AI deployment

Hi, I'm Tony Cameron. And I'm Benedict Evans. And what were we going to talk about today? AI enabled companies. Oh yeah. So. And is it potentially already a data phrase? Well, so I thought there's a sort of an interesting thread in here. I was at the N.O.F National Retail Foundation Federation in the middle of nowhere last year, last week. And I was talking to a lot of corporate CEOs and I spent a lot of time over the last 12 months talking to CEOs and consultants. And the strand I suppose to think about is that, like I think, slides and so the slide is like for any new technology, as a company, you have three questions. The first question is how does this change your operations? Like how we actually run our business? What do we do with it internally? The second question is, well, how would we buy it and build it and deploy it? Do we go, do we pay someone to build it for us? Do we buy it off the shelf? Do we go to a startup? Do we need to build our own custom thing? And then the third is how does this change our market, our product, the way we speak to our customers? And so simple lens to apply to that is, you know, imagine if you are a caterpillar and a newspaper company looking at the internet. Both of you have to think about what this does for your operations. Both of you have to think about how you buy, build, deploy this. But for caterpillar, the internet didn't fundamentally change the nature of your industry and the newspapers, obviously it did. And the same thing sort of applies to it. And then of course there are companies, those are both companies that are at each extreme and then there are kind of companies that are somewhere in the middle, like maybe an airline company. Like there's a change quite a lot about how you go to the market, but actually your industry hasn't changed. Or maybe like auto trader would be an example to think about there or real estate listing company. So you have also people that are kind of somewhere in the middle where there's change a bunch of stuff about your industry, but also didn't change a bunch of stuff about your industry. And how would you buy this and how would you deploy it? And so I think that's an interesting framing to think about. And then it kind of mapped out again what's been going on in AI in the last two or three years. Step one was give everybody K pilot and that didn't work too well because it's kind of all-clawed or attached to EBT because it's kind of like being in 1997 and giving everybody a web browser. Okay, this is the future, but it's not that's not what was the productivity to giving everybody in the web browser in 1997, right? Not particularly clear. And then you do a bunch of pilots and some of the pilots go to production and some of them don't and you automate a bunch of kind of point solutions in your back offices. And maybe you get cost savings, maybe you don't, but that doesn't like get to kind of fundamentally change the nature of your company. How would you do that? And that becomes a B&B CT McKinsey question or it becomes an Accenture Fuginizing Kindral PWC kind of a question because big companies don't have lots of people sitting around not doing anything and working out how you would completely change your company and then building it is really hard and this is why strategy consultants and system integrators exist. And is that question about you buying software or are you buying labor? What are you buying? You go to a bank or you go to Accenture and they help you work out what to do with this thing and how to do it. And so one of the things that drops out of that in the last few, the lot this year particularly is this idea of the AI enabled company which is we're going to do a new law firm but based on law on software and based on AI. We're going to do a new XYZ and new accounting firm and new architecture firm and new XYZ but based around AI. Or we're a PE firm and we're going to start a new PE firm and we're going to buy a bunch of companies in a XYZ space and roll them up and quote unquote, fix them or improve them or transform them with AI. And the thing I was sort of thinking about here is imagine you do a search and replace on everything I've just said and replace AI enabled with internet enabled or PC enabled or computer enabled or digital. And then you get this wonderful phrase digital transformation. And going back to my kind of my set of three bullet points at the beginning, the thing I kind of wonder here is is that the key point of leverage. And this was a thing that kind of applied with the whole software, it's a world thing from Mark Andreessen 10, 15 years ago. Does software allow you this new technology XYZ? Get to some key point of leverage in the industry. Does it unlock something that couldn't be done before? Does it get to the key challenge? Or is this something that will provide to consumer surplus marginal benefit, generalized improvement across the whole industry? And some people will get there before others and some people will take share. Maybe some people will take quite a lot of share. But in the end, that's not all the businesses which would be the extreme case would be here would be like, I mean, airlines would be an interesting conversation here. Like online travel became a huge thing. There's hundreds of billions of dollars, there's booking.com and Expedia. On the other hand, and that maybe that unlocks easy jet and Ryanair. And maybe the unlocks now at a new competitor. But at the end of the day, the business is buying and hedging fuel and owning landing slots and buying and leasing airplanes and filling them up marginal cost. So equally, think about near banks like Monzo and Revolut and all these people. The internet's a new route to market. You're starting software first so you can build a much better experience and you can build different products and you can be much more nimble and you can do all this cool stuff that the legacy bank struggle to do. Okay, great. What happened to the legacy bank? So kind of still there. Sometimes that's a regular is bad. It's like challenging example because of regulation. But like say, look at the UK Revolut and Monzo are these great things. Fine. But like, batteries and lois and halifax and all the halifax, boggles and lois and hfc and so on. That's all that. So this is sort of question of is this the new point of leverage that lets you transform this business? What's the right way to take that to market to implement change to build? Can you unlock something critical within this industry? Is this a new route to market? Or the end of the day is saying that we're an AI enabled XYZ kind of like saying we're an internet enabled law firm in 1998. And going back to the concept of that phrase like saying we're an AI enabled company, it was interesting thinking back of just like we didn't use to say we're an electricity enabled company or an SQL enabled company company. But we did say e-business, we did have a whole concept of digital transformation and I wonder is it because is it when we're in a period of confusion where we actually don't understand the world in which we are in that we go for those sentences such as AI enabled company. When we're not like why did we not why did we I guess why did we not have a moment in time where we said we're an electricity enabled company but we do have a moment. Well I think we think I think people probably did get that to the 20s or 30s. Yeah, I think I think of like do we go to these sentences when we're in a period of confusion where we don't actually understand and so we just need to etiquette. So some of this I mean some of this is you know this is you know you get two textbooks from the 80s and 90s bitten by partners at KPMG or Accenture or Anderson Consulting that are called things like change management and how you adopt digital because clearly the deployment of mainframes and data processing and spreadsheets and EOPs and so on did kind of completely change the economics of your business and that came with a whole bunch of challenges most of which were really internal management challenges not 10 at operational challenges not engineering challenges they were this is where you get phrases like pilots and lighthouses and quick wins and you know how do you manage into a stakeholder management how do you change a 100,000 person organization which are quite separate to does this actually change in HR industry or is there something that just everybody will do and it will well it will change in HR industry does it change your competitive landscape always just jumping something that we're doing our competitors will do we'll all get some cost savings out of it and some efficiency gains that will get competed away in price because we're all doing it and the end of the day as long as you don't screw it up this doesn't change the competitive dynamics very much or is this some key leverage that will allow somebody to come in and break apart this whole space and let everybody do something in some let someone do something in some completely different way the extreme I got this catalyst bizarre conversation with a prominent doomer on Twitter who said something on their setting they're like companies aren't going to throw at their ERP tomorrow and refile their accountants into a place it was Claude and he said well then they'll be destroyed and replaced by companies that do and I'm like okay so Sanco Bay is a big French glass company they're going to be destroyed by a new glass company that's using Claude to do its accounting you know it's a reductio ad absurd but of course we'll do this absurd but that's a sort of you know mundane boring fuzzy reality of what happens when you're a 100,000 person company or 50,000 person company and you want to change how you do stuff this is some of this is the thing we talked about on the sort of the last podcast I think which is the one about most people on tool builders which is you know imagine you are working in accounts payable inside a 5,000 person company and there's 150 people in accounts payable I'm making this number up and that first of all you're not the kind of person who's going to see we could do this differently if we use AI. Secondly if you were you're not the kind of person who's going to think in detail and know well this is how that would work and these are all the things that that would do in this completely new way like designing product inventing product knowing how product works is a different skill and thirdly most importantly even if you were like you Mr 25 year old Mr 25 year old working in accounts payable 30 That's not there's 150 other people and it's plugged into SAP and it's plugged into drive and it's plugged into workday. You can't just like go out and rip it out and replace it yourself. That's a project that needs to be determined centrally within the company. And then the company needs to sit and work out how they're going to do that. And that's a do we buy it? Do we build it? Do we work with the center to integrate it? Do we pay the center to change it? Do we hire somebody to build us a new thing? Like what? And where does that go in our budget cycle? And we do it this year or do it next year because we've got other stuff we do this year. And right now we're in a three year migration to SAP Hanna. There's a sort of, I suppose, stepping back from this like, it's a thing that occurred to me with the latest open letter from people who work at AI companies. Someone suggested the World Test for AGI is that any software that can actually replace researchers needs to be able to write open letters and sign them. But part of this is like you've never actually worked at a company that's been around for 40 years and whose main business isn't technology. And you have no idea what SAP is and you've never seen it. You've probably never even seen Excel to go up. And you think that the world is just going to change completely in 18 months. And so then you think, well, then the US government and quote unquote the international community or international measures can create treaties and new things that will change all of this. So you don't actually know how the world works at all. You don't know you're worried about this because you don't know how the world works. And you also think your solution is based on also not knowing how the world works. Now that's a slightly kind of glib statement. But yes, this clearly is going to have a bunch of major effects on society. But it's not going to be quick and easy and simple. And if you think it is, you would have to think that you would have to not understand that it isn't quick and easy and simple to think that like the US government can call a meeting and then everybody in the world will just kind of agree to stop doing AI for six months. Now I think that's an interesting point that we talked about previously also just like this concept of change management of just for the last what let's say 10, 15 years. SAS was, you talk about this a lot, SAS was bottom up. It lands with a team you expand and that works when the tool is replacing, when the tool is making an individual better at the job that they already have. It fails and gets complicated when the value requires real actual change management. And there's something interesting there of just bottom up gave us faster technology adoption. But we didn't get only up to a point and we didn't actually get organizational change from it. And so what happens when the companies are no longer buying software from it? I don't know, it becomes interesting when the value is enter and process. So that crosses the department, that crosses different budgets. That's not just making your job easier and better. And principle, if you're going to change the process across multiple departments, multiple people, multiple systems, that cannot be bottom up. That has to be a decision taken by the company. And it's like it. And it's such an obvious statement. It's like bizarre that you kind of have to say it. But of course, of course, you're not going to just replace workday with Claude. Now what you will do, this place that I think I always talk about is what this does do is it takes over from the stuff that you might have done in Excel or you might have done in email or you might have done in a spreadsheet or a shared folder or something. And it also adds a bunch of stuff that you might not have been able to do in SAP before and now you can or you might not have been able to do an third party product and now you can. But all of that, and the thing that occurs to me here also is we have this sort of, this conversation is about the state of models and the sort of on the basis of, well, you know, Chinese models are still three to six months behind the cutting edge. And you think, Jesus Christ, it takes three to six months to get somebody at an enterprise to reply to your inquiry about selling them some software. You know, the enterprise sales cycle is what, 18 months at best. You want to add a new thing and plug it into SAP and Stripe and Workday and your payment system. We're not going to do that in three months and we're not going to change it every month either. You're bit about, it only gets replaced if you can actually get the person to write the letter and sign it. It leads me to like, if everyone can buy the same tools, the same capability, the differentiated N is what actually can't be bought. So the, I guess the conversation that I have these days is what is it that can't actually be bought? The proprietary data, the trust, the taste, the word that's being thrown around, taste, the accountability, because you have to sign your name to the final product, I guess, or the output. Well, this was, I mean, this was one of the big law firms said that they're going to spend a couple of hundred million dollars over the next five years building their own stuff because they don't want to just use the same legal software as everybody else. Now, you could argue both sides of that and that might be like saying, well, we're going to build our own WordPress server because, and there was a moment you get back to the early eighties when there were like 50 people making spreadsheets and word browsers, as Boeing made Boeing, actual that Boeing sold a spreadsheet program. And like, and the selling point is, I'm not sure if it was unique, but I saw an ad for it and the ad is it's a 3D spreadsheet, which means tabs. I mean, we had presidential candidates at Nation Bill, the same we think we can build, build this software in house. No, you're fucking can't. Why? Well, there's two points here is age, there's age should you try and build it yourself, but the key, obviously, the key determinant at every point in the last 30 years is, is there something where you need to do something unique that's specific to your business? Or should you just be buying the same thing that everybody else uses because it's a commodity? Nobody writes their own wordpresses today. Some people do create their own CMSs. Maybe or maybe not a good idea. Should you create your own CRM? Well, it depends. Should you customize, or should you spend a lot of money to customize Salesforce? Well, it depends. Where is this the key point of leverage? If you're a JP Morgan Chase, then no, you don't use Stripe. You build your own systems. On the other hand, you don't build your own wordpresses and you don't build your email servers. So do you think the bottleneck, bottleneck, quote, and quote, has moved then from producing this to actually verifying it, like running a diagnosis, doing the audit? Is that the thing that's coming? I think it's difficult versus the producing part. I suppose this is all a very long way of describing how traditional or normal technology works. And saying, "A.I. has to kind of go into those flows of normal technology because there are, you know, the reasons why software is bought in these five or ten different ways and used in these five or ten different ways, do not change when you say, AI." True. Some of them do in some narrow specific places like, "Will I still do that in SAP or will because it's too hard to do it in Excel? Maybe now I can do it in Florde. Will I still use Nation? Maybe I will now do that in Co-Work." So there's clearly places where the boundaries shift around. But the core of why do we have one system of record for this rather than everyone doing it in a spreadsheet on their desktop doesn't change if now everyone's desktop has a copy of whichever AI tool it is. And that gets me back to these deployment questions which are, "Okay, do you buy it? Do you build it? Do you hire Accenture? Do you hire Bane? Do you work with Microsoft? What are the salespeople from Anthropic trying to sell me?" I have a friend of mine said, "Bend it. You think in slides?" So I have a slide. And the slide says, "All AI because I've been doing lots of core presentations." I get occurred to me that all the questions I get from a fit into two categories. The answer is either no one knows or how to do it at work last time. How did it work out? It says a broad class of question like what will happen to the science or the model scale, etc. Where the answer is no one knows. And then there's another broad class of question which is, "Okay, how did you deploy cloud? How did you deploy?" So how do you think about social? How did you think about mobile? How did you think about the internet? What will this do to middle class employment? Well, what did the internet do to middle class employment? What did PCs do to middle class employment? Completely changed all of it. And of course the next slide is, of course no one knew last time either. Like we didn't know how the internet was going to work. We didn't know how mobile was going to work. We don't know how this is going to work. But the core of it is, there's clearly a bunch of ways where this is not a normal technology. Most obviously we don't have a good scientific model of how the models work and so we don't know how much better they can get. But then you get into the nitty gritty of what happens when a company uses this stuff and the answer is, well, there are reasons why they do it like this that don't change with this new technology. And so it does kind of look like the way you deploy it. So why before? Do you think the buying process has to be reinvented before we figure out the employment? Like are we buying, do we even know how to buy these new tools? Because we spend a lot of time talking about the deployment. The deployment is the hot thing. But well, everything is the same as everything is different. So with SaaS, you move from buying the thing one off, doing an installation, doing integration, putting it on-prem, putting Windows 32 software on everybody's desktop, to thinking about single sign-on and a website and what how many seats have we bought? How many seats are we using? And what's the feature upsell? So clearly you change how you bought it and you change how you bought and sell software with SaaS. And clearly, there's a very narrowly, you're paying for tokens. At least at the moment, you have marginal cost of tokens and so the least if you don't have them, then the software company has them. And so there's a whole bunch of ways in which the way that you buy and sell this looks different at least at the moment. And you need to think about, well, you may need to think, well, do we want to be model agnostic and do we want to haces locally and do we want to use an open-source model that's one on our own, one on our own infrastructure? And so there's always, it's always different in some senses. And there's always stuff that's unpredictable. I mean, going back to what I was saying earlier, there were some industries where it was really clear that this was. wasn't actually going to change fundamentally the nature of your industry. And there are others where you thought that and it turned out to be wrong. And there was lots in the middle that evolved in kind of weird unpredictable ways like say the LN industry online travel working. And you're sort of right at the beginning of that process. But you do kind of have the same questions that you had last time. It's interesting because I don't think I've ever personally thought about it probably because I've never been in a position of having to buy, you know, for a large company this type of software. But I've never thought about the actual, the way we buy these tools and this software does actually impact how it gets deployed. Or how we're thinking about the deployment. Yeah, you know, and I had to have a coffee a few weeks ago with somebody who works through an extremely large company. And they are, their conversations about AI are, firstly, like how do I think about AI bias? How do I think about model cost? How do I think about open source? How do I think about locking myself into something because it might be exactly, I mean, I think this is you, because it was telling me this the other day, it's talking about e-commerce companies in the late 90s that were buying like, did you tell me this? No, I don't. The e-commerce companies in the late 90s would say, well, we've done a 20-year contract for our CMS. Because they thought that was, because that was how you bought software in the late 90s. How you bought software? I mean, my god, we're still doing this in sports, by the way, everyone just signed an insane 10-year, 15-year contract, broadcasting contract with Sky Sports. And like, if they say, "Rome, we're signing Sky Sports," it's going to be around in 15 years. And it will look anything, all the value will look anything like what it looks like now today. Exactly. But kind of finishing the point, I mean, what I was saying is like, part half of this, the, this VPs conversation was like, trying to understand how AI, the other half of it was, I need to think about quick wins versus Lighthouse user cases versus pilots versus building deployment versus, you know, you go to the index of your book on PC deployment from 1985 and they're working through that. Like, how do I think about, how do I deploy a big transformative piece of technology inside a big slame-moving organization with lots of different stakeholders? And some people who would very happily spend $1,000 of token as a day. And some people who would spend five grand of token as a day and not understand that they've done that. And then people who think this is all the way to time, which is exactly the way the internet looked all the way, social looked all the way, well, looked. And so half the questions are AI questions and half of the questions are logistic and all like, how does it be accompanied by software? And none of it is, well, we'll just give everybody Claude and go, "Hey, or Claude for X or Claude for Y." It's like, we're a law firm, well, we'll just buy Claude for law and then that's it. Like, then we'll just file all the associates. Come on. Because it's an interesting distinction tool, so just like co-pilot versus autopilot and those are different sets of tools and the output looks differently and then the labor. What's autopilot? I don't know. Isn't that the thing that's happening right now? There's co-pilots which sells a tool to a professional and they remain responsible for what they're building and autopilot feels more of just like, "Okay, this is a thing that's just going to go out and build whatever we say it." You mean, you mean, you mean, "Keywork" or "Codex"? Yeah. Well, so, I mean, I think there's a separate narrow problem here with co-pilot, which is emerged quite quickly. It's not really a meaningful tent. It's not a product. Co-pilot is rather like Watson. It's a brand name that represents and that represents every single PM and Microsoft had to add an AI feature as part of their quarterly goals. And so every third button at Microsoft, every Microsoft product is purple and has sparkles and they're all called co-pilot, but none of those 45 different systems, sort of none of them are connected to each other. But they will call co-pilot. So they don't use it, don't understand. It's not clear to the user that that co-pilot is not that co-pilot, which is not that co-pilot. And also the CIO is like, "Why am I paying for all of this stuff from my users? I'm using any of it." Or they're not, you know, "I'm not clear why my user is what they're doing with it." So that's a kind of a narrow deployment equally. It's another narrow kind of deployment problem. It's like, "We gave everybody AI. What does that mean?" As opposed to, how do we work out the places to deploy this? How should we end this? What do you think, the where's your head at? Once you've had a conversation like this for 20 minutes? Well, so I think there's a, again, I'm becoming something I say too often is like being at the internet in 97. It's everything is unclear and the way this is going to work is unclear. But the central part of it is that this is where entrepreneur, this is what entrepreneurs do. This is what startups do. They come in and work out, "Well, those use cases are, work out how to sell them, work out how to explain to people that you have this problem that you never noticed." And that is the function of a startup. The function of a startup is to unbundle. You unbundle Excel, you unbundle email, you unbundle Google and SAP and Salesforce and Workday. And then you turn them into a new bundle and you identify tasks or projects or problems that people didn't see before. This is also why you go to B&B, CG McKinsey. There's a sort of an interesting overlap here in that the function of an end between the function of a piece of enterprise software, an enterprise software startup and a consultant. In that both of what they're both at them doing is they're kind of looking at the company and saying, "If you realize that you were doing this and if you did it like that instead, it would be better." The difference being that you pay the consultant to work that out for you, whereas the entrepreneur has to work it out themselves and then come and pitch it to you. But it's kind of climbing the same mountain from different sites, which all of which makes it kind of unintentionally hilarious that people think this is the end of consultants. It's kind of a point I've come back to several times. Nobody is going to work this out by themselves and most people are going to pay somebody to do it for them. There we are. That's a nice optimistic way to end. Yeah, I think that's a great way to end. Good chat. I will speak good to channel, speak you next week. Bye.