AI Is Producing More Code, but Is It Producing More Value?

Speaker A: Foreign. So we're back with my friend Evangelist Moody's of Synapse Partners. His blog is at Corporate Innovation Co. We are recording this on July 10, 2026 morning off episode notes, where I will place links to all the things we talk about is on the Data Exchange Media. Hit the subscribe button if you're on YouTube and subscribe wherever you get your podcast. Topic number one, the rise of open models and our belief that open models will ultimately prevail. Evangelos, you start.

Speaker B: Um, a lot of this is coming out from the research that uh, I've been doing on legacy enterprises and the use of AI by them, as well as the work that our firm has been doing with, uh, such corporations and I would say enterprises have embarked on a journey for AI. Uh, most of them, if not all of them, are starting from frontier models. But very quickly their, uh, CFOs are starting to recognize the uh, very high costs that are associated with such models, uh, and also what it means for their own data. Um, they are, as part of this journey, uh, they're starting now to ask questions and in some cases already started utilizing, uh, models that they, A, they can modify with their data and B, models that will allow them to, to have full corporate sovereignty for lack of a better term, uh, as well as better control their costs. And um, obviously certain types of, uh, open models, uh, will allow them to do that. Uh, frankly, I also believe that uh, over time, and this is what we are advising these corporations, they should use such models as part of their journey. But the ultimately they will have to build their own models because they don't have to be big. In many cases, we're starting to show them that they don't need a massive model in order to be able to accomplish certain of their tasks using AI. Uh, but that's a lesson they have to learn over time.

Speaker A: All right, so, uh, I'm just going to uh, play uh, devil's advocate there. So first of all, doesn't this contradict what you keep saying on this podcast, Evangelos, that uh, uh, enterprises actually don't have their act together. So why in the world will they use open models? They should start with, uh, proprietary endpoints because they're beginners.

Speaker B: Right? And they do. Uh, and frankly, if you look at what are the two, uh, requirements that this kind of uh, entail is, number one, they need to start getting the right people. Uh, and the right people doesn't mean only AI people. They need to have the culture and the people that will allow them to embark on this journey and bring it to uh, completion. And number two, which is actually something that the enterprise has been going outside of it is they need to start bringing again their IT in house. Because if you're going to start running these models and have what I call before, what I termed before a corporate sovereignty, uh, you cannot be doing all this in the cloud. Right. The hyperscalers will love you to do that. Uh, and obviously that would be a step. But the way I see it evolving over time is we're going from full cloud, which is what we have today with a proprietary uh, frontier or specialist models, to a hybrid uh, deployment where some of the models may start running within the enterprise and others will continue running in the cloud, whether it is new clouds or hyperscalers, whatever. Uh, and then ultimately uh, the more capable of these enterprises will need to bring these models in house and run them in their own data centers. And in only in this way they will be able to have what I've called in some of my write ups, their AI factories which they will control and they will be able to build that from that expertise.

Speaker A: All right, uh, I'll uh, come with my first reason, which is not really a reason, it's more of a, of uh. In many ways it sets the stage for uh, why we're making this observation and that is the fact that uh, um, open models have gotten real in many ways. Uh, I think uh, if we were to have had this conversation a year ago we would say the gap between the open weights models and uh, uh, frontier models from the frontier labs would be on the order of 12 months. Now it's really more in the order of maybe, I don't know, three to six months. So that's why I think uh, and you know, I mean you can argue that well maybe that's just a gap in the benchmarks because these Chinese uh, uh, uh, open waste format providers are benchmaxing and they're doing distillation. But at the end of the day uh, the benchmarks just simply are a way for you to attract attention. But people will have to use it based on at least anecdotal uh, information including uh, my usage and the usage of many people. I know I uh, would say that by the open weights models of God and real, you can realistically offload uh, I don't know how to quantify it but maybe 80% of the time you can use these open waste models and then reserve uh, the frontier models from the frontier labs for the remaining 20%. It's more of uh, uh, the reality of where these open weights models are today is uh, one of the reasons.

Speaker B: So let me add to what you're saying. Um, the first point is we get a lot of questions, uh, because everybody realizes that the majority of the open models today come from China. Uh, and I think this is in my opinion it's a very interesting and a very deliberate and important uh, strategy that China is employing.

Speaker A: With regards to the two exceptions are Gemma from Google, which, which is smaller and maybe more optimized for edge devices, but for a lot of routine enterprise tasks it should work if you tune it. And then the other exception is Nemotron from.

Speaker B: Yeah, so I was going to get to that, but I was going to make a bigger point. So we do get a lot of questions about the risks of running uh, models that have Chinese provenance. Um, but the second thing with regards to what you're saying on the capabilities of these models and how they may apply to a lot of the tasks that the enterprise wants to uh, use to test AI and even deploy AI, is they're starting finally to be a realization after two and a half, three years now that we have generative AI exploding into the scene, there's starting to be a realization of how to match tasks requirements with a model's capabilities. And I uh, think again, uh, the more uh, deliberate of these uh, enterprises, the ones that really want to continue on this AI journey, uh, they're really spending the time to understand how to not only automate uh, a specific uh, task or a specific function using AI, but they're starting to look more holistically into workflows and processes. And once you see that and you try to determine how can AI be applied in an entire process, what steps of the process can it automate, how can it automate them, what is the value of that automation and more importantly how much will it cost me if I were to deploy an AI driven process? Then you start becoming much more aware of the importance of open models of what they can give you and what they cannot give you. Uh, and that's why I call it as an intermediate step towards this full corporate sovereignty.

Speaker A: Yeah, yeah, yeah. So the uh, realistically what you're looking uh, at is a hybrid stack, right? So where you probably use open waste models for a lot of what you do and then uh, for the remaining 20%, the ones that really require some deep reasoning or something, you use the uh, models from the Frontier Labs. By the way, for our listeners, when I say Frontier Labs, that's usually a shortcut for uh, anthropic, OpenAI and DeepMind. So do you have another reason or should I close up with my last reason?

Speaker B: Uh, I will um. Another reason for using them.

Speaker A: You mean another reason why open AI open waste models will uh, will uh, prevail and prevail meaning they will eat a lot of the workload.

Speaker B: Uh look I want to keep emphasizing these two points, cost and corporate sovereignty. Uh to me these are the paramount um, I mean the cost of using the proprietary models and the need for a corporation to keep its IP which for the purposes of AI is really in the data that the corporation has and how it uses that data in the processes that it employs.

Speaker A: So so for me that my last reason is uh, to own your own AI has to do with uh, tools and infrastructure and, and uh, what I'm about to describe is not quite there yet but will be there I would say. I don't know, I don't want to know, I don't want to give a time frame but maybe 12 to 24 months. So by this I mean um, so now we have the tools for, we have somewhat advanced teams doing pre training of specialized models. So Datadog is famously built a foundation model for observability data uh ranging from parameter count in the several hundred million to a few billion. Uh, I know of healthcare startups that have done the same for very specific uh workloads in the few hundred million to a couple of billion. There are advanced teams who are able to essentially pre train uh frontier models and I believe the tools for uh, allowing much more mainstream enterprises for doing that will come shortly. The reason I say that is uh, there's the infrastructure on the compute side. You've heard me talk about the park stack, right? So Pytorch, uh, uh, Ray Kubernetes but on the data management side now you have LANS file format for multimodal data. Uh so managing pipelines is going to be much more convenient for enterprises. Uh, that's free training. So then on the customizing. So we all know that supervised fine tuning is basically a solved problem now I mean I uh think uh, all you have to do is come up with your label data sets of examples of prompts and decide response. So the next level is reinforcement fine tuning. And uh, I have a post coming out I think uh, uh next week on Tuesday which I describe uh, many, many startups that focus on reinforcement learning. Many of them uh are building tools that uh, are specifically aimed at helping companies do this step of reinforcement fine tuning. And now so I believe that uh, the tooling would get there for doing both pre training and post training, not of the massive models, but of the more targeted, specialized models. And, uh, the reason more companies will do this is for the reasons that Evangelos described. Uh, but the other reason is really owning your own intelligence, specifically owning that compounding loop. Meaning once you deploy these models in production, you can really observe usage, uh, and, uh, uh, areas where the model needs improvement. You can get feedback from your users, and you can get that loop going of improving the model over time. And you own that. Obviously, a lot of what I described right now, uh, the tools are still in the initial stages, but as they say, the writing is in the wall that this will happen. It's just a matter of time. Uh, the democratization about, uh, pre training, post training, and this, uh, ability to, uh, extract, uh, benefit from this compounding loop.

Speaker B: So, um, I have written about the fact that, uh, the corporations that are serious about AI, uh, some of them are discovering AI for the first time, some of them are rediscovering AI, some of them enhancing what they have already been doing. Um, they really need to have their own factories, and there's no other way. But now, um, the quirk that I, the turn that I put into this is in order for them to get to what I keep calling corporate AI sovereignty, they really need to engage more than the IT department in building and expanding and maintaining that factory. Right? So this is not only an issue of technology. You know, build me something. Yes. You guys are responsible for it because you're in the IT organization or because you're in the AI organization. Um, we've been working with a couple of foreign organizations. Uh, costs are becoming a really big deal. Uh, tokens, uh, the cost of token, the usage of tokens, how to account for tokens in budgets, uh, are becoming a big deal.

Speaker A: Uh, and people are getting more sophisticated in the fact that, uh, uh, it's not just about token costs, it's actually task completion.

Speaker B: Right, so, exactly.

Speaker A: For example, uh, uh, zebus, uh, glm 5.2 is actually quite inefficient in terms of tokens. Right?

Speaker B: So that involves. Or that requires the engagement of the finance organization. Uh, it requires, as we've been saying, bringing the right people, uh, into the organization. Not only the people who understand modeling and understand AI, but people who can even do the deployment and educate the end users, uh, on this. And that means HR becoming involved in this effort. And then obviously the process owners. Right, the process owners are the ones that will be able to, uh, not only help but uh, the AI experts to uh, do the right thing and bring the right uh, technology and the right modeling uh, into this process. But also they're the ones who uh, show that there is an ROI from all of these efforts because these are not quick efforts and they're not cheap efforts. So, but they are necessary efforts towards that uh, ultimate goal that we've been, you and I have been talking about.

Speaker A: So, so yeah, I guess uh, to just uh, quickly end this discussion. So like I said, so the uh, infrastructure for pre training, post training um, is coming. Um, and uh, I think what you just described there in terms of uh, uh financial management of uh AI usage also requires infrastructure frankly because attribution is a key here. So if you can't attribute uh, you can't really manage. Right. So. All right, so topic number two is um, I guess uh, one of the, one of the main success stories for generative AI and agents is around programming and coding. So I recently uh, was curious about this because I was in an event where many of the Frontier labs were there and uh, the people who built the uh, I'm not going to name them because it was all off the record discussion but the people who built all of your favorite AI coding tools were there and they were super bullish about uh, uh, the space. So then I decided actually you know I'm, I like these tools too but let me see if I can figure out uh, what the data from recent uh, uh studies say. Right? So and uh, I wrote a long post which is quite uh, comprehensive where I tried to lay out uh, the argument for uh, being Buddhist. And then I would balance it with a kind of uh, counter argument, kind of red teaming uh that argument. But basically uh, it seems like at a high level there's some sort of attenuation funnel happening. So uh, the intensity uh, uh is kind of decreasing as you think about this more. So there's no doubt that developers are writing more code, right? So pick your multiple, right? So I said 2x3x more code, right? So um, and by the way the sources of data for this uh, meta analysis roughly falls into two things, right? So one is surveys of developers and as you know evangelist developers tend to overstate uh, uh things right? Because it's self reported responses, uh, right. And then the other one is telemetry. So the attenuation funnel works as follows. So there's no doubt the developers are writing more code, maybe 2x3x more. But then the next step is how much uh new software is getting shipped that's not at the order of 2x or 3x, maybe it's in the 30%. Then if you follow it all the way down to let's say you look at the actual uh usage of the software. So you go to the Apple App Store or Google Play or any other software distribution channel, as far as we can tell uh, nothing has changed. So the amount of usage and so now one can argue that hey uh, maybe the users of the software are not humans, they're agents, right? So that I don't know. But as far as uh, if one of the goals of uh, uh using these tools for example is to write more apps for your mobile phone, that hasn't seemed to have moved the needle substantially. But there's no doubt that uh, we all are writing more code. And I think actually uh, to be fair I write a lot of code now with these coding tools that are throwaway code right? So that maybe I need to crawl a website or need to do some data munging. Just one and done. So obviously now if I were to measure how much more code I'm writing, yes it's probably 2x or 3x but all the way to shipping, that's not actually 2x or 3x, that's more in the order of uh 30% according to these studies, right? And then carried all the way to the usage of the end users. The needle hasn't moved. And in my analysis obviously uh, I have a lot more data around uh the bugs it introduces, security challenges and so forth. But one way to think of it is really this attenuation funnel, right? So 2x3x more code, 30% more shipped software but all the way down to the users of the software. Not much change.

Speaker B: So um, let me add my perspective if I could here uh, which is really as I said based on uh our firms engagements um, uh first of all uh, um there was a huge excitement about what this could mean to the, to the enterprise, uh and the backlog of projects and that excitement is being moderated. Uh there is a realization of the type of technical debt that uh, is being introduced and what it will mean down the line. There is the uh, issue as you started describing Ben in terms of how is this code, uh evaluate it and which of the code passes the evaluation criteria that the enterprise ah has established. Um at the end uh, what we are seeing is the more sophisticated at least of the enterprises we're working with um, are using these tools to a uh, do a lot of prototyping and understand what new applications could look like uh, and I will say, uh, again, the more sophisticated of these companies are smart enough in my opinion, to use these tools on new applications as opposed to going back and trying to modify or enhance existing applications which ingesting uh, a very large quantity of enterprise code, uh, before doing something. We're seeing some efforts to uh, use uh, this application, I mean these tools to um, uh, convert in a sense refresh uh, from an older language to a new language, the periphery of an older enterprise app, as opposed to going to the core of it. Um, but in general I think the excitement of what can be done, um, is being moderated. Um, Keep talking about the enterprise. I'm not talking about.

Speaker A: Yeah, so what do you think of that attenuation funnel? Does that make sense?

Speaker B: I think it makes sense. I uh, do not know. I mean actually uh, we have not done a quantitative analysis on the work that we have been involved in. Uh, but uh, I would say um, it makes sense.

Speaker A: And by the way, here's one intuition for you listeners. Yes. So 2x to 3x more code, but only 30% more shipped software. And in many ways that makes sense, uh, uh, intuitively because basically people still have to review this code.

Speaker B: Yeah, right.

Speaker A: So the processes for review may also be automated somewhat but uh, uh, people have to be careful. Right.

Speaker B: So uh, let me add two more points very quickly here Ben. The first one is that uh, we're starting to again on the use of the tools. We have been seeing maybe for the last eight months now, uh, startups that uh, are able to produce uh, version 0.5 of their product using these tools with a lot less people. And that's important from a capital efficiency perspective as far as startups are concerned. We're also uh, uh, so that was a revelation. The second point, which is not so much of a surprise is that companies that we will characterize as digital natives, ah, make a lot more efficient use of these tools than the uh, the legacy enterprises by and large. I mean this is a very broad statement, uh, in terms of how they use it. They tend to have much more modern code, more modern practices and that makes the use of these tools a little easier. But I'm willing to say that to your uh, funnel, I don't believe that we will see uh, significantly different um, results uh, between how uh, uh, a digital company gets value out of these tools versus a sophisticated legacy uh, corporation.

Speaker A: So, so a couple of things on that point. Right, so, so in the, in that long study that I published, right, so there was a DevOps, uh, research uh survey. Right. So uh, and uh, they did find that AI functions as an amplifier of your existing team. So meaning if you have mature internal platforms, clean workflows, really disciplined engineering practices, you extract uh, outsized benefits. Uh the other thing is that uh uh code based maturity also decides where AI helps or hurts. Right. So um, I guess let me kind of also tie this back to well then what should I do? Right, So I think you know, because the attenuation funnel starts with that 2x or 3x right? More code. Um the temptation obviously is to reduce headcount. Right. So um, and I think that might be, that make might make sense short term but over the long term maybe uh that could hurt you because you do still need as evangelist and I have talked about repeatedly here, uh you still need uh the more code you produce and write the more people you need to actually maintain and watch over all of these systems that you are either using or your customers are using.

Speaker B: Um then you reminded me of something which I think uh some of at least listeners who work in uh uh established companies, corporations may uh, may already be feeling uh as you know there is a, the the word K shaped uh is because. Has become very popular on many contexts. But um, on. Because of the cost that I talked uh about uh a few minutes ago, we're starting to see a type of a K shape budget allocation. So uh, because of the costs that are associated with some of these AI projects particularly CO generation, we're starting to see corporations taking money from uh certain IT projects, maybe new also certain applications, um, migration of applications and moving them towards uh AI projects. Some of it is to plug holes, some of it is because though they are seeing important results and that may relate both to the uh modeling discussion that we had earlier today uh but as well as the program and the use of AI tools for particularly for CO generation. So um, be aware of that. That is not um in last year uh it was all new money going into uh AI projects. Now some of these projects are starting to uh expand and consume a lot more capital. Uh we're starting to see this K shaped budget allocation.

Speaker A: By the way this ties to our previous discussion around uh tokenomics which basically you should actually focus uh on task completion, not token consumption, the total cost of uh task completion. So uh, here you should measure shipped value not activity and activity can mean many things, right? So code, uh number of lines of code written, uh tickets result and things like this. But basically shipped value is uh the main metric. By the way. Speaking of which evangelist, one of the Things that people talk about in this area is the J curve, which is basically as for people who don't know. The J curve, uh, is describes uh, a phenomenon where you know, in the M, on the short term there's a dip in productivity and before that it skyrockets, uh, uh, once again, right? So in this case people are saying we're still in the Jake, the J car before it turns up in the following sense. Right? So right now people are still understanding how to use these tools. They're overwhelmed with the new uh, code they have to review and things like that. So then uh, people are speculating that there will then be a J curve where okay, once people stabilize and really understand how this works, boom, productivity will explode. But this is obviously hypothetical. We don't know if that's going to happen.

Speaker B: Right. Well we, we are. I mean I have written about this uh, for, for those uh, the listeners who go to my uh, blog. Um, I have my, my most recent posts. I mean I've been exploring this concept that I call the AI stakeholder squeeze. Most of the squeeze which is coming from employees and uh, senior management, starting with the CEO, is uh, exhibited during the dip of the J curve. What I'm starting to study, uh, now again looking at our customers and broader data from uh, several analysis firms, um, is whether uh, the uh. Depending on the strength of that squeeze and the forces that create it, uh, whether the dip becomes shorter or deeper or shallower. I mean so it has certain characteristics which vary from company to company. So uh, stay tuned for uh, ongoing analysis. But uh, it's becoming a lot more interesting than I thought it would be when I first started ah, doing the

Speaker A: research and uh, uh another thing that I was able to uncover is that obviously um, uh one of the explanations for why maybe we're dipping down right now and maybe there's hope for going back up is that uh, this is an entirely new developer experience. Right? So uh, people are adjusting to these new tools and, and one of the things that people frankly miss, uh, including myself is uh, the state of flow. It's hard to have a state of flow when you're chatting with a uh, uh, chatbot or a coding agent before obviously by state of flow I mean you're coding and uh, you're just on fire and you just uh, uh, uh, able, able to work through problems. Right? So, and one of the reasons is obviously is because you are able to focus and lock in and things like that. But now I think with these tools you tend to get more distracted. I Think your attention is not as uh.

Speaker B: Actually I thought you were going to say something else also, which is not only that you have to deal with a new regime of tools but uh, within that regime there are new tools every day and you. Which have different flows. Right. So um. Well no, that is also creating stress.

Speaker A: But most people frankly they lock in. Right. So I'm going to use cloud code or I'm going to use cursor. In my case I'm going to use open code. Right. So uh, I think that uh, I think there's still some shopping around obviously but for the most part I think uh, uh, that would be crazy to keep trying something new every week. Um, and uh, obviously there's also the notion of governance around these tools. Right? So um, how do you govern the use of these tools? Because I think you've uh, talked about this in the past, the whole uh, what's the term where bring uh, your own AI. It could be that these developers love these tools when they're at home and then they're banned from using it at work but then they find a way to work around that and basically use it.

Speaker B: Essentially you have a dark IT or shadow it, I'm sorry, uh, uh, with uh, regards to AI. So I called it shadow AI uh, for you.

Speaker A: So to close the loop on this, I guess I started out with this attenuation funnel which was a bit uh, not pessimistic but realistic. But I think overall I think uh, my sense is that these are going to be uh, tools that people will embrace and learn how to use and uh, we'll see if uh. So if you follow that attenuation funnel all the way to the App Store apps, will we see increased usage of apps because of these tools? We haven't seen it obviously at this point maybe the way uh, maybe these tools will allow people to author even more compelling apps that people will download and engage even more. I don't know. All right, so last uh, topic, uh, what evangelist refers to as uh, Neo Consulting. So go.

Speaker B: So again um, going back to this uh, point of how do corporations uh utilize uh generative AI, AI in, in general, um, we're starting to. And uh, I do not know really whether it was the enterprise that motivated it or the providers of this uh, frontier models and particularly frontier models and associated tools are motivating it. But we're seeing the emergence of um, a new type of consulting firm, uh, um, uh, whose primary employee is the so called forward deployed engineer, uh, whose goal is to go into a corporation and help the corporation um, utilize in a better way and faster and see results uh faster uh with the um uh with the tools that uh Frontier Lab or an associated partners are providing. So we've seen most recently Microsoft allocating two and a half. Obviously Palantir started that.

Speaker A: We're not talking so uh, Evangelo. So in the words of Neo. So there's the Neo clouds which really were new clouds. Neo labs which were really new labs. So this NEO Consulting, are there names of actual new ones or they're just new practice, new groups within old ones?

Speaker B: Well m. Look Microsoft started a new company, put two and a half billion in it. Uh OpenAI and Anthropic started new efforts and they took money uh from uh private equity firms like TPG or uh Blackstone.

Speaker A: There's no new Accenture or new wipro New info.

Speaker B: Well uh, there are a number of startups um, starting with Distill which is Palantir, um uh, alumni, um and a few others that we have in our ever expanding database uh of uh AI startups.

Speaker A: By the way, startups and the word consulting usually don't mix together well.

Speaker B: I mean uh, uh look these are, you can think of it as services companies that combine some combination or combine software expertise or AI software expertise with uh, people who are capable, who don't only have maybe an MBA which was the traditional path for uh systems integrators to hire uh people um, but have a lot deeper expertise uh of AI and of specific tasks and they are embedded into an enterprise's um, operations in order to bring in uh the new tooling. What is interesting to me is not only the emergence of um, this effort.

Speaker A: Uh, is there an easy way to think of NIO consulting as a clearly different category from Accenture McKinsey?

Speaker B: So yeah, I was going to say

Speaker A: a one sentence description of what makes a consulting company in, into a NEO consulting company.

Speaker B: Uh, I will actually say first and foremost as far as I'm concerned is the business model because a lot of these companies efforts employ outcomes based compensation uh as opposed to times and materials. And that has uh, very big implications to uh certain services providers, particularly the outsourced providers from countries like India or Indonesia or the lower cost providers but also uh, it has implications to the more established firms. Accenture McKinsey. Because now you have to uh, now before you can recognize that revenue, uh the client has to um, uh recognize results to see and as we know in AI a lot of this is prediction based. So you're predicting that something is going to happen. It's not like integrating SAP into your uh, operations uh where you can see results immediately. So how are you going to how this uh. NEO consultants are uh, structuring their contracts so that uh they can provide value to the customer but also start seeing revenue from that customer. Revenue that they can recognize is uh, a challenge that um, CFOs are uh, working on.

Speaker A: Uh, um. So Neo consultants by definition in your framing focus almost entirely on AI?

Speaker B: Uh absolutely. Today this is, I mean you see them from Anthropic and OpenAI to Microsoft to Palantir to obviously the startups, they're all um, focused on AI.

Speaker A: And then uh, uh you brought out in the outcome based pricing. I bring up two seemingly contradictory terms from the world of consulting. One is uh, being on the bench. Right. So meaning uh, you're a consultant then you're not on the project. That's never good. And then secondly it's highly relational in the following sense. Right. So if I'm, I don't know, I'm making, making this up. I'm JP Morgan, I'm working with McKenzie, I want senior partner X, you know, some, not some no name person that uh, so there's a branding even among the, among the people inside the company. Right. So are those two things gone from this?

Speaker B: No, no, no, actually they're not. And I think if you look at the efforts by both OpenAI and anthropic, um, they're bringing, they're partnering with some of the established, both management consultant consulting firms and IT consulting firms. Um and uh, in fact so to your second point on the partner, uh the engagement ah team is always led by a partner. Uh and the seniority of that partner depends on who is setting up the firm I guess. Um, but one of the things that we've heard from corporations and we're facing it as uh, in part of our advisory practice is um, is that they want the senior partner with what you call the senior partner to be far more engaged in the delivery of that outcome uh as opposed to what was happening before in the past. In the traditional let's call IT consulting model the senior partner is responsible for selling the engagement and then providing some supervisory uh work and obviously ultimately delivering the outcome, the results to the report or whatever.

Speaker A: The main work they're doing is uh, getting these companies to become uh, AI companies. So in other words to actually not just AI strategy but they actually go in there and help them with AI.

Speaker B: I will put it, Ben, uh, I will put it as they are helping the client take the best advantage of the AI tools that they are promoting. So obviously the uh, the, the OpenAI, uh, Neo consultant is promoting the OpenAI tools and is doing the same, Microsoft is doing the same. Um, but they're trying to, they're making sure first and foremost that the uh, enterprise as part of this AI journey that I keep calling, is making best use of the uh, vendors, uh, tooling.

Speaker A: So what you've described to me, the ones from OpenAI, Anthropic and Microsoft, I don't think, I don't.

Speaker B: In Palantir, right?

Speaker A: Yeah, in Palantir. So I don't think of them as NEO consulting firms. They're basically just services companies for that particular company. So NEO consulting to me would, would uh, be someone like Accenture. Ah. Or wipro or Infosys standing up. Uh, uh, a practice that will um, mimic what Anthropic and OpenAI are doing, but they're much more neutral in the fall in the sense that uh, we'll help you do what you need to do, but we don't, we're not tied to one vendor.

Speaker B: Okay, so let's be clear. In the large scale consulting From Accenture to PwC, Capgemini Pay, PMG, all of those guys, they have part even to this day, let's leave AI out. They have partnerships with large uh, SaaS vendors. Right.

Speaker A: So uh, you know, I'm just using them as an illustration. But NEO consulting, NEO means new. So this is a new consulting company. Focus on AI. I'm gonna go into your company and help you with AI, but I don't have a, I'm uh, not tied to one vendor.

Speaker B: Well today they are, some of them are tied to a vendor and I think again you have some of them

Speaker A: are from vendors themselves.

Speaker B: Yeah.

Speaker A: So, so to me they're not consulting firms, they're just solutions engineers for that firm.

Speaker B: Uh, yes. Uh, even though you have to appreciate the following is that as they look at a process to automate a process, as we've been saying all along in this podcast, AI is not 100% of, of the process. Right. So they have to know about other tools uh, in order to be able to complete, to complete the particular uh, task and provide the outcomes that will allow them to get paid. And um, again the. So to me the.

Speaker A: So they're, they're still acting as consultants but when it comes to recommending the AI tools they only recommend their own.

Speaker B: I would, I would imagine that, I mean, and I think that what you will start to see between let's say a startup like Distil or, or uh, HAN10 systems. Um, what you'll start to see is they will differentiate themselves from efforts from anthropic or from OpenAI or from Microsoft by saying we are much more independent in terms of what AI tools we can introduce to your firm and we'll

Speaker A: even use open weights to tie it

Speaker B: back to the first potentially. I mean uh, you have firms like UMI for example that, that can provide that type of, of capability. Right. If they, if they choose to. Uh, anyway so, so my, my point is that the, the NEO part comes from their AI expertise or the expertise of their people. Right. Which is centered around AI and processes. But second and more importantly the business model, this outcomes based business model is completely new and how it is going to be received is going to be very interesting. I mean when we talk to some of the companies we collaborate with in India which, which provide either engineering services or IT services, they are feeling a lot of pressure to reduce their cost because now they have clients in Europe and in the US Saying that if I use AI, uh I can do certain things much cheaper. This still to be by the way,

Speaker A: uh, this new business model being outcome based, actually placed directly into the hands of uh, Anthropic, OpenAI and Microsoft because basically they don't really care about uh, billable hours. They just want to make sure that at the end of the day you're on their AI stack.

Speaker B: I actually think again my personal opinion is that it plays much better in the hands of hyperscalers like Microsoft. Uh, AWS for example established a similar unit.

Speaker A: Right.

Speaker B: Not leave them out because of their balance sheet. Uh, in the case of anthropic AI they have to, I mean they have to get money in order to support uh, this effort. Right.

Speaker A: When I say that I uh, mean uh, I, my benchmark comparison is not necessarily the hyperscalers but the existing consulting companies.

Speaker B: Yeah.

Speaker A: Which, which focus on billable hours and yes.

Speaker B: So if, if the business model takes hold is going to have a big impact on the services industry. Um, it is not clear uh, how far and how much the target customers will embrace it but I know that more and more are asking for it actually Remind me, I'll give you a quick anecdote. Uh, when I was running IBM's AI division this is back in the 90s, we were offering both services and product in my organization and I remember giving an interview in London to and uh, as part of the uh, event um, I said something like um, we will be willing. We feel so m certain about our capability that we will be willing to do a rev share type of an arrangement with a client. This is pretty Watson, by the way. This is 90s again. Just, again to put into perspective, as soon as the, uh, the following day the newspapers print the interview. Actually quite, uh, big newspapers in London. Uh, I get a call from uh, IBM CEO. Uh, that was the last time I talked about, uh, outcomes based and, and rev share type of uh, uh, payment.

Speaker A: Um, so what happened? So what happens to the existing consulting companies?

Speaker B: Um, I think again, that's what I'll say if this thing. So the existing consulting companies have two challenges. Uh, uh, the first challenge is retraining, uh, as many of their people as possible to be able to understand, uh, the needs and provide the AI advice and capability that enterprises need. Um, um, that may lead to a lot of, uh, reshuffling, uh, maybe even some reductions, but most reshuffling. But if the business model takes hold, then I think we're going to have a significant impact or could have a significant impact on their financials, right? On how they recognize revenue. I mean, company like Accenture, 400,000 employees plus.

Speaker A: Uh, yeah, because the reason I asked is in the past, if billable hours means I want more hours that I can bill, but now you don't need as many hours. You want to get the job done as quickly as possible, uh, for the customer. So then I know I don't, you know, the ideal scenario is I have a lean team that can do it.

Speaker B: So here's how you need it. So first and foremost, customers are asking us, and they're asking competitors and all that. And we're a microcosm of what's happening in the broader consulting, uh, ecosystem. But customers are asking for lower per hour prices, rates, uh, because they know that all of us are using these tools in the course of performing a task. So they're saying if you're using Gemini, uh, Anthropic or you cannot be charging me X, you have to charge me 30% of X, 50% of X, you know, some, some smaller amount. So, so that impacts, um, uh, consulting services regardless of whether you're doing AI consulting or meat and potatoes consulting. Right. You know, you're integrating, as I said, SAP or Workday or something like that. The second part is what you do with AI projects. And now you have to both deal with a different type of consultant with a different kind of knowledge in order to be able to compete.

Speaker A: Also, like I said, because you're not motivated by hours, what you really want is a lean and productive team.

Speaker B: Right. Those teams need to be managed differently, need to consist of different kind of people. Then at the end they need to have a different business model. So you may end up seeing the legacy firms, like in Accenture, like in kpmg, having a two track, uh, business model, A, uh, different model that is more time and materials for the more traditional, uh, tasks, but still at reduced rates because of the use of AI in completing those tasks. And then a separate new business model for the tasks that involve delivery of AI.

Speaker A: And for our listeners in India, what happens to the big Indian IT consulting firms? Yes, what happened to them?

Speaker B: I think they are already, uh, seeing a lot of pressure.

Speaker A: Um,

Speaker B: I think for the public ones, I think you will start seeing, uh, missing revenue targets. Um, and, uh, uh, also how they hire, who they hire, how many they hire. Uh, I think this changing, I mean, we're seeing it because some of them are asking for our advice, uh, on how to proceed. Um, but we are, uh, we're seeing, uh, we're starting to see the impact and hear about the impact.

Speaker A: And with that, thank you, Evangelos.