Oct. 2, 2026

Tiffany Willcox & Graeme Cox: The Gap Between AI Spend and AI Return

Tiffany Willcox & Graeme Cox: The Gap Between AI Spend and AI Return
Tiffany Willcox & Graeme Cox: The Gap Between AI Spend and AI Return
AI Pathfinder for Private Equity Podcast
Tiffany Willcox & Graeme Cox: The Gap Between AI Spend and AI Return

AI Trends: Closing the Value Gap with Attercop | Part 1 of 3

Billions have gone into AI experiments, yet for most businesses the return hasn't shown up in the P&L.

This is the first of three conversations exploring why that gap exists and what it takes to close it. I'm joined by Tiffany Willcox and Graeme Cox from Attercop, an AI consultancy that works with PE-backed businesses, who share the lessons and frameworks from their work.

Key Topics

  • The widening gap between AI spending and returns
  • Importance of data foundations for AI success
  • Challenges in scaling AI from pilot to production
  • Business transformation required for AI value
  • The concept of AI boost for practical implementation
  • Private equity's role in AI adoption
  • Technical and strategic hurdles in AI projects
  • Measuring AI impact on business metrics

Takeaways

  • Most enterprise AI pilots deliver little to no measurable impact without proper infrastructure.
  • Closing the gap requires focusing on data quality, business value, and scaling processes.
  • AI initiatives often fail due to poor planning, data issues, and lack of business transformation.
  • Private equity firms can leverage AI as a repeatable value creation tool.

Links

Tiffany Willcox
Graeme Cox

Attercop website

AI Pathfinder Private Equity AI Strategy Meet-ups:

https://www.aipathfinder.co/private-equity-ai-events

Have a question or would like to suggest improvements, please contact steve@aipathfinder.co

SPEAKER_00

Hello and welcome to the AI Pathfinder for Private Equity Podcast. AI Pathfinder helps private equity firms make sense of AI and make it work. It's an expert network built on insight, experience, and connection, bringing the right people and ideas together to turn AI from something firms are curious about into something that delivers real results. If you'd like to attend one of our regular AI strategy briefings in London and New York, please check out the show notes for details. Today's episode is the first in a special three-part series I'm running with Atacop, an independent AI consultancy based on the south coast of England that works with businesses and their investors to turn AI ideas into commercial results. Across these three conversations, we're going to dig into what actually works and what doesn't when firms try to move AI from experiments into daily use. The question every investor and portfolio leader I speak to is grappling with right now. In this first episode, we're focused on a specific trend, the widening gap between what businesses are spending on AI and what they're actually getting back. We've all seen the headline start from MIT's State of AI and Business report last year, that roughly 95% of enterprise generative AI pilots deliver little to no measurable P ⁇ L impact. Not much has changed with growing frustration amongst our group of real ROI. In our own AI Pathfinder benchmark survey in June, found that only 40% of PE respondents agreed that their data foundations are good enough for scaled AI adoption. Six in ten know the groundwork is not there yet, so there's a real disconnect between ambition and infrastructure, and closing it is arguably the defining challenge right now. I'm joined by Tiffany Wilcox, co-founder and chief commercial officer, and Graham Cox, founder and chief executive. Tiffany leads the commercial side, Graham, the data science and delivery at Atacop. Welcome to you both.

SPEAKER_01

Hi Steve.

SPEAKER_00

Lovely to be here. Tiffany Graham, talk me through what you're actually seeing in the field. Where is the gap between AI investment and AI return and why has closing it become such an urgent issue in 2026?

SPEAKER_02

Well, so Steve, the spend has been enormous, but it's mostly been experimental. For most businesses, the return's not really shown up on the PL, and the that that that gap is the story of 2026. What's changed, I guess, is patience. Boards and investors have run out of the appetite for an open-ended AI budget, and they now want to see the line on the PL and not just another pilot. Closing is uh closing that gap is what we've been working on, and and taking an idea or a stalled pilot and delivering value back into the team, in particular the finance team, that they can actually see.

SPEAKER_01

So, from a technical perspective, what I see the biggest problems in AI initiatives are that they are either poorly thought through in terms of actually understanding what the return on investment is likely to be. So frequently taking the route of going there, of looking at what appears to be the sexy option rather than the one that's actually going to deliver back to the business, and or not really thinking through the data requirements before getting started. I think the latter is particularly an issue in what is still this very new world of generative AI. You know, I come from three decades of working with artificial intelligence and machine learning. And the single most fundamental tenet of my career has been that uh garbage in delivers garbage back out again. And actually, even in this world of generative AI, where AI is good at handling unstructured data documents in a SharePoint file share or whatever, that remains true. And for so many organizations and use cases, the knowledge engineering that's required to get the best out of AI is dull. It doesn't deliver return on investment, and it doesn't feel like the thing that you should do in order to get the product out there. But it is the thing that stops you from getting value back out of your investment. Again, that is exacerbated in generative AI by the fact that small scale demos, which are very easy to produce with Gen AI, and we see non-technical staff across PE and Portcodes everywhere creating these wonderful little apps and use cases. They work great for small amounts of data, because over small amounts of data, by which I mean less than a few hundred documents, you can deliver very, very good results. As that scales, AI's ability to actually find the nuggets of gold amongst the weeds becomes harder and harder. And the return you get in the delivery of that app in terms of its accuracy, the completeness of its data, and therefore the utility to the business drops dramatically. So, in summary, companies need to plan what it is they're doing and why they're doing it, not just because it looks great, but because it actually delivers a return for the business. And working at scale and working in production still remains much, much harder than delivering very cool demos.

SPEAKER_02

Yeah, the value the value is lost in that last stretch, I think, because the models are great, as you know. You know, the models are there there's plenty of them and there's plenty of of the technology available, but it's that last stretch and getting the use out of it and the value and being able to articulate that value back to show that that there's benefit for using it.

SPEAKER_00

Yeah, the um democratized sort of nature of of this technology is is is been brilliant, but it's then focused everything on on that on the technology. Uh, and that's not necessarily really where the problems lie, is it? Or where we should start, for sure.

SPEAKER_01

It it's absolutely not where you should start. Uh business value is always where you should start. And the business value of many generative AI solutions is not obvious. I think, you know, just riffing on that a little bit, that I see a lot of work on the ground where individuals and teams are building small-scale automations in terms of how they produce their documents for work or how they process uh data out of the CRM or how they prepare their marketing materials or whatever else. And it works brilliantly as a way to speed up an individual. But going from that personal accelerator, AI as a tool, to agentic AI, where AI is working as a peer within an organization and is actually executing processes and making decisions is a massive leap and not one that can be done without significant planning and significant business transformation as well in many cases. It is not possible to automate a human, but it is possible to automate some of the things that a human does. To be able to extract those processes and make them automatable is not just the technical challenge that I spoke about earlier in terms of data quality and presentation and governance and control, but it is also about the transformation that is required to extract that process and make it self-operating, which requires the entire business around that process to change to work. So AI for value inside a business is most often more about business transformation and understanding where ROI can be found than it is about the underlying technology.

SPEAKER_00

Yeah. Yeah. Um hearing that loud and clear. Well, this is a gap you're you're aiming to close. And you've mentioned that you even put a name to it, AI Boost, I believe. What does that actually mean then? And why does it even need a name?

SPEAKER_02

I mean, it's it's something that we've lovingly called AI Boost for for exactly that reason. So we, as we've just articulated, it's moving from a promising idea or a proof of concept to something in daily use that changes a number of business, a number that the business cares about. So we've we've seen and we work with some extremely knowledgeable teams and often with an AI team who know what they're doing, but they just need extra firepower. So approaching it from an AtiCop perspective, we are able to help the teams identify the use cases that they're working on and the challenges that they have and even supplement or support them in creating the framework in order to move that idea from concept into reality. I think there are no shortages of ideas here and there are no shortages of what people want to do, but the experience of a data science and how you actually implement these things into production and get them working is is still a is still something people need to learn. And we're fortunate enough that we do AI very well every day. Um and we're able to boost other teams with that knowledge and support them to get over the over the that hump for want of a better word. Graham?

SPEAKER_01

Yeah, I think most of the AI boost work that we've done has been with software teams. So in a lot of SaaS businesses, for example, and other teams that have significant software engineering capability, which might seem surprising on the face of it. But there are there are two things to touch on here. The first is that we are being brought in by the private equity investors owners, and we're and that is being done specifically in order to accelerate the AI initiatives and capability within the portfolio company in order to deliver, in order to create value in that business, uh to keep up with the competition or to generate new revenue generating capabilities or revenue retention capabilities from the underlying data that exists in that business. The reason why we're brought in plays back to my previous comment around the difference between demo and production. So every competent software engineer in the world pretty much now uses AI on a daily basis. If I ever find a software engineer who doesn't, I would be very, very surprised. They use it at very least in order to help generate code to support them in the process of software engineering as a minimum. Most software engineers have put some effort into developing at least proof of concepts and tooling that uses generative AI because it's so simple. You just call an API, send it some data, and the generative AI comes back with a result. Why would you not be doing that? But again, as I said before, the difference between being able to do that as a demo or an individual tool and doing that in production for customer-facing features that are revenue generating or revenue supporting is a big, big difference. It relies on implementing something which is fundamentally alien to the nature of a software engineer. And obviously your listeners are going to be very familiar with this, Steve, because it's something that we all bang on about all the time. But AI is a non-deterministic medium. Software is a deterministic medium. Software, if if a software feature doesn't work, it's a bug. If a if an AI delivers a hallucination or an incorrect output, it's not a bug. It's literally a feature of the AI. Because AI is probistic. It is stochastic in its nature. It never gets everything right. That is both its biggest strength and its biggest weakness. And understanding how you can harness probabilistic models in a deterministic software stack requires a lot of understanding of the customer and where that customer value is, what they actually will accept and what they don't, and how to correlate that outputs with the expectations of the user. So two things that drive us in doing this work into otherwise extraordinarily competent software development teams. The first is that we understand where value can and cannot really be achieved in using machine learning models or generative AI in software stacks. So what will actually deliver back for the business. And the second is that we understand in detail how to harness these things and deliver at scale, how to get the data right, how to do the knowledge engineering, how to harness models, how to build evaluations that will actually allow you to understand and control the behavior of the model. These are the things that are not easy, and these are the things that ultimately add to the top line of the business.

SPEAKER_00

I just wonder if you could just walk through as a I'll say detail, but briefly the steps, I suppose, what I'm asking for here is in terms of if I run a business, we've got three AI ideas on the table. What should happen next? Shall I take that?

SPEAKER_01

So it's a great question. Um so the first thing is understanding what the business impact is that you're trying to uh achieve and how to build an ROI model around that. That is that is um particularly when AI a relatively new capability within an organization, it is a big a big thing and needs to be considered carefully. So again, if we look at most of the AI boost programs that we have undertaken, the the software engineering teams and CTOs that we are working with, they're they're great people, really skilled, understand their domain extremely well, but they're not AI engineers. They don't have the detailed knowledge and understanding of AI. And because it is easy to do something with generative AI, it is easy to fall into a way of feeling of it's just another bit of tech, it's all just fine, we'll just, you know, we'll read the manuals, we'll play with it, it'll all be okay. That is that is a poor way to understand what the investment is actually required in order to get sustained value out of AI. How do you how do you build effectively? How do you build for performance and scale? How do you understand the uh the token costs and the other back-end costs that come with use of AI that are not present in normal software engineering? How do you manage model performance and model hallucination? How do you operate and run models within within businesses? And here I'm talking about not just generative AI, but also machine learning models for prediction or forecasting, things that allow you to extract value from the underlying data set that exists within the organization. Understanding the true cost of those new skills and what that investment is required is part of the equation. The other part of the equation is actually being able to put a reasonable model around how you are going to derive value from the delivery of these features to your customers. Can you actually realistically charge more money for introducing AI into your stack? Or is AI now just a matter of keeping up with the Joneses, which it is for a lot of businesses? You don't do it, you get left behind, you you are seen as destroying value in the business and risking your customer retention in the process. If that is the case, how do you manage the additional cost of both the people cost, the token cost, the operating cost of binding AI to a system and understand how the value is delivered back into your organization? So it is both a technology discussion around how to do it right and what the real cost of that is, and a commercial discussion around how do you actually justify that and how does that deliver value back to your customers and ultimately to your stakeholders as well.

SPEAKER_02

And Steve, I'd I'd add as well, you know, I I don't think this is a new, this is a new concept. I think people are perhaps a little bit AI has been a little bit romanticized in the fact that it might solve all your problems and isn't it great? But really, you wouldn't put in a new bit of technology without understanding what the impact to your business is, and certainly from a cost perspective, how much more it's going to cost or whether there's indeed a cost saving. So, you know, AI is no different in that perspective, but the the crossover between tech and commercial is actually far more commercial than it is technology. So, you know, validating those ideas and and truly testing the feasibility of them are really worth looking at. And more often than not, we're we're helping people to identify that these are these are the one or two areas that you can definitely add value from. And the rest can rate, can wait, because I think people often go for the big, bright, shiny response, but that cause often causes the most problems further on in the down the line. So it's a real, it's really feasibility, value, making sure that you understand what that is. And and at the end of the day, what's that metric that it's going, how's it going to change? And the answer sometimes might be that it's not.

SPEAKER_00

As you know with AR Pathfinder, we'd like to hear some real real examples. And I wonder if you could take a real piece of work, obviously not naming names, but that that captures what we've with what you've been talking about here.

SPEAKER_02

Yes, that's a good idea. So we've done a load of them recently. Uh Graham, so if we do a recent technology business in the hospitality sector, shall we say? Um and sitting, they were sitting on a load of operational data, which they just really weren't using. They had an idea to put the data to work and it installed at the proof of concept stage. And that and that was sort of where we then got brought in.

SPEAKER_01

Yes, absolutely. So we have we have so many examples here. Uh I'll come back to the specific in literally just a second, but just to set the scene here. The uh the the uh again, so with with this work, we've we've done quite a lot of work on product software products that are the core revenue generator for a business. So a lot of SaaS businesses, a lot of businesses for whom the underlying um uh um software is is a core part of the service that is being delivered, managed service, professional service, whatever it might be. What there are two key areas that we um generally can make a big difference. And in this uh in these examples, the first is uh this is going to play back to everything I've been saying already. The first is being able to exploit latent, untapped data uh inside the organization to generate new capabilities, functionality for the customer, and revenue generating opportunities. And the second is being able to build an AI-enabled stack into their software that will work at production scale with high reliability, strong governance, good observability so that you can trust the AI that you're building into the organization. So, with the the organization that that Tiff is referring to, that that is uh an example of the that exploitation of latent data. So fundamentally, this business provides services to its customers that allow them to manage their own data, which was very, very helpful to uh building a customer base when they started some 18, 19 years ago, uh, allowing the customer to basically build their own taxonomy, their own definition of what their data meant so that it made sense to them, but ultimately led to them not being able to scale the bit the back-end data of the business. They couldn't compare one customer behavior to another customer behavior because the data sets became so different and disparate. So they'd ended up with this really successful business serving organizations across Europe, really well scaled, really successful, loved by its customer base, but unable to exploit this huge data set inside the organization because of this lack of data engineering that had been planned in a pre-AI world. Um so in a just a few weeks, we went in and gave them a path to engineering that data and built machine learning models that would help them understand how they could build forecasting and prediction models from that data going forwards that would uh that would give them a whole new capability for their customers. It is going to then require them, now we have finished that work, it requires them to actually start to do that hard work to build up the the data set that they need from that foundational information over time to scale it to the point that they can deliver that value. But what we left them with was a an ROI model, evidence of how it could be done, a model training approach and a data engineering approach that they could follow, and hands-on training on how to how to do all of that. Now, obviously, they can come back and call on us to um to continue that work, and we are indeed supporting them in uh in the next phase. But the primary goal there is both to give them the capability, sorry, the to give them the the the the the the business and technical model of how value can be achieved, and to give them and to teach them how to fish, not to give them the fish, to actually put them on a road where they can start to learn and understand what taking on AI and machine learning means, how they build that into their company, how they model the costs of that in the back end, and how they derive the value from it over time as they bring it to production.

SPEAKER_00

Let's bring it back to the private equity um firm now. Why should this approach matter to them specifically?

SPEAKER_02

Yeah, so uh I I'll go and then Graham, please do chime in. But AI is now a repeatable value creation lever, right? You can run it across your portfolio. Um the way you did with you the way you would do with, say, pricing or procurement, rather than a bespoke gamble in each company. We know that it it does work. We know that when implemented correctly, that it really does start to drive value. As Graham said earlier, either within your own operations or a customer. Facing externally, it it's AI will show up in the metrics that that investors already track. So your margin, your revenue growth, your capital retention, um, in time, the next exit narrative will almost certainly, in fact, without a doubt, have AI as a as a narrative. So I guess having it in in there, it de-risks the AI line instead of the open-ended spend. When we come in and we help them deliver these solutions, we're actually making them tangible examples of what's worked, where they've moved the bar, the target, the timeline, um, and and they've and and it's become something that they can use rather than a concept, which it has been for so many people. And from a diligence and a value creation planning perspective, it gives you a way to read where the company really is. And, you know, the the boost is something we've spoken about. Well, one of the other very exciting projects that we work on and that our team love is actually the replication exercises. So, you know, more often than not, people are going, well, is this business going to survive in the next year? Is AI gonna just totally wipe it out? And so it's really, you know, again, from a value perspective, how we go, we are able to help them answer those very specific questions. And it and it's, I know it sounds so simple, but there's so much work that goes into actually doing that piece of work. But anyway, coming back to your question from an investment perspective, I think it sits so intrinsically within all of your value metrics that it's something that you have to take quite seriously, which is why what we do helps move the bar. Graham, is there anything more you'd add?

SPEAKER_01

Yeah, just very briefly. So, you know, so we're we're I think it's a really interesting model that we've been working on that we're actually going to be talking about with the private equity company concerned at at your next event, Steve, around how this AI boost has been executed driven by the private equity company rather than just as a uh a soft introduction to the portco and suggesting, oh, you might like to talk to some people who know what they're doing. Where the the PE company had taken a view that the portfolio needed to up its game broadly in its execution of AI across the portfolio. And we actually agreed uh contractual terms and a master services agreement directly with the with the PE firm, uh, allowing the PE to essentially foist us on a number of their portfolio companies and say, there you go, these guys are going to take your most challenging AI problem and they're going to show you how to fix it in the space of a month. Get on with it. Now that is and that is an introduction that could go horribly wrong for a lot of companies because a lot of portcos have just balk at that. And to be clear, the private company in uh in in question here does it its majority or full owner of all of its portcos. It's not a minority investor, so it does have control here. But nevertheless, it it the the ability for us to execute on that and make it a success is dependent on a our our our focus on what really matters and value creation, and therefore our ability to engage directly with the CEO as a starting point to help them understand why this almost enforced spend is actually of great value to them. And the second is that we are we are humble, uh we're not your kind of stereotypical consultants, and we are focused on delivery. You know, we're fundamentally AI experts, AI geeks, if you like. We love our subject matter, is what we live and breathe, it's what we do, and we're and we're good at that. We don't pretend to be good at everything else. We work really, really well hand in hand with CTOs, technical teams, IT managers, CIOs, where whatever the environment is, to help them get something done that is a challenge for them. No ego, not a focus on slide deck, much more a focus on practical delivery capability and ROI, uh, and a laser focus on just being good at our core subject. And that's what makes it work.

SPEAKER_00

We're coming up to time, and I just wonder if we could just sort of almost summarize really what we talked about here. If a listener wants to act on this next week, what are those first steps?

SPEAKER_02

Okay, and then Graeme, you add, so I would say pick a number that you want to move and one use case that sits behind it. Don't start with the technology, start with the change that you're trying to make. Be honest about your data. There's the other thing. I mean, Graham said it early on, like garbage in, garbage out. I've been researching the concept of knowledge quality recently, which I'm I'm I'm it's just blown my brain in terms of the next level of quality and how people are approaching it and managing it. But the fundamental is that your data, you know, most of the effort lies within curating your data. So find out how ready it is, look at it and really dig into it. And then lastly, there is no shame in getting help. So scope it out properly before you spend it. The expense of mistakes almost always happen at the start. Um, and as we know, AI will confidently tell you the very wrong answer. So the more you can spend time making sure that the data it's referring to is in the right space, in the right format, in the right manage being managed in the right way, honestly, that's the best thing you can do for your business before going in and implementing, would be what I would say. Graham, what would you say?

SPEAKER_01

Yeah, I just echo echo that. Um, there has to be a clear focus on value creation. I mean, you know, I say that, and then there are there is some level of AI initiative that is pretty much mandatory for almost every business. You know, there is a matter of just get involved, because if you don't, you're just gonna get left behind. But broadly, for uh for significant AI investment, there has to be a return. Being really clear about what that return is and how that is measured is is really important. And being realistic about it up front, indeed, around areas like can we really charge for this? How will this actually uh, you know, where is the benefit really a bit dark ridgeable? Uh, what line does it really land on? What is the real budget required, not just in year one as a project to make it work, but then in year two and three in order to operate it? All of the elements that go into understanding, you know, a mature model for AI adoption, that is the starting point. Technology is very much the the secondary concern, but it is nevertheless, it's a technology implementation. We are now, even though we're only three and a half years into the generative AI revolution, we are many decades into understanding how AI and machine learning models really work. And the engineering patterns to derive benefit and to control AI in production are mature, well-understood patterns. They are just different from traditional software engineering. So it is not just a bespoke, it's not just a a journey into the unknown. There are people out there that that can guide, that understand the well-trodden paths, that can help organizations learn what it means to operate models and cap AI capabilities in production in in revenue-generating environments. And either acquire that knowledge, hire that knowledge, or invest in that knowledge before you get started with significant AI initiatives.

SPEAKER_00

Tiffany Graham, thank you both. Really useful to hear how you're thinking about closing the gap on AI spend and AI return. Really appreciate your time. Been our pleasure, Steve. Thank you very much.

SPEAKER_02

Thanks, Steve.

SPEAKER_00

Well, that was the first in a three-part series with Atacop. We'll be back over the coming weeks with the concluding two episodes. If you'd like to find more on Atacop, we'll include the links in the show notes. And if you're navigating this territory yourself and would like to be part of the private equity AI strategy network and events across London in New York, please get in touch. That's it for now. Thanks for listening, and we'll see you next time.