73. The AI Impact Episode

0:00Welcome to The Difference Engine, the show for tech founders, investors and innovators.

0:08what are we talking about today well today we're looking at the topic everyone is talking about and few seem to be able to grasp the impact of ai yeah and we're going to channel steve jobs and hopefully forgive the english here have you think different as we look at some historical context to ask what can vacuum cleaners teach us about the use of ai also some up to the minute AI marketing advice, which we have allegedly co-created with the mighty analysts at Gartner. You may be surprised at how much we agree with them. For once. But first, we attempt to chart a course through the saspocalypse, which many see as AI's first legacy.

0:53We've mentioned before, in the US, that venture capital has seen its allocation shift somewhat. If you look at premier firms like Sequoia Capital and Andreessen Horowitz, for instance, their portfolios are increasingly dominated by AI investments. It feels like the classic SAS playbook is losing momentum. Yeah, SAS has obviously defined the last two decades in tech. Companies serving different types of applications like Salesforce early on and then Slack and part of Salesforce. Now, and Shopify have built entire categories around cloud-delivered software, but those markets are now maturing, and most new SaaS startups are only incremental improvements within the existing categories rather than category-defining breakthroughs.

1:42Yeah, exactly. The 2010s, if you can remember that far back, were dominated by SaaS category creation. I mean, if you could define the category, you could dominate it, identify a problem, build a strategy, a narrative before building the actual product. Yeah, the problem with AI is very different. And to date, all the AI winners that we've seen, the early AI winners, if you will, appear to be inherently horizontal. horizontal. When companies like OpenAI or Anthropic release systems such as Claude or ChatGPT, or even new model capabilities like Claude Work, the impact is immediate, but it's also across multiple software categories, customer support, coding tools, marketing platforms, knowledge management, you name it.

2:28To me, that's the issue that's causing the slide in these highly specialized vertical SaaS valuations. Why should we care? I mean, clearly a whole lot of value disappearing out of the market is an issue but from our perspective all this makes category ownership much harder if your startup is built on top of open ai anthropic or google deep mind the underlying intelligence layer is controlled by somebody else now that of course limits how defensible your category really is it's you know looking back it's like the early days in PC software. A whole load of application providers appeared, but they were building to PC-DOS, or later to the Windows operating system, which was ultimately controlled by, guess who?

3:18You know, Microsoft, the Redmond giant. Now, Redmond was quite happy that third parties built applications that encouraged more people to buy copies of the operating system, cash cow. But That was until the revenue stream slowed on the operating system. And then Microsoft, of course, did what anybody would do. It slowly but surely bundled all the third-party app suppliers right out of existence. Yeah, and of course, this has been noticed by our friends in the VC and private equity world. And if they've started adjusting how they evaluate startups, which, of course, as a reminder, mostly are valued on future earnings and not what they're doing today.

4:01So if those future earnings start to look even a little bit foreshadowed, they have an issue. During the SaaS boom, which they enjoyed for probably a decade or so, the sort of questions you ask yourself was, can this company dominate X SaaS category? Now the question is more like, what proprietary advantage, they also say competitive moat, does this company have? Is it data? Is it distribution? Is it their models? Or as we found out from the dinner we hosted this week, is it their brand? And I'd also add that there may have been room back in the day for five or six players in exactly the same category, which made fast following a very valid strategy.

4:40You know, you didn't have to be the absolute dominant. There's plenty of money fighting for the crumbs from the table. But now, unless there's genuine differentiation, life has become much, much harder. And as you mentioned, moats, the reality is the nature of moats has changed too. You know, if you think about it, SaaS companies relied on workflow lock-in, you know, combined with amazing sales teams. And we've had some of that experience on this pod in the past. And, of course, those amazing networking events, which we all remember with great fondness. By the way, if we could remember them. And there was also the cost of switching, which was often very high.

5:23But, of course, what people did as they were building those companies was ensure that there was integration into ready-made ecosystems. Now, AI companies are different. They rely on different things. They rely on training data, the availability of compute and memory access, and model context, and all quality. You know, we're generalizing a little bit here for this, talking about AI companies. But one thing we've noticed is the speed of category formation is a step change. So a SaaS category could take years, potentially decades, a decade or so to emerge. And from an investor point of view, this is lovely.

6:01This meant many years of compounding growth and optionality on when they could exit. They could ride something that was still growing and year two, year three, use secondary share sales, which the public never got a sniff of, to return their funds to their shareholders, et cetera. They had time on their sides. But AI categories appear and saturate in months or even weeks. Let's look at AI Coding Assistant, GitHub, Copilot, now ironically owned also by Microsoft, which was a snip at over$7.5 billion. That practically redefined the category overnight. But there's new players. Now we've got Lovable, we've got Replit, we've got Codex, we've got OpenClaw, and the whole vibe coding revolution has sped everything up.

6:47We've noticed that this sort of disruption leads ultimately to what we term category inflation. If you think about it, if you're living in IT every day at the moment, every new startup now describes itself as an AI platform. And we know the reality is many of them are just thin wrappers on top of foundation models. The sort of thing we've seen before has been paradigmatic shift in enabling technologies in IT. You know, when I wade around in the world of tech M &A during my stays, I've noticed that this leads to a curious M &A challenge, particularly in Europe, where individuals believe they can, out of the blue, create technologies that transform the SaaS experience.

7:35they believe that their consulting value-add channel will result in others paying huge sums for their apparently game-changing IP, which, of course, they've created in just a matter of weeks. Sorry, Europeans, no. Stuff like that might occur as an outlier in Silicon Valley, where there are layers of relationships and previous connections behind that decision to buy rather than to build. Anyway, lots of this happening. but I digress. Let's go back to that wrapper problem. Exactly. So we're focused on European category creation. And we recently spoke to a company in the code, let's just say the coding space, mentioned Vibe Coding and what it would do to their market.

8:19Didn't want to know because they had nice local, regional, current, live relationships with, let's just call them national flag carriers. So they're missing the point that the European wrapper problem is real. If a platform provider, let's say Microsoft or OpenAI, decides to ship the same feature as you are relying on overnight, you lose your differentiation overnight. And in the US, you can see this happening at rapid speed. You can see aqua hires, choke point acquisitions like OpenAI buying OpenClaw, Meta buying MaltBook. So OpenClaw for a rumored billion dollars, speculated numbers. and Maltbook for, again, speculative, alleged hundreds of billions.

9:08Who knows? This stuff happens lightning fast in the US. To date, this is not the game that we Europeans can play. Yeah, and these characteristics just reinforce the fact that venture funding is polarizing between a comparatively rarefied AI space and everything else, of course, including good old SaaS application layers. You know, high capital flows into and between foundational AI companies and infrastructure players like NVIDIA, OpenAI, Anthropic, and so on, while the funding for applications is comparatively sparse or non-existent to date. Yeah, seeing a lot of that in offered companies on the private market.

9:54It's ironic because during the SaaS dominant era, it was the application layer that captured most of the value. And infrastructure providers, they just had to feel like they were commodities. It's now the other way around. You've got a name for this, I think. Oh, of course. Yeah, yeah. So we call this the great software switcheroo. So the question becomes, if SaaS categories are saturated and AI is controlled by a few foundational players, where do the next categories emerge? Yeah, and again, I think some of that came out from the dinner that we hosted this week. But my guess is that AI-native workflows will be the new application layer.

10:37Instead of embedding AI into existing SaaS tools, that's, you know, all that AI washings, sort of discredited, disruptive new companies, and there could be a lot of them, actually, will build completely, design and build brand new workflows around AI from the start. Yeah, and this is paradigmatic shift, and that means there has to be a shift in thinking. And in this case, we think that the category shifts from something based on the idea of software tools, which we've been used to for decades, to something closer to digital co-workers. Yeah, exactly. And not the sort of co-workers we see sort of lamely offered today.

11:15So instead of tools that help people in theory work faster, we're going to see systems that perform entire parts of their workday autonomously. So coding is the obvious one, then research analysis, even operational decision-making. And it's the operational decision-making where the humans are involved that is the crunch point. The new challenge in category design is that you're no longer just selling software or even selling its benefits. You're asking companies to trust an AI with decisions about your business process. Yeah, and that introduces, dare I say it, governance, mature themes like oversight, and of course, trust.

11:58And the real category battle here in EMEA might end up being who defines that operational layer that sort of governs or manages the AI systems, because that's a place where a lot of value can be added or subtracted, and real category power can be exercised. One example which we helped launch in this space is a British company called Azoma. It's Agentic Commerce Protocol AMP, which stands for Agentic Merchant Protocol, is in fact, is a category definition. And it creates a new system of record. It's not a platform in the sense of an LLM, of an infrastructure platform. It's a layer with a real value to its ideal customer profiles, who, by the way, include customers like Mars, Perfect Head, and several other well-known brands.

12:44It sits above Amazon and the other online shops, and gives control back to the brands and the merchants, who, let's face it, spend billions to create customer pool, only in some cases to hand over all of their creation, all of their creativity, all of their brand power to these massive hypermarkets, who they can probably go ahead and advertise rival products against them. And that's why we've got the Azoma CEO, Max Sinclair, on a future episode, sharing some of the lessons about what this new type of category design involves. So I think if we're to draw some lessons from this, is that the next category might not be dominated by the smartest or, as has often happened in the past, not the smartest but best distributed AI model.

13:30Yeah, right. It might be the company that best defines how AI actually operates inside large organizations. And that could be the savior of some of the massive consulting organizations who are, and I'm thinking of PwC at the moment, who currently are beginning to and maybe aggressively re-evaluate their top to bottom structure for service delivery. Yeah, well, that might be good news for PwC, but it's also going to be, or could be, good news for the likes of Accenture, IBM Consulting, Deloitte, Capgemini, Cognizant, Tata, Emphasis, Wipro, and, of course, their employees, but only if they can reskill fast enough.

14:18Well guys, in this increasingly volatile world, many of us are looking for digital sovereignty. And now, more than ever, our national health providers, those in the law and order community, governments and public sector organisations want more control over their communications. They want a digitally sovereign solution, which can never be interrupted based on external demands, no matter how powerful they are. And a service which means they are not reliant on a single vendor. Opting for open standard software means a competitive vendor ecosystem and solutions which are designed to interoperate with other organisations infrastructure.

14:52This is where Matrix, the secure communications open standard, plays a vital role. Most European governments already use matrix based solutions for communications and Element is the leading provider of matrix based secure communication solutions. It's already working with the European Commission, NATO, the United Nations and governments in France, Germany, the Netherlands, Sweden and many others. Element provides sovereign, interoperable and end-to-end communications. If you're responsible for a nation's trusted and resilient communications, take a look at the Matrix Open Standard and Element at element.io.

15:25We've used Element ourselves and find it fast, secure and so easy to adopt. If you want digital sovereignty, you have to check out Element. See our show notes for more information and get on board with the new secure world of communications today. All right, so today's lesson from tech history is two paradoxes and a law. Well, everybody's been talking about agentic this and agentic that, a lot about what AI agents are going to do to knowledge work. We're going to try and look back at work-related lessons from yesteryear and see if we can tackle some of the principles of work that seem spot on for today.

16:01It all starts with a classic work of women's history, a book from Ruth Schwartz Cowan. Definitely worth reading if you can get past its title, not terribly snappy. Its title is called More Work for Mother, The Ironies of Household Technology from the Open Hearth to the Microwave. It did win the 1984 Dexter Prize from the Society for the History of Technology, a very august body. And you know how much I love musing on the history of tech and what it can teach us. Now, in this case, what I particularly love is that it describes one of the most counterintuitive results in technology adoption. Labour-saving household appliances don't actually reduce labour because we've spent most of our lives listening to a whole load of promises and concerns about machines or machine adoption eventually leading to the sunlit uplands of endless leisure for us humans.

16:59It cannot be a coincidence that this paradox about labor-saving devices not saving labor is about domestic tasks. The reason I say that is even the most highly inventive nations on earth, the Japanese and the Koreans, who have got aging populations and real concerns, can't find solutions for these sorts of problems. Exactly. And this phenomenon, which we described with the very snappy title, is now thankfully known as the Cowan Paradox. Now, to me, that cemented Ruth's place in tech history. I reckon, though, that it's pretty useful as a guide to predicting what AI agents will actually do to work.

17:39Are you sure? I know you love this stuff, but just give me the TLDR, too long, didn't read version, please. All right. Okay. Okay. So take the vacuum cleaner, or Hoover, if you want the Boomer generic, or perhaps even now the hurrah for the UK, Dyson. Now, before electric vacuums existed, cleaning rugs, if you were flush enough to afford anything resembling a carpet, was a huge ordeal. It involved moving furniture, rolling up heavy lumps of textile, taking them outside, hanging them on a line, and literally beating the dust out of them with paddles. Sounds good for your mental health, actually.

18:15Probably good for your mental health, very good for your forearms, and cheaper than a gym membership. But if you think about it, it took multiple people to do this. And sort of because of that and other factors, it happened maybe once or twice a year. Now, of course, that's where the idea of spring cleaning eventually came from. Right. So just cleaning, therefore, was infrequent because we're humans and it was inconvenient. It was exhausting. It was time consuming. It was effectively expensive in terms of labour. Getting all economic there. Yeah, in economic terms, labour was comparatively cheap back then, though.

18:48Still quite knackering, though. Then electric vacuums appeared in the early 20th century. This new category of technology, as is what happens when new technologies do burst on the scene, actually required the perfection of a number of innovations, from domestic electricity supply, smaller motors, new materials, and so on. But the point was suddenly you could clean the rug or increasingly the fixed carpet actually installed because of the arrival of the Hoover, right where it was sitting. No lifting, no helpers. That's handy because given the time post the Second World War, women were entering the workforce as, and I mean the non-domestic workforce, what we would call work.

19:34So logically, the time needed for housework had to be reduced. And here was a technology ready to deliver this very time-saving, time that could be used like even more value-added tasks or even leisure. But that's where you're wrong. It didn't. It absolutely did not. That's actually mostly determined by three things that changed at the same time. What happened because you got these handy Hoover Dyson vacuum cleaner things was that the frequency of use exploded. Once cleaning became It became easy, society changed, social expectations shifted. So manufacturers, of course, drove demand for their product based on the stuff you see coming out of PNG and lever these days, scare tactics about germs, disease, and dirt.

20:29So instead of cleaning once or twice a year, and everybody did that, so that was fine, households were guilted into vacuuming weekly or even daily. So you're doing it more. So second, the actual labor got reassigned, right? So think about it. Beating rugs actually required physical strength. So, you know, men or hard help often did it. Vacuuming, of course, was, I mean, seen to be and actually was easier. So it got reclassified as lighthouse work, which meant it became the wife's responsibility. How convenient for the patriarchy. Anyway, third thing that happened was the helpers disappeared. And this was a product of post-First World War where there weren't that many people around because they'd been killed somewhere in the mud in Europe.

21:21And middle class households that used to employ domestic servants couldn't find them or found there was an alternative. So for that level of heavy cleaning, we now have an electric servant replace the human. And of course, if you think about it, the same thing happened with domestic laundry. You know, you used to ship it out for somebody else to do, or, you know, the woman of the house would spend hours and hours on the tub, working away, cleaning the clothes. But then what happened was laundrettes opened, so all the capital was put into putting communal machines in. And that became a sort of halfway house before everything got cheap enough and small enough.

22:04And we started to own our own laundry machines in the UK, certainly by the 80s. There you go. The labour's disappeared. It actually merely transferred the labour, but not the time taken, because the launderers, mostly housewives, needed to mine their washing and transfer it to dryers. The Cowan paradox, there was relatively little time saving. All that happened was the format of the work just moved, and in fact, some of the tasks multiplied. I can see the argument you're making. In fact, I witnessed that just this week. Somebody was talking about SEO roles in a consumer packaged goods companies.

22:40And yeah, that work needs to be done. It's just less important and probably broken up into different tasks. So looking at keywords, endlessly updating HTML pages, this is not a job for humans. And thankfully, with AI, the nature of this work is actually changing. This sort of shift changes the very meaning of quote-unquote work and reduces the need for certain job roles in paid employment, just like it did in households. Yeah, that also exemplifies how quickly labor shifts, because there was a point when endlessly updating HTML pages was a job for humans. You know, just like doing the laundry, a job you really wouldn't want to do.

23:21The Cowan Paradox basically says the time spent doing the work, however it evolves, basically does not change. So empirically, time use studies from the 30s through to the 50s showed that housewives, men were frankly really not doing their share of work then, the housewives were actually still spending about 51 hours a week on housework, basically unchanged before any appliances arrived. Right, and that's the heart of the Cowan paradox. Labor-saving technology doesn't necessarily save labor, it resets expectations. Do you think that means we can expect to see more production of HTML pages and more human checking of the output maybe?

24:02That could be a result. I mean, Cowan argued that the real effect of technologies is raising the standard of the output, right? And obviously that's a very subjective idea, but it certainly is true if you think about it. When something becomes cheaper or easier, society demands more of it, right? So is that beginning to sound like the current AI debate? Now for our second law, and this is one that many, many people are talking about in the age of AI, and it's the classic efficiency paradox described by William Stanley Jevons, who published the theory of political economy while he held the chair of political economy at Guessware, our alma mater, Manchester University.

24:45Hurrah! Yes, and there's a whole building called the Jevons Building, named after him. Anyway, the Jevons paradox. Right, back in the 19th century, Jevons noticed that as the newfangled coal-fired engines became more efficient, total coal consumption actually increased. Yeah? efficiency lowered the cost per unit, which encouraged more use of those new fangled engines. Same pattern again. Induced demand, added roads, and you just increase the traffic, improve fuel efficiency. And guess what? People drive more miles or choose bigger Chelsea tractors, as we call them in the UK, SUVs, as our American friends call them.

25:27And they seem to have negated the theoretical overall reduction in fuel consumption due to more efficient engines, because they're massive lumps. They're great big bricks, which you're trying to throw through the air. So what this says is sort of in quasi-economic terms, efficiency gains get absorbed by demand expansion and don't deliver reduced effort. So your argument is that these AI agents that people are touting are basically the vacuum cleaner of knowledge work. Well, they certainly hoover up dull tiles like SEO, but should we not expect a life of leisure just yet? Yeah, exactly. They'll make certain tasks dramatically easier but it won't reduce work.

26:06It will increase the volume and expectations around knowledge output. Now, I think we're already seeing early evidence of that, and guess what? The academics are on it. The Harvard Business Review, a very august journal, which I'd recommend you all to read, recently published a study by Aruna Ranganathan and Jinki Maggie Yee. Massive apologies for the pronunciation, but we're mere Brits. Anyway, these two academic stars spent eight months embedded in a 200-person tech company studying what actually happened when employees adopted AI tools. Now, this is interesting. And let me guess, the result wasn't less work, was it?

26:49It sounds like, if it wasn't, that the AI in question was Microsoft Copilot. Don't get me started on that one. Not commenting on the tech. given I spend most days working with somebody who appears to be half man, half co-pilot. So yes, yes, no work reduction. What they saw was three clear patterns. You have to think these things are all deeply rooted in human psychology, right? So what happened was that there was task expansion. Tasks expanded, and we'll explain this in a moment. work boundaries started to disappear because of friction reduction in actually getting the stuff done and there was an explosion in multitasking presumably also driven by the idea that people felt they needed to do more of the AI to keep their basic jobs so if you think about task expansion what happened was that they observed that people didn't just do their own jobs faster they started doing other people's jobs.

27:52Product managers, right? So this is what happens when you get vibe coding. Product managers began writing codes. Researchers started handling engineering tasks. You know, God, God, work that previously required hiring somebody new got absorbed by existing employees, right? Right, right. So in a way, the AI agents here are like the old domestic servants. They're being replaced by, if you like, housewives, that's you and I, doing a lot more work yeah but it's also but you know to use extend that analogy it's also the housewives justifying their existence uh for want of not knowing what else to do um right so the second thing was that the work time boundaries disappeared i mean what is a work yeah what is a work time boundary i've never i've never had one in my life so in this more apparently structured environment the work time boundaries disappeared and this was because apparently the ai tools sort of feel conversational sort of human-ish.