How the corporate AI productivity lag is costing billions

0:00Welcome to Path Founders with me, Mike Butcher. We like to investigate the code behind the entrepreneurs, the capital that powers them and the consequences. For the last two years, we've been told that generative AI would make work faster and more productive. What happened? Well, actually, very little. Now, new research suggests that inside large UK organisations, AI is actually creating a brand new kind of drag on productivity. Employees are spending almost as much time checking AI outputs as they do generating them. The report suggests that this is leading to a£29 billion productivity leak, rising anxiety and so-called AI burnout.

0:46Now, normally, this kind of report comes out of people like consultancies who can often be sceptical of AI, and it might well eat into their business, so they have a bit of an axe to grind. But it's emerged from an AI startup and someone who helped actually define the modern AI interface. William Tunstall-Peddo is the founder and CEO of Unlikely AI, a British AI startup, and was previously the creator of Amazon Alexa. William, welcome to PathFounders.

1:15William Tunstall-Pedoe:Thank you very much, Mike. Good to be here. Well, right at the outset, your report estimates that this£29 billion productivity loss is obviously huge. First of all, but what we need to know is how did you arrive at that number? So it's based on the gap between the potential of what AI could be and where it is due to trust issues. So what we identified is that there's a ceiling, there's a trust ceiling that comes from the fact that AI solutions aren't trustworthy enough to be fully adopted. So there's very, very high levels of adoption of AI. Pretty much every business that we interviewed is using AI or trying to use AI or doing projects.

1:59It's true of pretty much every business in the sample.

2:04William Tunstall-Pedoe:But, you know, the majority of the senior leaders that we interviewed don't fully trust AI. only 22 % of those businesses are getting significant return on investment from their use of AI and the substantial trust ceiling. And we extrapolated to get that 29 billion figure from the value associated with that trust ceiling. If that trust ceiling wasn't there, there'll be 29 billion additional value to the UK economy. And is that sort of over a year or is it over a time period or just the opportunity gap? Yeah, that's per year. OK, so£29 billion per year being lost actually because of AI, oddly enough.

2:45I mean, I think we're all familiar with people using AI for fairly menial things like, you know, composing their LinkedIn posts, alas. But when it comes to deploying AI in an enterprise setting, it effectively sounds to me like what you're saying is that everybody who has to use AI now is almost like having to check the work of an intern and it's doubling their workload. Is that right?

3:11William Tunstall-Pedoe:Yeah, that's not quite true. There are definitely use cases where AI is being used and sometimes being used fully automatically. But there are many, many use cases, particularly in regulated industries, where it's just not reliable enough, where you either have to spend a lot of time checking it or it simply isn't good enough to be used at all. And if you're automating something, it's your brand. In regulated industries, is there are very strict requirements for the quality levels of what you're doing. And the technology, the current technology just isn't good enough in many, many cases to automate systems in these enterprises right now.

3:47OK. And one of the stats that you pulled out from this report was that 87 % said that they trusted AI, but 65 % still feel anxious using it. So there's actually, they're actually leaning into the tools. it sounds like, but they're still concerned about what its output. What's your sort of read on all of that?

4:11William Tunstall-Pedoe:Yeah, so I don't think it's fully trust. I think the majority of the people we serve is don't fully trust AI. So there's some level of trust, but the majority of the senior leaders don't fully trust it. And that trust gap is what's preventing the full opportunity to be grasped. Okay. What's going on under the hood here? Is it LLMs? We've kind of pushed LLMs to expect them to be sort of be all and be all of everything that we're applying them to. Yeah, I think that's right. I think people see this magical technology and it is magical. It does things that were previously completely impossible for computers to do.

4:54William Tunstall-Pedoe:you know over the last three years AI has largely conquered language which is like one of the huge huge barriers that AI had not previously fully conquered and that's unlocked all sorts of access to data that computers previously couldn't really do very much with it's it's opened up all sorts of other potential use cases but the technology doesn't work well enough it's it's incorrect some of the time it can't explain properly it hallucinates it occasionally produces things that are completely invented. And that creates a very substantial trust gap. And yes, so often, if you are going to use AI in these important business situations, you need to spend substantial amounts of time checking, checking what it's doing.

5:36William Tunstall-Pedoe:And that creates all sorts of cost, it may mean that the use case isn't viable at all. And it also creates some, you know, psychological burden of having to having to kind of, you know, spend that time checking and not being sure that the the outputs are correct well it is such an interesting phenomenon isn't it because you and i both have also been in living through the machine learning era which was a sort of a previous iteration of ai perhaps not generative though of course um and you famously um were you know uh came up with the original technology which formed the the base layer of amazon alexa and there where you said you know how high is the eiffel tower it would recognize that through machine learning and natural language process it recognizes those words and it has an answer for that you know an exacting answer for that but we're really we're very much through the looking glass with generative ai aren't we yeah i mean i think where you're where you might be going with this is different types of ai so so the statistical there's a statistical deep learning statistical machine learning world of which generative AI is very much an example.

6:49William Tunstall-Pedoe:And in that world, everything is less than perfect. So when you train an AI model, it gets better and better as you train it, but it always flattens off before 100%. So it's always wrong some of the time. Every paper published by academics on AI talks about a benchmark, talks about getting a new record on that benchmark, you know, and the record is 87.2%. And that's the glass half full version of it, the glass half empty version of it, because if it's only scoring 87.2 % of the benchmark, it's wrong 22.8 % of the time, which is the glass half empty way of looking at it. And for many, many applications of AI, you know, that's unacceptable.

7:31William Tunstall-Pedoe:There are applications where being wrong some of the time is okay. You know, you've got an improvement. there are applications where you have a human in the loop where the AI is doing something and the human is is you know catching those errors but if you're automating a really important business process if your brand's at stake if there's money at stake if if you're regulated and the regulator is going to you know cause you all sorts of problems when these errors occur then it's often simply not good enough and similarly these statistical techniques you can't really explain So, you know, a deep learning model you can think of as a formula with, you know, hundreds of billions of weights in it, hundreds of billions of numbers.

8:10William Tunstall-Pedoe:So a formula with hundreds of billions of numbers in is probably the most unexplainable thing imaginable. You know, the researchers who build them don't even understand how they work. There's an entire academic area of study trying to figure out how deep learning works, how generative AI works. So it's a technology that's doing remarkable things, but it isn't reliable enough, isn't accurate enough, and therefore isn't trustworthy enough at the moment for many of these applications to be adopted. And there's also lots of there's lots of examples of of people creating POCs, that sort of proof of concepts.

8:49William Tunstall-Pedoe:So they trial the software. It appears to work. It does a few magical things. But then those those tests, those experiments never make it into production, never actually fully automate the business because it's then discovered that they don't work well enough for these leaders to take the leap and automate the business with it. So what's your thinking about how we're going to kind of cross this chasm? Yeah, so this is the problem that my company is working on, Unlikely AI. We're trying to solve that with radical new technology. And the broad class of technology that we're pioneering is called Neurosymbolic.

9:28William Tunstall-Pedoe:so neuro means the statistical machine learning software that i was talking about that's that's you know able to do new things uh but is intrinsically unreliable uh and unexplainable and symbolic is shorthand for the other types of software that we use you know just like like the spreadsheet for example um you trust your spreadsheet uh you trust your spreadsheet to add up two cells correctly um it you know when you add two cells correctly up in a spreadsheet it doesn't work 87.2 % of the time, it works 100.000 % of the time. And if it ever added up two cells incorrectly, you would regard that as an extremely serious defect, you would sue Microsoft, or you would complain like mad about it.

10:11William Tunstall-Pedoe:But with machine learning, with generative AI, we accept, or we're exposed to a level of error. So what Neurosymbolic aims to do, what Unlikely AI is is blending those two types of software together. And the aim is to get the capabilities from generative AI, the ability to understand natural language, to process complex data, but combine it with the explainability and the extremely high accuracy you get with other types of software. That's an extremely hard thing to do, but it's something that we're pioneering, something that we've succeeded in doing, something that we're now commercializing. Right.

10:48Well, admittedly, I should have pointed out that this is research that you commissioned, obviously it is to some extent backing up your argument for your company, but to be fair, it was commissioned from Censuswide Research and it was a survey of a thousand business decision makers. So that's statistically relevant, it sounds like to me.

11:09William Tunstall-Pedoe:Censuswide did the research. So we commissioned it, but it's independent research done to a very high standard and all the methodology was published and open. Yes, yes, yes, absolutely fair comment. But also, I mean, if it's the case that Neurosymbolic ought to be the way forward to combine the power of generative AI with the accuracy of previous more maybe machine learning approaches or statistical based approaches, why is it that we're not seeing more of that approach? do you think? It's very difficult is the answer. I'm making it sound simple, just let's blend these two types of software together and get the best of both worlds.

11:55William Tunstall-Pedoe:It's super hard. And I think the other thing that's going on is the big AI labs that are pioneered generative AI. They're packed with scientists, but these scientists think they're purely in the neuro world. So their entire world is in the machine learning world. That's the way they think. Everything they're doing is trying to improve these models. And because this is an intrinsic property of the technology, it's not something that's at all easy to solve. So, you know, the industry is very, very aware of these problems, but it isn't something that can really be solved by iterating that technology.

12:28William Tunstall-Pedoe:It requires a sort of different approach, a different way of thinking. And that's where we come in. Okay. So obviously you're taking this neurosymbolic approach. Do you have any commentary to make on the the approaches being taken by the likes of Jan LeCun for instance who's left Meta to work on world models um would you characterize this as still being in the neuro neuro world as opposed to symbolic or not really approaching a combination of the two yeah no 100 that's in the neuro world it's doing something different it's a sort of world model company it's it's it's got a different approach it's got a different architecture for the neuro but it's still completely newer.

13:06William Tunstall-Pedoe:Yes. OK. Right. Well, you know, we know that unlikely AI is raised in the region of 20 million dollars to date. I was checking again before we recorded this and you haven't done any other rounds. Are you sitting on a gold mine and nobody's waking up to this issue? So, yeah, we of course, we always have conversations with investors, but that doesn't mean I can disclose them. But yes, hopefully we'll have some news to announce in the near future. But yes, obviously there's a potential goldmine for what we're doing. It's an astronomical opportunity. I mean, the big AI lab, big AI companies have literally invested over a trillion dollars in AI infrastructure very recently.

13:53William Tunstall-Pedoe:And that trillion dollars of AI infrastructure is in anticipation of very large scale adoption of AI. and some of that adoption will happen even with these trust issues. But with the trust issues solved, that adoption will be very significantly greater. So, you know, what we're unlocking, not just for the UK economy, but for everywhere, is much more adoption of AI because it's much more useful when it's trusted, when it's accurate. So it's an absolute big opportunity, yeah. A lot of that build out, as you said, we're talking trillions now, is into compute, GPUs, energy, effectively. Is there any kind of like, can you show us, open the curtain a little bit for us to see behind the newer symbolic approach?

14:47And would that, is it going to require the same level of energy in computing GPUs?

14:52William Tunstall-Pedoe:Yeah, so that's a great point. So the symbolic part of Neurosymbolic runs on CPUs, not GPUs. And that's orders of magnitude less energy than GPUs. But it does still use the GPU. So it's not a replacement. It's not going to unwind the need for these data centers. But there's definitely a very strong green dimension to it and a cost dimension to it and a lower energy requirement dimension to the technology as well. Well, that's a very tantalizing look into what might be coming down the pipe. Just finally, sort of closing this discussion out for now, you know, how fast do you think we can close the gap on, you know, the productivity gap?

15:43or do you think a lot of companies are going to be banging their heads against the wall trying to make a lot of their AI solutions work for quite a bit longer?

15:53William Tunstall-Pedoe:Yeah, I mean, we hope to roll out our solution. You know, we already have customers. We've got a very good relationship with some big customers, and we hope to roll that out at scale, you know, over the next year or so. But yes, I mean, it's a big, every single business is trying to adopt AI. you know, that's every single business across the planet. And, you know, there's an astronomical amount of activity trying to do this. It will inevitably take some time before the full benefits are here. But they are happening for sure. Well, I'm sure we'll be returning to this subject in the near future.

16:27But for now, thanks so much, William Tunstall-Peddo there, who's the founder and CEO of Unlikely AI. Thanks very much for joining Pathfounders.

16:41We'll be right back.