# AI Strategy: Distinguishing Incremental Tools From Fundamental Leverage

**Podcast:** Another Podcast
**Published:** 2026-11-10

## Transcript

Hi, I'm Tony Karen Brown.
And I'm Benedict Evans.
And we're going to talk about AI.
What else?
Yes.
Wet labs, doom, giant one-eyed aliens.
The end of the world and civilization as we know it.
I spent a lot of time talking to big companies about AI or listening to them talk about AI.
And there's a sort of, okay, we're doing this stuff now.
And next year, but what's the fundamental change conversation, which is very like, as I say, in a lot of contexts, it's like looking at the internet in 1996 or 1997, which is kind of where we are with AI.
It's a very big deal.
We don't really know how all of this is going to work and what the building blocks are going to be, nor what it will mean for particular industries.
There's an old joke that newspapers thought the internet was going to be great because their printing bills would go down.
The challenge here is, you know, you can say very vague general things like, you know, distance will mean less and physical assets won't matter in the same way, which is like kind of true, but not very helpful.
Or you can try and make super specific predictions about individual industries and, you know, do a big like consulting exercise and make big matrices and you'll get it all wrong.
But there's sort of something in between, which is to think, well, let's go push a little bit further on what are the kind of conceptual building blocks to think about what's exposed and not exposed.
Yeah.
And it feels like every week we have like a good couple of questions that sit at that weird middle ground intersection.
Yeah.
We've got a few this week.
Well, so there's a sort of a high level point, which is it looks like LLMs are going to be able to do.
anything, at a minimum, anything digital where you have enough training data and can validate the response at scale.
Maybe that's a duplication.
If it's digital, you probably can.
If it's not digital, it gets much harder to validate the response at scale.
But if it's digital, if you can capture it, if it's explicit and not implicit, if you can test the results at scale mechanistically, you can almost certainly automate this with LLMs.
And so that's one point.
That's a kind of a one fairly high level way to think about it.
And you push that one step further and you can say, well, we're going to have much faster, much cheaper and automation of many more things.
And just at a kind of base level, that might change a lot of industries.
And one can kind of dig into what that means and the Jevons paradox and price elasticity and so on.
But it struck me that there's a sort of there are two further questions to ask.
And one of them is a question that applied to the internet in particular as well, and probably PCs, which is you can make, and we've talked about this in a different context in the past, you can make an AI-enabled law firm, or you might have created an internet-enabled law firm.
Is the AI or internet or the PC getting at the key point of leverage in your business?
And clearly there are some industries where the internet unlocked something fundamental, maybe not obvious to begin with, but it unlocked something fundamental that created completely new possibilities, new market, new competition, demolished your model.
And there are others where it didn't.
So for the sake of argument.
So you mean like the services and the product completely changed the business versus?
Does this new technology?
actually get at the core of your business and unlock something, maybe what you thought was the core, maybe something else, and unlock something crucial.
Or is it additional tools?
And the answer is sometimes it isn't.
So if you go back and think about the internet, you could say that obviously the internet was transformative to the music business and the newspaper business and so on.
But the other extreme, if you made heavy machinery, if you made construction equipment, if you were Siemens and you made gas turbines, it's kind of a useful tool.
Everyone, you get email now.
Email's great.
But it doesn't actually change the business of making 20 or 30 gas turbines a year and selling them to power utilities.
And then there are people who are kind of in the middle.
So you could think about, say, banks or airlines where the internet unlocks neobanks and it means you can make a new consumer banking business without needing lots of physical branches.
Or now you have, if you're an airline or a hotel, now you have these new aggregators and new routes to market and new people who want commission from you and new ways to sell and get customers.
But at the end of the day, your business is still owning airplanes and landing slots and buying fuel and filling of seats.
And yes, Revolut's a giant business, but bye.
my parents haven't switched from their bank and the other banks are still there.
So it was an entry path.
It's a similar business for a new audience that's looking for more online digital banking first and only and take it from there.
It's marginal change.
It may be a new entry point.
It may be a new route to market.
It may be a new competition, but it doesn't actually completely overturn everybody else.
It doesn't mean that all the incumbents go away in the way that we saw with, say, with newspapers or record companies.
It's a new way of doing the same business.
Yeah.
Yeah.
So you might be John Deere.
You might be a bank.
You might be a newspaper.
And I think that's a kind of a way of, obviously, it's never as simple as that.
And as a consultant, everything has to be three bullet points.
But that's kind of a way of asking, just saying the internet may or may not actually change everything.
And probably making an internet-enabled law firm was a great way of getting customers.
But at the end of the day, Freshfields and Linklaters are still there.
The big law firms are still there.
And so this is sort of the first question is, well, okay, how does this change everything that, what happens if you're doing everything faster and cheaper?
The second is, does that actually unlock some fundamental point of leverage in your business?
Does it actually change the thing that matters?
I had this conversation on LinkedIn the other day with a lawyer because I was sort of wondering, what did legal databases do to law firms?
Because In 1980, you want a precedent.
You send an associate to the law library for a day or two.
And then in 1990 or 2000, you have a database and you're paying for the database, but you can get the precedent much quicker.
And the answer is actually that was a really small part of the business.
So who cares?
I don't know if that's true, but that was the answer.
But it might be.
That's kind of the point.
Maybe the internet.
the thing that the internet changes isn't important.
And this may be the case for AI as well for some industries.
Like this doesn't change anything that matters.
Or it changes everything, but it doesn't create new entrants.
It's just like you could say that computing changed everything about accounting, except it's all the same companies and all the same people billing all the same customers the same amount of money.
So you're saying sort of like in some cases, it's a useful tool, but in some cases, yeah.
And in some of the product has shifted.
Exactly.
And then...
Actually, I was at an event yesterday for people who make software for the insurance industry.
And I gave a version of my new AI presentation.
And one of the first questions was, like, we all heard a bunch about machine learning 10 years ago, and that didn't happen.
So what's changed?
Why did InsureTech, why did machine learning not completely change the nature of the insurance business?
There's another sort of chart I sometimes use, which is to compare the impact of Uber on taxis with the impact of Airbnb on hotels.
where Uber was transformative to taxis, Airbnb was entirely incremental to hotels.
And there's a whole bunch of reasons why that was, starting with this insurance event yesterday was in a hotel.
And it wasn't in an Airbnb.
Oh, that's a good point.
That's such a good example.
And, you know, when you travel or when my wife travels for business, you arrive at eight o'clock at night and you want a gym and room service and security and you're going to leave to client side at six o'clock the next morning.
No, you're not staying in an Airbnb in the suburbs of Charlotte.
So it depends.
I think the third question is, and it says that's that's a point that those are both points that sort of apply to every new technology in one way or another.
I think the third question is much more specific and conceptual about the nature of machine learning and LLMs, which is these things are telling us what most people would probably say.
And sometimes...
You mean it's giving us an average of how people think?
This is what a good answer would probably look like.
Sometimes that's what you want.
If you're doing your taxes, you want what a CPA would probably tell you.
Sometimes it's what would a million people probably say.
I was used to describe the last wave of machine learning as giving you infinite interns, like listen to every call coming into the call center, tell me if the customer is angry.
Ten-year-old can do that.
And that's one way to think about all of the cyber stuff that's going on, which is it's not that this thing is amazingly clever.
It's just there's 10,000 of them and they're really fast.
You spin up 10,000 agents and they try every possible combination of every possible thing until they find...
the path.
That's your million interns as million cyber interns, perhaps.
You could also say that's part of scaling.
And your question here is, is the desired answer outcome that average or is it something else that you're looking for?
Yes.
Do we want something that isn't in the training data?
And this is, you could argue, the sort of the essence of creativity is something that is different but good.
And the...
The theoretical problem there is that you can very easily imagine a model making more generic pop music or more generic images like that.
More difficult or certainly a very different level of abstraction would be for a model to say, hey, we've all been listening to disco and prog rock and ABBA and the economy's turned down and it's time for something completely different.
So here are the Sex Pistols.
Like more stuff that sounds like the Sex Pistols is a much easier problem than inventing the Sex Pistols.
Would it know this isn't in the training data, but people would like it?
Can you build a model of taste that can evaluate something new?
Or can you only evaluate how well it matches the things we already know people like?
This is, of course, I often go back to thinking about AlphaGo.
that AlphaGo could invent a new move, but it could test it.
This is the validation point.
It could make a new move, a new strategy that people had never done before, that people thought was a mistake, but it had an internal scoring system, so it could test whether that was a good move.
What would that equivalent be for making something that would appeal to people?
How would we know people would like that, but it doesn't match the training data?
There's a, the other side of the brain, I never remember if it's left or right, there's a discussion now of like, can you make new scientific discoveries?
Or can you only expand based on what we already know?
So I think there was a paper in the last week or two of someone sort of trying to give a model what we knew about physics.
before Einstein?
And could it make the jump to Einstein?
I'm not sure whether you can conclusively test the answer to that.
And of course, people didn't make the jump either until Einstein happened.
But there's a sort of theoretical question of how far can you get from just extrapolating?
You can certainly get some kinds of new discoveries with a million interns.
And again, this is a conversation about theoretical mathematics that was happening last week.
I am totally unplaced.
to take a view on this, but there is a sort of question of, are there some kinds of maths that LLMs are good at finding new things?
Is it constrained to certain sorts of problem where the answer is just test everything?
Or is it something else?
And I wonder if we already know this.
If we're aware of the instances and the use cases and the moments where we want and we're looking for the average or we're looking for more of the same, and what are the instances where, no, no, we need you to go beyond the average answer, the normal answer, we're looking for something completely new and different.
I wonder if that's going to shift how we approach these things.
Yes, I think you could probably propose that if you were at Anthropic and you'd ask Claude, do you think it's a good idea to start a wet lab researching human diseases at the same time as telling everybody we need to slow down?
AI research.
I'm pretty sure ChatGBT3 would have said, yeah, probably not a good idea.
Seriously, read the room, guys.
There are some questions.
We'll get your million interns to read the room.
Hopefully you get to the same answer.
In fact, you probably only needed one intern to answer that question.
Like somebody that you hadn't already shared a group house with.
But kind of zooming out.
I think this is a, these will be the wrong questions.
Yeah.
I mean, I always remember when I was a baby analyst looking at mobile in 2000 and the European mobile operators had just spent 110 billion euros on 3G licenses, which at the time seemed like a lot of money.
Not so much now.
And every client would say, okay, well, what's the killer use case for 3G?
What's the killer app for 3G?
That was the question.
And is it video calling?
Will people do video calling?
is it games?
Is it games?
Is it WAP?
Is WAP the killer app for 3G?
And the answer was that the killer app having the internet in your pocket anywhere you went was having the internet in your pocket anywhere you went.
But it had to be done in a bunch of ways that were completely different from what anybody was thinking about then and would basically take 10 years.
None of this is going to take 10 years, but it was the wrong question.
Or is it the wrong way of thinking about what the question was?
Go forward five years.
I've got a research paper from like IDC or Frost & Sullivan or one of those companies that's got five pages doing a forecast of how much money mobile operators would make from location lookups.
Like they would charge you 10 cents every time your phone asked for a location lookup because you would be doing it from the cell.
And they said, no, there wasn't like GPS is a thing.
There was no.
So these are all like, we don't know, but...
Do you think we fully understand?
We don't know, but it's interesting hearing you say, look, it was going to be the thing and it took 10 years.
Do you think we fully realise that none of this is going to take even probably a year or two?
Because I'm hearing your conversation, even with your guys yesterday, of just like, okay, this is where we're at today and this is where we think we'll be in a year.
We have no idea.
It's moving way too fast.
Yeah.
And it's moving fast in multiple dimensions.
It's moving.
And of course, the difference from the internet and from mobile and smartphones is that with those technologies, as I've said before, we understood what the physical limits of the technology were.
So we didn't know what the next iPhone would be, but we knew like it wouldn't fold.
It wouldn't fold.
Because we knew that technology didn't exist yet.
And we knew, you know, it wouldn't have a one-year battery life or a retinal projector or something.
Whereas with this technology, we don't really know what it's going to look like in two, three, four, five years' time.
Like what is physically possible for the models to be able to do.
We also, but back to 1997, you know, is it browsers?
Is it app stores?
Is it this?
Is it that?
Is it consumer video?
Is it an agent met as launched, as a consumer agentic platform?
Great work.
It's a lot of new habits.
So asking people to form all at once.
And so there's a big difference between saying everybody's going to have this and probably quicker than last time.
And actually knowing, well, what does that actually mean and which companies get broken apart?
And this is why I kind of come back to these questions.
Like, if you're making an AI enabled law firm, I know why you're doing that, because it's really hard to change how an existing professional services company works.
But is that actually going to change the fundamentals of what it is that the customer is buying from you?
I mean, maybe I can argue both sides of that, but we don't think we know yet.
Is the AI itself going to give you answers that are different enough or that add enough value?
And this is why you see a couple of the big law firms saying they want to make their own tools because what they don't want is for everyone to be using the same tools that gives the same answers.
That might be a little bit like an investment bank saying we want to have our own spreadsheet or we want to have our own model.
Well, you probably do have your own valuation models, but you use the same software to build it.
And what is the point of leverage?
Do you want the average answer?
Maybe the average answer in itself is a point of leverage.
That is one of the leverage points.
I actually just want it done the same way as everybody else, but I'm going to do something else.
But those are the sort of puzzles we had with the internet or with mobile.
Okay, you're using the phone, you're using mobile, you're using location.
Okay, is that the key thing that matters?
Or is the key thing that matters somewhere else?
And that seems like that's the same question this time around.
And I like the way you'd framed both of those three points that were framed of what do you get and what do you not get with a million interns?
And where's the question that actually matters?
I thought that was a nice way of framing how you think about this.
Yep.
Until the next time.
Well, until next week, until something minor and big change.
That's funny.
I go back and look at my old presentations and sometimes and again, like, well, that was right.
That was the wrong question.
Yeah, because that's the name of the game right now.
What the right questions to think about this week.
Maybe next week we have a bit more information.
The questions are actually no longer valid or relevant.
Yeah, I like the Groucho Marx line.
These are my principles.
And if you don't like them, I've got others.
There you go, folks.
These are our questions this week.
And if you don't like them, tune in next week.
We'll have others.
That's a good point to end.
Good to chat.
Get the shot.
See you soon.