Why Most Companies Still Get AI Wrong | Anthropic & xAI Investor

Artificial Intelligence is transforming every industry, yet most companies are still struggling to generate meaningful ROI from their AI investments.

In this episode of The AI Visionary Podcast, host Joel Azariah sits down with Matthias Hendrichs, Managing Director, APAC at Pit, former Apple executive, founder of a US$1.5 billion technology company, and investor in leading AI companies including Anthropic and xAI.

Drawing on decades of experience building, scaling, and investing in technology businesses, Matthias shares why most organisations are approaching AI transformation the wrong way and what CEOs, boards, and business leaders must do differently to create sustainable competitive advantage.

If you're responsible for AI strategy, digital transformation, or business growth, this episode provides practical insights that go far beyond today's AI hype.

In This Episode

✅ Why most companies still get AI wrong

✅ Why AI transformation is an organisational challenge—not an IT project

✅ The AI ROI gap holding enterprises back

✅ How AI agents will reshape the future of work

✅ Why workflow redesign matters more than choosing the latest model

✅ The biggest misconceptions about enterprise AI

✅ How CEOs should think about AI investments

✅ The future of SaaS in the age of AI

✅ Will AI replace jobs—or redefine them?

✅ The investment outlook for the next generation of AI companies

Timestamps

00:00 – Why AI transformation isn't a weekend project

00:25 – Meet Matthias Hendrichs

01:11 – Why most companies struggle to generate AI ROI

07:15 – Why AI transformation is different from previous technology waves

11:52 – Will AI eliminate junior jobs?

14:52 – The future of enterprise AI over the next five years

17:15 – Is the SaaS apocalypse real?

23:48 – The biggest risks of AI transformation

26:30 – DeepSeek, Anthropic, xAI and the future of AI investing

30:40 – What every CEO should know about AI risk

35:39 – Why AI budgets are accelerating

40:14 – Is AI really replacing jobs?

42:57 – Why companies still get AI wrong

47:29 – Do graduates still have a future?

48:55 – Are trillion-dollar AI valuations justified?

53:49 – Rapid Fire Round

56:30 – Final thoughts

About Matthias Hendrichs

Matthias Hendrichs is Managing Director, APAC at Pit, a former Apple executive, founder of a US$1.5 billion technology company, and an investor in leading AI companies including Anthropic and xAI. He advises enterprises and investors on AI, digital transformation, and technology strategy, bringing a unique perspective from both operating and investing in some of the world's most influential technology businesses.

The AI Visionary Podcast is a platform where AI leaders share bold, unique and tangible insights about pathways to realizing meaningful ROI from AI, pitfalls to avoid and their success stories.

These conversations go beyond the hype to uncover what it takes for organizations to be successful in their AI transformation journeys from a leadership, talent, strategy and operating model perspective. 

Co-hosted by Joel Azariah and Milind, new episodes release every month via You Tube, Spotify, Apple Podcasts and various other Podcasts Channels

Links

YouTube Channel

https://www.youtube.com/@TheAIVisionaryPodcast

Spotify Channel

https://open.spotify.com/show/3Bt27yjIzXP9Yr1JyqzpTv?si=feq6M3RTR1ywJS_694-Q7A&nd=1&dlsi=ce2a21ba58db44b6

LinkedIn Podcast Page
linkedin.com/company/the-ai-visionary-podcast
Joel Azariah LinkedIn Page
...

Transcript

SPEAKER_03 0:00

This is not a weekend project where the IT department installs a new browser version while everyone's at home and Monday morning you open your laptop and and you're you're done with AI transformation. That's what they breathe into the neck of the CEO. Um and the CEO gets very panicked. Now the junior part is is already fully um covered by the models. So now you could say, well, we don't need the juniors anymore.

SPEAKER_02 0:24

Welcome everyone to this episode of the AI Visionary Podcast. Uh our guest today has been part of the global leadership team at Apple in his role as the head of the App Stores commerce and services business in the Asia Pacific region. He's also co-founded a $1.5 billion startup uh here in Singapore, covering Asia. And then more recently, in his in his uh latest avatar, he's been an investor in the front of AI companies like Anthropic uh and XAI. So he's not just sitting on the sidelines of the AI revolution, he's actually funding it. Welcome, Matthias, uh, to this episode of the AI Visionary Podcast.

SPEAKER_03 1:09

Great to be here. Thanks for having me.

SPEAKER_02 1:11

Great. And if we can kick off, right, I think you know, in your recent interview with uh CNBC, you spoke about how, as an investor, you know, profitability is one aspect, but even more important is how the customers of the companies you invest in are generating ROI. Um, from what we can see on the ground here, some of the companies may be struggling to generate that ROI. What do you think is causing that struggle? And you know, what can they do to generate meaningful ROI?

SPEAKER_03 1:43

Right. I think um, yeah, that's a question or a topic that's very close to my heart. Um I found that with all the advancements that we we have in AI, we we sometimes tend to get carried away by by maybe the features and the newness and shininess of the models, of the tools. But ultimately, what decides whether this is this is really useful is is not um maybe the few uh people on the edge uh of early adopters, but it's it's large enterprise adoption or large um adoption and and with that generating impact on enterprise level, right? And my typically my way of um explaining that is if if a drug company, let's take a Pfizer of sorts, instead of uh them taking 10 years to develop a new drug, if they use AI in their processes um and and and in their research, um, then might only require five years instead of 10 years. And if that can be largely uh attributed to the use of AI, then I think these valuations that we see with OpenAI and Anthropic and the likes, um, then then that is makes sense to me economically. But um, if what I see today um the benefit is largely on the personal productivity side, um, writing nicer emails, um having maybe some calendar suggestions, um, that is that's of course nice to have. That is not worth a trillion dollar valuation of the Frontier Labs, right? So that is that is the gap that that I see. Um, and and when I speak with enterprises uh around the world, I see that adoption is um on the first kind of wave, adoption was making or giving access uh of LLMs to to employees, right? The so-called enterprise uh chat uh versions of the anthropics and chat GPTs and so on of this world, which is a great starting point, right? I I don't want to um negate that, but but I think that is only a starting point, right? And and the real question that is still yet to be proven is how can we create value? And I think in my experience over the last 30 years with transformation, not just AI transformation, but any digital technology transformation in a company, these transformation processes are hard. Um, and they are never, although they're maybe technology driven, they are never a technology-only effort. Um, what does that mean? Um, if you want to really transform a company with the latest technology, in our case with AI, you need to look at process by process, workflow by workflow. This is not a weekend project where the IT department installs a new browser version while everyone's at home, and Monday morning you open your laptop and and you're you're done with AI transformation. It doesn't work like that. It's it's hard work. It's a lot of um uh a lot of work, um, a lot of manual work. We still need people. So AI transformation does not uh work uh you know on its own. So the AI agents that maybe are being deployed afterwards as a as an outcome of the transformation, um they they still require the human beings um to get them into the company, into the processes. And that's where you see the open AIs of this world investing quite significant money in services company, right? I think the last two, three weeks we've seen five, five and a half uh billion dollars being invested into various kinds of consulting and services ventures. So um, six months ago everyone thought consulting was dead. Now, now we see, well, humans are still very much required at least to get the AI into the workflows in the companies. Um, what happens in five years one once you know all of that is done is a different question. But I think that is that's why this is so hard. Um, you need the domain expertise uh plus the technical slash AI expertise. Yeah. Let me give you an example. If you talk about processes and big corporates have so many processes, right? Um I just pick a random one. Let's take vendor onboarding, yeah, something that does not sound very uh enticing, uh, but is a fundamentally broken process in every single company, probably on this planet, right? It's a multi-stakeholder uh endeavor, um, takes always way too long. Um and if you want to streamline that, if you want to say, well, let's let's um use AI to optimize this and maybe even largely run this process, how do you do that? You you can't have your AI forward-deployed engineer uh to do that because um he or she has no domain knowledge uh about vendor onboarding uh and also not the specific industry background. So um it's a multi-stakeholder, multidisciplinary approach that's required, and that's why adoption is still very nascent. Long answer.

SPEAKER_00 7:15

Excellent long answer, and I have follow-up questions. Sure. Yes. So um as Howard Marks uh uh in his memo has been pointing out, there's one difference between this technology and every other technology, and that is autonomy. Yes. So on the example, uh in the example that you just gave, there is uh uh uh the claim that the forward deployed engineer doesn't know the process or the best practice for procurement for that industry. Yes, but the frontier AI that it is working with, it knows. Yes. So uh now there is that uh last 10% of customization for that company. This this part is correct. But the amount that the AI already knows and uh it is a technology that helps actively in its own implementation. Yes. And it is also the technology that helps you actively discover that last 10%. Yes, because of the way you are going to implement it, yes. Uh where testing and evaluation is front and center. So all failures will lead to continuous learning. And the speed with which that last 10% gets discovered is completely different. So uh the transformation process is uh at a high level is very similar, uh is what is required, which was required in the previous technology uh transformations. But the process of doing that, what is important is quite likely going to be quite uh unique this time. Any thoughts on that?

SPEAKER_03 9:05

Yeah, I think you're you're spot on. Um I think it's um indeed a very different process and way of working. Um I still think it's it's um more difficult than most people expect it to be. Um and um just by looking at previous kind of transformation waves, um the technology part was never the reason why any of these projects failed. Right. So it's um we still have the humans, the humans in the loop, um maybe less so uh because of the the agentic nature of the entire process. Um, but we still have the humans who ultimately um maybe don't well, they still use it or or orchestrate or oversee it.

SPEAKER_00 9:50

Yes.

SPEAKER_03 9:51

Um, and um that that is that is a factor that that also needs to be considered.

SPEAKER_00 9:56

Uh excellent point. Uh one more uh follow-up question on this one. Um quite unlike previous uh transformation efforts, one of the essential activities that we do, yeah, uh when in the previous transformation, as is to be human training, etc. The amount of human intervention required and the number of humans required in the new to be process, which are likely to be completely dynamic, yeah, process was a control mechanism we invented 20, 30 years back to c to control the uncertainty so that we can standardize. Yes. But uh now process control gives way to to making goals, yes, and evaluation is much more important. The objective is much more important than the process itself. So one of the things that happens is um that the number of people required to oversee so human in the loop, human not in the execution loop, because then it's uh scapegoat in the loop. It is human on the control loop. Correct. Yeah, so the number of humans, the capability of humans also is going to be different. And if um recently uh the controversy around Cloudflare, yes, one of the interesting things that the CEO mentioned was that for 1100 intern positions, they got one million resumes. So if uh as a leader uh responsible for that transformation, I have a choice. Yes, you want to change, fine. You you will be part of the loop. You don't want to change, I have yeah, uh a lot of options. So what do you how much do you think this perf this perspective actually holds? And how much do you think that there is an issue with this kind of a view?

SPEAKER_03 11:52

Yeah, this this is a great question, right? Um, and I think so. Number one is what is uh what is the role of of humans, right? So as as we just said, it's is clearly not in the execution loop, but in the oversight loop, right? Um what does that mean for jobs? And I think the the first job um category that that uh will be affected here or is already uh heavily affected is software development, right? Because that's where uh LLMs um um are very, very um natively good at. Um and usually in the past we we used to have the the junior developers and and then you you you you do your coding and and you become a senior developer where you oversee um um and and coordinate teams and so on. Now the junior part is is already fully um covered by the the models. Um so now you could say, well, we don't need the juniors anymore. Um or you you take a different stance and you say, Well, actually, coordinating the agents that is actually the most junior task, right? Maybe we don't need the terminology of junior and senior, um, but this this is kind of the entry level, the first level is coordinating agents. So maybe that is what um somebody, let's say, straight out of college or university is doing. So it means you still need obviously certain domain knowledge, uh coding knowledge, um, to oversee um and and and check and probe and define and and and um architect, um, but you also need almost kind of managerial um skills, right? Um which typically college graduates in the past at least would not have to have, right? Um so so then the nature of work changes dramatically, right? So um I think in the initial discussions I saw a lot of people said, well, we only need the senior people, um, we don't need the juniors. But then the question is, how do we get the next batch of seniors if we don't have juniors, right? So so that that was the kind of conflict. But now I think the the thinking has evolved to say, well, actually we still need juniors, but the juniors have actually a very different um uh set of skills and and and tasks and and and responsibilities. So I think that that is actually very interesting, right? Because everyone becomes a manager, yeah. Um maybe, maybe even of mixed teams, right? So maybe sometimes there's still some humans doing some part of the work, right? Maybe not so much in the software development piece, but uh in in other areas. Um and then they work alongside with with agents, and and um you might be the manager of a mixed team of agents and humans. Yeah. So very interesting new dynamics.

SPEAKER_02 14:51

Yeah, it's interesting how you mentioned the nature of work will change uh quite a bit. And we're seeing that impact also on certain industries, right? Whether it's software as a service or consulting, like you uh mentioned earlier. How do you see this playing out, Matthew, so the next three to five years um as AI's impact keeps on increasing?

SPEAKER_03 15:14

Yeah, I think the next indeed three to five years will be the kind of ramp time, right? And and we we discuss the the how an enterprise will adopt, right, in in this kind of new, more um agentic and and automated way and then self-learning way. Um, but that there will be a ramp time, obviously, right? Because as of today, there's uh very little penetration, and then I also think it's probably three to five years around the world um to have significant, I don't know, let's say get to the 80% level or something like that. And I think here we still have these services firms that that will help. And I think that's why the open AIs and Anthropics and so on invested heavily because they realized they need just uh some boots on the ground to um to facilitate the process, um especially because and that that just so many enterprise businesses around the world. Um but I do indeed think that after these three, four years, yeah, when we reach the 80% mark of adoption, um, things for these services firms will change dramatically, right? So I think number one, there will be probably quite a boom and and almost like a golden age for these services firms in the next three to five years before they going extinct.

SPEAKER_02 16:36

That's a big statement. So it's kind of interesting, and what I want to share with you is from speaking with the partners of these various professional services and consulting firms, they also have a similar view that in the next three to five years, they do see a lot of partnership with frontier AI companies. So, for example, uh OpenAI has a partnership with BCG, and this is for the precise reason that you just shared is you know to drive adoption and they need that uh domain expertise to be able to integrate. But beyond that, it's actually difficult to see what happens after the adoption 80% is reached.

SPEAKER_00 17:14

Beyond uh this uh the service industries, yes, one of the things uh uh that people uh have been talking about is SaaS Apocalypse. Yes. A couple of years back, uh Satya Nadela once uh in one interview once mentioned that all uh enterprise applications are crowd databases surrounded by business logic implemented through code. And he hinted that it is quite likely that the business logic code layer will be uh entirely taken over by AI. This was two years back, so yeah. But uh this is somehow already kind of presaging what we are discussing about SaaS Apocalypse, and quite often uh I come across senior um senior leaders saying that why am I paying so much license fee to this CRM company or that? Why can't we just use AI to develop our own? We are not using all the features. How realistic, in your view, is uh are these views? And where do you see the sweet spot where reality will actually settle?

SPEAKER_03 18:23

Yeah, I I I have a slightly more balanced view here. Um I think there are you and you mentioned CRM, right? CRM products like like a Salesforce um are hugely complex software, pieces of software, right? Um to reflect the natures of large ginormous enterprises with uh various business lines and and business logics and and so on. Um to think about um you could vibecode a CRM system of that complexity. I think the is it impossible? No, maybe not impossible. Um you would need to have one of the smartest people on the planet to prompt uh the and and and architect that. And of course, you you would leverage AI to do the prompting partially, but but also that is probably would be one of the most innovative things that have ever been created by any of the LLMs. Number one. But I think that that part could, in theory, be done. Um thinking about how much tokens that would cost, yes. Um you you would say, well, you know, as a Salesforce enterprise license is not cheap. Um but uh burning millions or hundreds millions of tokens on this might also not be cheap. Um you would still need to host it somewhere. Um and and it needs to be um compliant with all the regulatory uh requirements, etc. etc. So long story short, right? I I don't think um for all um current SARS uh offerings they that they will go away just because we now have the the the magic vibe coding in our hand. Um I think that there are some obviously uh quick wins, easy wins, also in the in the maybe in the piloting and in the the product development side. Um but I also give you an example that that I saw um where maybe the vibe coding is is a little on on the potentially on the dangerous or risky side. Um so I've I've seen salespeople who maybe didn't get the resources from product to to make uh enhancements to to the product itself. They would go and vipe code, let's say, an extension to the product uh themselves and and um with the best intentions and would then give that kind of extension product uh to the customers um who also initially were quite happy because it um on the on the surface fulfilled exactly that that additional gap that was was discussed. However, when engineering then reviewed uh this this wipe-coded um uh extension, there were um gaps in in terms of security and and and yeah, um so all kinds of issues um that then needed to be uh corrected. And obviously the salesperson uh does not have the background to understand whatever the the vibe coding tool would would give him or her, um whether it fulfills you know the company's uh security standards and whatnot. So just the ability having the ability also doesn't mean that it's it's it's always good, right? So and I think that is the same. Um so that was in this case was a product extension, so not necessarily very SARS, but but you could say that there are a lot of SARS products um that augment uh today's kind of product. Of people in the company, right? And you could say, well, some of them maybe we just wipe code our alternative instead of paying a monthly subscription. The difference is the most of these tools where you pay the monthly subscription, they're well established, well audited, and obviously you have a proper engineering team to ensure compliance. So I think that that that still I would keep in mind before completely abandoning all of our current SaaS subscriptions.

SPEAKER_02 22:32

But if looking at from an investor lens, obviously is there like a feeling that you want it to happen in a way where it benefits these frontier AI companies so that they actually progress to a stage where you don't need so many SaaS products?

SPEAKER_03 22:47

Well, for me, I think that that there's a split, right? So I I looking at the market, SARS companies are um down across the board, basically, right? And I think for some of them, maybe justified. Um for the likes of the sales force and so on of the world, I think they they won't go away so easily.

SPEAKER_02 23:06

Yeah.

SPEAKER_03 23:06

Um and and maybe for them it's a wake-up call also to say, well, you you just need to keep innovating and you need to keep um you know um being excellent at at what you do. Um uh whereas there are others probably um that are maybe already a uh a clot skill away from from being completely taken out of um you know existence. Yeah. Um for for them, maybe it's it's just too late and and you know like the Figmas of the world, for example, right? So I think Figma is struggling a lot with with Claude design and and and others to to stay relevant.

SPEAKER_00 23:48

Since you mentioned this uh security-related risk, yeah, this is one uh thing that we've uh discovered. Of course, uh with the news of models like Mythos, which are also uh by themselves seem to be quite good, and this will be probably surpassed in capability very soon. But this is just one of the risks. What other risks do you see of uh AI transformation of organizations, which organizations should also pay uh attention to?

SPEAKER_03 24:18

Yeah, I think one uh is is in the category of the the product or the outcome. Uh and and the outcome can be, I think, several fold, right? One is uh if it's a software outcome uh product, um, as we just discussed. The other is if it's let's say knowledge outcome, right? And the the risk being here is that um maybe the results are not factual, they are hallucinated, right? And and we've we've seen we've seen that also, right, with consulting firms uh generating entire studies, and and um um part of that was was just hallucinated, right? So so obviously the the the risk is that that maybe you you you still need to be in the loop to some level, yeah, especially if if your business is the knowledge business, um you you can't just uh completely go hands off. Um so I think that that's that's on the outcome level. Um I think another another level I think that that we see is is on on the on the systems on the system side, right? And and I think that's where mythos um shows vulnerabilities of your entire company, maybe. Um and and the the question is how much can that be exploited, right? And and with that, interestingly, also who gets access to these models at what point is is I would say almost national security relevant, right? And and I think that's why recently the US uh government demanded that they need to get access to every model, I don't know, 90 days in advance or something like that. Um again, uh good question. Um, why would the US government only get access to that, right? And and and but um that's maybe another level, right? Um but but you see, I think government influence um on these AI companies, um, also when it comes, for example, to to China and what what influence they exert is is quite interesting, right? Because um these things aren't they go beyond enterprise influence. Yeah.

SPEAKER_02 26:29

I think you touched upon you know different geographies and having influence on how these AI models are adopted. So a couple of years ago, we had that deep seek moment where you know it came out and then was identified that okay, it's actually possible to develop this at a much lower cost.

SPEAKER_03 26:44

I think this was just last year, actually. Okay. It feels like it was five years ago.

SPEAKER_02 26:50

So from an investor lens, is that something that's okay, actually this is progress, or or no, actually this can be uh a competition for some of the companies maybe that you've invested in. You mean the the overall the pressure from Yeah, this view that okay, there's a different AI landscape or different ecosystem that may be evolving completely differently that could pose either competition or could be showing a sign of progress, for example.

SPEAKER_03 27:19

Well, to be honest, I think with AI or with let's say LLMs, not AI specific, but LLMs, um there's already very little competition, right? That there's probably only five to seven players around the world that that are even on on a relevant level, um, which is very limited, right? Um, and that's that's not great. Um I I was very happy when I saw when when this deep seek moment happened because it kept uh especially the the US models on on their toes, right? And and um they they knew that well, that you know, that there are other companies. I think it didn't matter so much whether that was China or US or it's just somebody with a very different approach to it um was was just shaking the tree. And and I think that's genuinely good for consumers and for every every user um to to keep everyone really at the forefront of innovating and and and challenging themselves as well. Yeah.

SPEAKER_02 28:20

I think to your point, absolutely I agree because we're seeing uh not seeing, but it's already in play the circular economy that they call it, you know, where companies are very well connected to each other to the extent where there may be you know very strong interest to determine a certain type of result and influence people. So we call this, I think in Silicon Valley or elsewhere, they call it token maxing to use certain models or products to the extent where proves that there's a demand there that may be real or not real. What's your view on that?

SPEAKER_03 28:55

Yeah, it's interesting, right? And and so so number one is is the dependency because there are just so few players, right? And and if you go through the different layers of of the AI industry, right? From maybe we should need to extend it to to even starting with energy, right? Energy, then data centers, compute, the chips, the CPUs, the uh the GPUs. Now now there's there's a CPU comeback, right? We saw Intel, uh Intel is is back, right? Um and and and then going to the LLMs and and then to the application layer, right? And ultimately everyone needs the the GPUs, um, everyone needs the compute, uh, everyone for that needs the energy. So it's um there's high dependency, right? And and especially when you then see how I don't know, in Nvidia invests into anthropic or open AI, and and uh they they use some of that money to to buy the the GPUs. Um that's the circular kind of spend, right? Um at the same time, I think people like uh Jensen Huang from NVIDIA um suggesting that that uh a developer who who you know um makes hundreds of thousands of dollar salary should also spend hundreds of thousands of dollars in token, right? Just to justify basically his his worth, right? Or worthiness. Um so it's a it's a strange time, right? Where you don't go, especially when when energy and compute is limited, you would say you go for conservation, right? And you go for efficiency, uh, but you rather go for all in, right? So um that that seems to be the name of the game at the moment.

SPEAKER_02 30:40

That's why sky is not the limit, they're gonna have data centers in space, right?

SPEAKER_00 30:44

I like that. Yeah. Yeah. Well, it's a very interesting risk perspective that you brought uh uh brought up, that uh only from LLM's perspective, yeah. We have only so few of those, yeah, and maybe also geographically uh yeah uh constrained. But as soon as we move out of LLMs uh to slightly more emerging ones, uh let's say we talk about world models or embodied AI, there are even fewer players in those in that space. And what uh immediately as a of course, since you are an investor in many of these companies, uh I will also ask you uh two-part question. One, yeah, uh, what is your view on investing in these companies? But from a business leader perspective, this clearly uh is a business risk, a concentration risk. Sure. Yes. And uh how do uh business leaders go about thinking um and trying to mitigate such a risk?

SPEAKER_03 31:44

I think so. From a let's start with the investment piece, right? Um especially if you only have very few players, you actually probably want to, uh at least I, for my own uh investment thesis, um you want to get broad coverage, right? Because you you don't know um what what happens. I mean, we we all wake up to to new models and news um every morning, right? And and it seems the the cycles are getting faster or shorter and shorter. So that's why I personally started to to invest or use vehicles that give me broader exposure. And that is also more difficult these days than maybe in tech investment in the past, because tech investment in the past, most of these companies would be public listed, right? And and uh obviously um almost every everyone has has these apps or can at least get apps, these apps for trading on their own phone, right? So public trading um is is very, very easy. Um private, um, privately held companies is not so easy. There are other vehicles that can be used for that. Um but um yeah, the the barriers is is a different one. But I I I would say getting broad exposure because you don't know um where things will go, which are the players, what will happen. Um we're still at at an to some extent at an early stage, uh, although the valuations are already very, very high. Right. Um so that's that's maybe my view from a from an investment angle. From a from a business and and leadership perspective, I think you also actually need to hedge your your your bets, right? Um because things think things happen so quickly and change so quickly. Um you want to keep um more more connectivity across a company uh a number of players. Um maybe also try them side by side, right? Or you pick certain workflows that that um where where you choose a certain uh provider versus others. Um I think companies like Open Router are very uh interesting, right? Where where um I think companies technically or not technically will start to use different models for different use cases, right? And then they will route different uh requests based on on the complexity of the request and so on to different providers. Um so I think that that will even increase further. And I think that will go beyond just using um one provider's kind of um fast and and low end and low-cost model. Um that will go across to maybe even in some parts of the world uh using Chinese models alongside American models, um in a mix of open source, on-premise, in the cloud. So I think we will see quite a sophistication, especially for large enterprises, um, developing these kind of um uh frameworks and and playing fields.

SPEAKER_02 34:56

Yeah, just to bring back a conversation outside, Matthews. What I'm seeing in in my world is because companies are so eager to prove themselves as AI first companies, especially at the CEO and board level, they are kind of jumping all in into collaborating with uh AI companies or companies that are building AI products and solutions to showcase that we are partners with this company, so that kind of makes us an AI company. But that's what you're saying is a potential risk where you know companies that don't understand the technology properly and what it means for them, it could have a lot of implications, both positive and negative, going forward.

SPEAKER_00 35:39

That's right. One of the interesting news um uh again from an AI uh adoption and AI transformation perspective that uh came out recently that one of uh one very well-known company burned through um their full year AI budget in four months. Yes. This is very similar to uh akin to what uh Joel was just mentioning. Yes, everybody got access to these tools, and uh now uh everybody's uh KPI is linked to it, you do this, or you are not getting promoted and not getting bonus. And now uh the AI budget is finished. Of course, this shows some gap in their understanding um of AI transformation and the journey uh roadmap. Uh any advice from your side? Yeah, uh, what is a smarter way to go about this?

SPEAKER_03 36:32

Yeah, that's that's an interesting one. And I actually the other day had a conversation with a CFO um who asked basically the same thing, right? Um, how do we budget? What's the right budget? Because in in your example, that a company ran out of uh their budget or their allocation um in four months could actually be a good I based on on that information. I don't know it's good or bad. It could be either, to be honest, right? It could be um so if they also achieved their targets in four months instead of 12 months, I I would say that that's a pretty pretty good result, right?

SPEAKER_02 37:11

Everybody gets promoted.

SPEAKER_03 37:12

And in that case, I mean I I would just revise the targets, right? Um the question and and that links back to what's the ROI on all of these things, right? Would they just burn through tokens because they token use is is more expensive, because the model is more advanced, but um the outcome is still the same, right? So the outcome is maybe only at the four-month level, um, but the the cost is at the 12-month level. Um so which to be honest might be the case. I don't know in your example, right? But uh, I don't know the case. But um so I think here it's it's it's still so early to to understand and and assess what are the right KPIs, right? Um, but I would I would try to link things more to an outcome than to a to a cost. Um because ultimately we want to influence outcome, right? And and um and and we don't know yet how much a certain outcome will cost us. Obviously, it it that there's a certain point where it's prohibitively expensive, right? Um, and that's where maybe you need to have a threshold where you then then draw the line. Um, but I would look at it to define the out the desired outcomes, um, and and maybe have a couple of milestones leading and yardsticks to that, right? To to make sure that that you don't uh realize at the 95% level that that you're yeah um used your entire annual budget and then it's uh maybe not April but March, right? So so I think that that's that's the new discipline. But frankly speaking, I don't think or I don't see anyone having clear answers and frameworks how to build this yet, right? Because these things also change so quickly. I think as we said, that the deep seek moment was one and a half years ago. I think it was January last year. Um just looking at what happened in in that that those last 15 months or so um is is is I think the the the real token usage and the token cost uh increased tremendously, I think um maybe in the last six months only, right? Maybe six to eight months. Um so, but also the productivity of the tools increased a lot. Um now the the the biggest risk, I guess, is using these tools. And I also see that actually, using these tools, um, or let's say LLMs, not tools, um, using the wrong, let's say, let's say that the most you know um powerful, let's say a claude opus um model for a task that should be uh a Claude haiku or something, right? Where where um you would then get a lower outcome because that the additional cost that you spend for the opus model is not reflected in the in the outcome. The outcome is still still the same. Yeah.

SPEAKER_02 40:13

And so you mentioned talking about our outcomes, Matthews. One of the key outcomes that most of the senior leaders are measuring is cost. And so to balance that cost, we've seen a huge on social media about jobs being impacted because of the costs. Right. How do you see this playing out? So you've been, you know, business leader yourself, leading very large organizations. How do you see this playing out in large enterprises as the cost kind of goes up, at least for some time? Right. And then they're having to balance and showcasing it in a positive light for the shareholders that we're actually cutting costs on the other side when it comes to human capital.

SPEAKER_03 40:53

Yeah, I think this cost discussion is a very difficult one, personally, right? I think um looking at efficiencies, again, coming from a process or workflow level, right? Um, I think there are efficiencies that that you can um realize, and efficiencies means you can do the same um at a at a lower price. And typically that means the the the savings is on the on the headcount side. Right. Um but I think this this is a very dangerous way of looking at this uh only. I always think, you know, when when when when we look at AI, we're looking at capabilities, we're we're building capabilities. Um of our we're increasing um maybe our capabilities also on the product side, on the on the um offering new type of products, new type of services, right? Um so everyone who's is uh worked in in kind of restructuring or or optimization work knows that one percent increase on top line uh give gets you so much further than one percent decrease on the bottom line, right? So the that basically means cost savings. Well, everyone needs to clean their house and keep your house tidy, right? We know that from from at home, uh, but and we know that from the company. Yeah, but but the real financial gain that that we we should strive to to to create is on the top top line, right? And and uh I think for me AI is not is not a tool, it's is is more a way of working um and a capability, and and it is surrounding all of us. So we are living in an AI-enabled world. So let's build products and services for an AI-enabled world, and and where AI can maybe help us generate more business and and more revenue, and then the the efficiencies uh they will matter less.

SPEAKER_02 42:57

Well, that's great as an overarching goal, but what we're seeing here is companies are saying that they are actually realizing cost efficiencies by cutting jobs. What do you think is really happening here?

SPEAKER_03 43:09

I think I think that there are two parts, right? Because enterprise adoption is still so early stage. Uh I think a lot of the layoffs that are now conveniently being labeled AI related are actually not AI relabeled yet. Um, because it's too early. There are very few use cases that would really allow people to completely step away from their role. You have maybe some in, let's say, call center case uh examples, right? And also maybe not so much in this part of the world because labor costs are very different, right? To to displace a call center in the Philippines, um um, not so easy uh to do that in the US or in in high, high uh labor cost markets, maybe. Um, but I think number one, that there's a lot of mislabeling. Um, and it's very convenient because uh the the uh global news you know feeds that narrative um i think that's number one um num number two is um i think we we need to work on cases where where the maybe the efficiencies that the work we talked about the software developer who um now becomes a manager instead of being a junior right so so you need to upgrade the work i i don't think we necessarily um uh in in three four five years time will all sit at the beach because there's no more work to be done um so bad news yeah so i'm so sorry um but but we will do very different type of work right we will be more the orchestrator um we might all have our fleets of of uh agents working uh for us on on various things and and maybe a company instead of what what you needed where you needed 50 people can be done by with five people right and and and the rest is is um age agents um so so i think that there's more more upside that that we need to create rather than than just pure cost efficiencies how do companies travel that journey because right now they are very focused on realizing something immediate and like you said labeling it differently yeah and that and that's what and and and I'm I'm also advising boards of companies and and most boards go first and ask about like what what are the efficiency gains we can have right that that's the default standard question um and and and that that's what they breathe into the neck of the CEO um and the CEO gets very um panicked um because there are oftentimes no no short answers or quick quick wins of that sort right um and and then they they just downsize certain departments but really not because of of AI but more because yeah you need to keep your house in order to some extent. And then doesn't that create more resistance to adopt AI from the folks who are there in the company exactly after the absolutely and then that's why earlier when I said um technology is usually not the reason for failure right um if you have uh an AI solution that is enabled um but the people who who still are still there right we don't have a zero people company um so that there are still people um who then work with these maybe highly agentic um solutions but they are people and and if they think they will be part of wave number two or three or four um they are very hesitant to to embrace those um which then renders maybe the entire AI initiative um worthless uh and then the the narrative becomes well ai didn't work so yeah we see that quite often exactly exactly so so this would be actually so in the in the history of of humans that that's a very uh repetitive pattern yeah and so interestingly on the point where you said for folks to be there to go up the ladder you need to hire young people but right now we're seeing folks who are in university or trying to start their career the the labeling of whatever is being said in the media is they they're not required anymore or those skills or we don't need fresh graduates to do certain things that AI can do.

SPEAKER_02 47:29

Again is that accurate or what what are you seeing on the ground?

SPEAKER_03 47:32

If if you see how many junior software developers open AI and Anthropic hire you would wonder like well they they have all the models they are AI first um yeah you you you still see them hiring all these people um so um I think in a recent interview daro Dario Amodei from Anthropic said um the the difference the question is not whether we hire them or not but how um effective uh and and productive do we make these people right and and maybe um without ai it's a one one to one relation and now maybe it's a one to whatever 10 or 20 yeah um that at every even junior resource that's that's being hired um so maybe on a macro level you could say do we need as many on a macro level um and maybe may maybe that then the answer is maybe no um but um I mean even we we have autopilots in in on on planes for decades right uh we we still have uh people on the plane who know how to fly a plane um and otherwise to be honest I wouldn't get on board um so so I think this you you still need people to to learn the skills and learn how it works um and and be there even if they only design the system but somebody needs that that domain knowledge oh completely agree flipping the view so we discussed just now how hard it is yeah for companies to go through this journey and make the transformation and realize the value flipping the view to these frontier frontier ai labs yes uh how uh do we justify trillion dollar valuations yes and uh what is it that they can do uh yeah to uh uh to help their end customers on this ai transformation journey what is it they're actually planning to do right now yeah I think the the justification for the for the valuation is uh is a tricky one right and and I think currently it's there's there's a lot of hope uh um in in those valuations um and frankly speaking even if if you go to public markets right so so it's it's always the question is how much hope and how much future earnings do do you uh bake into these kind of valuations right and and and I think personally for the right uh reasons there's a lot of hope for entropic open ai for any of these companies right because they I think they're all doing the right thing um which which one will win or how many will win nobody knows that's that's how markets work right um um is it worth a trillion is it three trillion is it 500 billion nobody knows right so this um what what you see i think the the latent demand now leading up also to these the ip potential IPOs that that we see um of anthropic and openai and and the first one being SpaceX um and in and I think in June um there's a lot of latent demand from the public markets right you you see that from uh the secondaries um where where people family offices and and and and investors try to really get positions in in those those businesses so I think that's that's fine um I think the question is what as you said what will they do in return right so so and and that is that is partially so that they're partially limited to some extent and and you see for example in anthropic also having a very different approach to an open AI. OpenAI started very much on the consumer side um having a um user base of I think close to a billion uh uh users these days um on on the on the consumer side um and and the the business was a little bit of an afterthought um more recently they they changed that um also probably in line with what they see the with uh the competition and with anthropic where anthropic had always been very much business focused from the from the beginning and they're now openai is now trying to go further also down that route but the other interesting part is if you look at this part of the world Asia Pacific right where openai started opening their first office two and a half years ago in round numbers in in Japan and then you know establishing local presence uh across the region umthropic has been almost two years behind that yeah so so they by now they also have um offices in in in Japan Korea Sydney India um not yet in Singapore um but um so the difference is and and both companies are worth close to a trillion dollars almost the same yeah um for one of them they focused only on product and only on enterprise the other had a broader focus and went more the classic let's say tech um expansion where you would you know go with your distribution around the world um it seems for anthropic the focus to only one sector the business sector and also the largely US uh geofocus has not stopped them from being a trillion dollar valuation valued company right so you could almost argue and say well maybe the the global distribution with its complexity especially here in Asia Pacific that these markets being very different from one another would require a lot of focus um and they decided to put that focus rather in doubling down on enterprise only and on the model and and the product itself rather than having a broader distribution. So quite quite interesting to see these quite different approaches of the two leading players.

SPEAKER_02 53:48

Excellent we need to change our pace a little bit so we're gonna go to the next section of our podcast episode which is called Rapid Fire spontaneous quick answers um very fast. So first one for you Matthews open AI anthropic or Google Deep Mind if one of these was to exist in the next five to ten years which one would it be Anthropic. Which one happens first AGI or AI bubble burst AGI if if you had to if you had the power to shut down one AI com uh company for the good of humanity which one would it be I feel you're leading me onto something yeah um XAI Human intelligence or artificial intelligence which one is overrated both will AI eliminate or create more jobs will XAI outlive Tesla yes last one most overrated metric in AI right now token usage great we we're at the end of this episode thank you so much Matthias it's been a pleasure a fascinating conversation and thank you to everyone who's joined us um please look out for more episodes of the AI Visionary podcast on YouTube, uh Spotify and Apple Podcasts and don't forget to subscribe. Thank you.