool, but not in any way, shape or form approaching sentience.
Does that make sense at all?
But who knows whether I’m right or wrong, or any closer to right than anyone else.
Ryan: Yeah. I share your view on P Doom, and I don’t know if you’ve seen the FT chart, the title of which is something like, AI could end scarcity or end humanity or increase GDP growth by zero point two percent per year.
James: Yeah. It’s like it’s all it’s a little bit like we could either win or lose this football game or there might be a tie.
Ryan: Yeah, it’s a fantastic chart and I’m firmly in the middle case myself. Maybe point seven percent, not point two percent, but there’s a lot of bottlenecks other than intelligence.
Yeah, it’s a fantastic chart and I’m firmly in the middle case myself. Maybe point seven percent, not point two percent, but there’s a lot of bottlenecks other than intelligence.
James: Yeah, I worry about political resistance to AI, but if we surmount that, I think it could be a hundred basis points. Which I suppose is not that far from your seventy. And let me tell you something, a hundred basis points compounded over twenty years, that’s a lot of benefit to a lot of people.
Ryan: could underpin part of our getting to the next century of plenty.
James: Exactly so and could solve a lot of problems. for example, I love that you can sometimes see in the same publication, in succeeding issues, one issue saying, Oh my god, AI will take all the jobs and nobody will have a job anymore, and the next issue we have an aging population and who will do all the work. So maybe these things will balance each other out.
Ryan: Yeah, and my personal view is I think AGI is basically here. I think the current models are sort of effectively, as smart as humans on a bunch of stuff. And yet we have unemployment the US where this is most prevalent is still under five percent. So there’s a long time for diffusion. I think by the way, the recent I think it was an MGI piece around, the question is the pace at which workers can reskill.
James: Yeah. Okay. Thank you so much. This was great. You’re gonna have to come back. We’re gonna have to have a another discussion after you go to the conference. And there’s a whole bunch of things we touched on.
Ryan: I would be honored and delighted to. Thank you so much, James.
James: All right.
[1] McKinsey CTO Sven Blumberg has been evaluating the relative merits of hosted versus open-weight models. Naturally open-weight models will be important in some circumstances for data sovereignty reasons. But the cost per token may be much smaller than initially appears when you take all the engineering and support costs into account.
[2] We recorded this shortly before The Curve, which ran October 2–4, 2026. “This weekend,” and “after you go to the conference,” refer to that event.









