Ep.103 - From AI Hype to Business Impact

Speaker A: Welcome, um, to another episode of Digital Value Creation where my brother and I cover the latest trends in AI value creation.

Speaker B: And I'm coming from an AI focused hardware company and I'm coming from an

Speaker A: AI focused software company. And that gives you different perspectives on the same issues. One of the things we want to go back to every time is, um, value creation, how businesses get real value from AI. And most companies still struggle with this. And uh, is this the stage of innovation condition? We're in the normal pain with it, or is it a warning sign? So today we'll get into the conversation about separating high from reality and try to break down the AI adoption cycle and what it means for different companies. Hopefully it gives you a different way to think about where we are in this journey and how to measure business impact.

Speaker B: Today's session is very timely as people talk about slowing down, scaling challenges and disillusionment for AI. And we believe that this is the time where companies should focus and to some extent double down, but have to be very selective there.

Speaker A: One of the things we are going to include here is there was a Deloitte study that found that only 40% of the businesses deploying AI have a formal, uh, ROI measurement framework. And that matters because if you don't measure the impact of AI, uh, you're spending the money. Somebody somewhere is going to ask the question, where's the beef? So we'll cover some of this. But why is this okay? I think we're in a stage where this may just be okay, Arpad.

Speaker B: Yes, uh, because if you think about early innovation, whether it was the Internet, uh, or E commerce, at the beginning, it was hard to justify or specifically measure the impact. Uh, on the long run though, these AI investment will start to be measured up and benchmarked against other investment opportunities. So while at the beginning jumping in, experimenting technology might be okay. Ultimately companies will need to have KPIs and specific way to measure AI initiative, whether it's revenue growth, cost reduction or customer customer impact.

Speaker A: One of the things, um, you and I talk about a lot is are we in a hype cycle? And of course, if you listen to Gartner, we're in the middle of a hype cycle when it comes to AI and AI agents, AI inference, compute and all the latest terms. And um, one of the things that come to mind is there's something unique to the tech industry in hype cycles. If you think about the other industry, like shampoos and the consumer products industry or the latest pizza trends, the there's really no Equivalent of a hype cycle. So why do we have hype, this acceleration and deceleration that happens with every new innovation and is that a problem? And um, I actually think the tech industry thrives on hype cycles. As new technologies get introduced you need early adopters and this happened with Internet, with blockchain, now with AI. And it feels like it's over marketing but it's also a way to get more people mind share about what this technology can do. But it also bothers me slightly being in the tech industry all my career that it feels like we always exaggerate early in a cycle um, the potential of a new technology and later in the cycle it's proven to be prudent. Um, so we go through those kinds of uh, cycles uh, of over promise and then under deliver and then over deliver the original promise. I um, think we may have said in an earlier discussion is, is ah short term AI is overhyped but long term is going to be under hyped.

Speaker B: Yeah I think part of it is probably the nature of how this tech industry is funded. I mean there are a lot of startup are really starting off with venture investment so there's so much pressure to make sure that they make a big bold promise that attract investors and uh, that creates an artificial inflated expectations around what this technology can do.

Speaker A: It's really interesting. I always go back to Jeff Moore's book Crossing the Chasm and uh, in any tech company but it's also any tech innovation you have the early adopters and then the prudent latecomers and every technology goes through this cycle and the prudent buyers are actually very important solidifying the technology, making it real. So it's happening with AI. A lot of companies spent ahead of the capability already and now they're coming into um, into realizing the benefits probably earlier than the mainstream. But if you look at the other industries like manufacturing, healthcare, finance, they're more cautious. From the beginning you don't spin up a factory uh just on the promise of a new manufacturing technology. And like I said you don't start with new shampoo either. So this kind of um, uh cycle is interesting to me. One thing a friend of mine who works um, in the shampoo industry did mention to me that 80% of of new shampoos and soaps and similar products are failing. So maybe they arguably they have their own way of product introduction and failure. Uh but the point is I think um, uh the hype cycle that we're in the middle of is unique. I think it Serves a purpose and we're going to move into a more uh, traditional ROI and value based evaluation of um technologies. But you thought about the risk of it too?

Speaker B: Yeah, I think there are both short term and long term risk with how we embrace this hype cycle. I mean this ethos of high tech, of move fast, break things, can lead to overinvest in technologies. And I think with AI we saw that already that um, whether it's the famous example of creating a custom overtrained GPT that was just superseded by a standard release or an update and, or saying that um, certain technologies evolved as agents and other capabilities start to appear. So there's definitely a risk of overinvesting and uh, also risk of creating this AI project with a clear business alignment so they become like science projects or IT or technology only initiative. On the long term though, there's an erosion of trust, which is why some companies then miss out when the technology really solidifies and both fail to deliver on the short term. And since investors or executives get disillusioned, miss the time when the real uh, technology eruption can happen, it's interesting how to deal with that, how to balance this innovation and pragmatism and um, talking to other executives who are trying to navigate the same challenge. But we find that companies who make progress focus on selecting pilot programs that has clear measurable objectives. And the other is realizing that where we are in the hype cycle, as Thomas mentioned, uh, we are probably at the peak now in a peak, going in with uh, our eyes open, knowing that some of these technologies will fail. So take a portfolio approach about what we really embrace. And when we start to get a disillusionment overall in AI, that's really the opportunity to look at what use cases actually worked out despite all the disillusionment because that really opens up an opportunity to jump in and scale. So all in all for me the key takeaway is this is unavoidable on high tech. As much as we would love to avoid uh, the hype cycle, it's more about how we navigate, how we embrace that we are going through that and not disengage from the technology just before it starts to deliver its promise. Now Thomas, I know you're passionate about the new hype that is coming up where uh, companies are adopting AI agents.

Speaker A: Agents. Agents. Um, I spend a lot of my years uh, in software automation. I know you spent your uh, career in hardware automation. Um, every time there's a new technology you sort of go back, is this still automation? We were automating 30 years ago, 30 years ago, 10 years ago. And there are different iterations. And when AIG agents came about, obvious question was, is this still the new iteration of rpa? Is it programmatic automation? Is it algorithmic way of if then else, um, type of automation that's been around a long time. And AI agents are different. The basics of it, for those of you don't know, the idea here is that you take human out of the middle. So instead of you interacting with a piece of software to perform certain tasks, you give it instructions and it figures out what to do and it does it on its own without ideally any human interaction. And that will be the ultimate promise of an AI agent. It has reasoning ability, it figures out these are the steps I want to take. And then it has agency, it has a way to interact with the computers to actually perform those tasks. So we're in stage one, uh, of this AI agent realm. And uh, with the latest announcements from Anthropic and now with OpenAI, there are going to be a computer interface to interact with underlying system. So that will be a big deal when it happens. Um, there are many versions of this at a basic level when you can send an email, that's actually a form of agency. So um, a language interface, a chatbot can actually create an email that can be sent. So there are simple examples of that. There's already agents popping up in customer service. So I think we can move from um, um, GPT or, or Genai. Instead of generating text, generating answers, it generates action. So I think we're getting there. And next year it's an obvious prediction, not even a big prediction, that it will be the age of agents. There are going to be more and more sophisticated ways of doing it. Uh, Arpad, how do you think about it? What makes AI agents different for you?

Speaker B: For me, uh, first of all, what you said is proactive. It is able to work independently but, but equally importantly it can learn. So if I'm looking at rpa, it was scripted, it was rule based, so it executed certain things versus agents will be aware of the context and can adopt the context, which is a huge shift. And while there were supervised RPAs that had some level of human collaboration, uh, a lot of agents are actually designed to work alongside human and, and collaborate with them, um, which I think open a door for a lot border adoption.

Speaker A: We're walking a fine line here because we don't want to sound like AI agents are here. Although I'm seeing many examples. You and I talked about it, you Were part of a CIO forum where companies deploy their agents already to do customer, um, service interactions. And they look at the use cases that you're asking about or tickets and they can determine the answers and they can actually initiate transaction. They can dispatch a field service agent, uh, which is a combination of autonomous action, but also a possible set of skills or scripted responses but you can select from. So it's not open ended completely, but it can perform a set of tasks within certain boundaries. And I've seen this in finance, you say pay these invoices and um, the agent can figure out what kind of tolerances, what are the guidelines. So it can read a policy saying under 50 bucks we just pay it. We're not going to overanalyze the invoice or we're going to pay everything within 60, uh, days. And if it's running up against the deadline, it's going to release payment. So it can observe rules without being explicitly programmed to do so. We've seen this in hr. There's a lot of job application systems that are emerging now on both sides, the applicant tracking systems, but also, uh, people applying for jobs that are using agents. So there are going to be a lot of new technologies and I think the question is going to be what are the processes where you tolerate mistakes because mistakes will happen, but then again humans make mistakes too, so I'm actually cool with it.

Speaker B: However, where I see the difference is our approach to managing that risk need to evolve. Right now a lot of focus on AI tools is how can we detect and potentially augment risk when the AI tool create an output, but when an agent takes the whole end to end process, then those mistakes can multiply or um, add because. And because of that, our control framework need to evolve. We need to figure out how can we maybe insert control point in each part of the process. Because the AI workflow is different, uh, because of the speed and scale they can work. So a lot of our typical enterprise controls are geared towards human pace or have some traditional checkpoint that, uh, an adoptive AI system that is context aware might not trigger. Thomas, what do you think? What do you hear as you talk to your peers about what companies do to manage these risks?

Speaker A: Um, we talk a lot about guardrail.

Speaker B: So

Speaker A: I think we talked about this in the prior episode. We want, um, AI systems to be error free, although human systems are not error free. Right. And we try to find analogs. If I ask Johnny to perform a certain task, I give Johnny certain boundaries. Make sure that you pay attention to these set of things. You follow a policy or standard operating procedure, uh, there could be restrictions in terms of payment or size of payment that John can process before uh, he comes up for additional approval. So there are guardrails in human processes and operating procedures. So the debate is like how do we create similar guardrails for systems? And the natural tendency is let's just put human back in the middle, let human double check, uh, the process, the data, um, uh, the final output. What if we apply the same logic and apply systems. So let's have one system check another one. So let's have dynamic monitoring, uh, or enhanced monitoring system that can check every step of the way how an agent performs a task. Maybe another agent's only job is to make sure that it follows a procedure, it follows a certain process, or it makes sense out of an intermediate output, saying does it make sense based on the input? So you can create these, um, sometimes called adversary system. It's a misnomer. You basically have one system check another one. Um, there could be um, there could be multiple um, um, agents involved where um, certain agents only check for certain things. Like it could be one um, agent could just be checking semantic responses or intent. Like you know, it's a customer service agent. One agent is performing a response to a customer. Another agent is only looking for negative signals and that could be triggering an intervention, maybe ultimately getting to human in the middle or, or maybe trying different strategies. Um, um. There's a lot of question about explainability with the process. Anybody who listens to this knows by now we don't quite know how things work in a generative AI, uh, system. So explainability matters. And there are emerging set of tools in this area, um, that can reconstruct why a, um, large language model made a certain decision. So um, there are tools like Shap and Lime and others that do this and there are many other techniques that, that could be deployed to help with the help with these, what I would call guardrail processes or validation processes. So it's important to look for those without reverting just back putting back human backs in the middle. So that's one thought I had. Um, and um, I think we'll put in some of the show notes some additional thoughts on AI specific guardrails because I think it's important to have a checklist. So Arpad, why don't we insert that in the show notes.

Speaker B: I love the idea. In addition to what you highlighted in these AI systems, it's even more important to have audit trails because as you said, these are not explainable.

Speaker A: Systems.

Speaker B: A lot of times auditors look at business controls upfront and a lot of times we use systems to enforce these business controls because systems are rule based. In this case there is a level of probability. Uh, and when we think about audit trail, it's not just the outcome of the workflow but every step of the way, uh, how can we make sure that the workflow stays within certain parameters and ideally even trigger some actions when they go outside. And I think for companies in addition to these guardrails, two additional point uh is very critical. One is continuous learning. I mean understanding how these systems work, understanding how humans designing this system need to adapt, but also how the data driving these systems need to be controlled and continuously updated. Super critical. And uh, last but not least, there's a cultural shift. I think the technology will be more ready for agency than companies. As you said, there will be a tendency to just either try to insert human in the middle or let these agents run and just deal with the fallout. So I believe that there is a need for evolving, both educating executives, business teams about how to use this, but also evolving AIOps. I mean right now we talk a lot about how we design these systems. These systems are not set and um, forget systems, I mean they are learning, meaning they are changing, they are adapting. So it's very critical when we think about AI systems, about how we can continuously tune, manage and operate them.

Speaker A: Let's pick areas in the business where we can deploy the technology to learn and accept higher tolerances for errors and mistakes so we can learn from those mistakes. Because that's another way we can um, uh understand the thinking process, um, how the system comes up with its own rationale of answers because we're looking for that. We're trying to understand how the system actually going to reason through a process instead of making it scripted. Uh, so it's a really tricky balance. But um, what are our takeaways for today? So we need to prioritize explainability. So we just talked about it, um, and as much as possible transparency of how decisions or reasoning happens in these models, how agents decide to do certain things and follow up. At a minimum we need audit trail, we need to say what happened. Um, most models like um, like O actually doesn't disclose this reasoning process so there is transparency issues in the underlying LLM. So you don't know exactly what the logic is but you can reconstruct it, reverse engineer it if you will based on uh, what actually happened in the process. Um, use different technologies to detect problems in the process. Could be Interim steps could be uh, uh, decisions that were made, numbers weren't added upright for some reason the those kinds of flags, it has to be caught. And you can do this with technology. And um, even though we're not a big fan of just experiments for experimentation sake, but pilot workflows in process that are low risk like we talked about.

Speaker B: Arpa.

Speaker A: Ah, what were your key takeaways for

Speaker B: you just to build on your last call? I think it's super important that AI is not a wait and see technology. Don't stay on the sidelines. And uh, when it comes to experimentation, it's not a waste if what you are experimenting is aligned with your strategy and have a likelihood to have a tangible impact in the next say six to 12 months. That's kind of the litmus test I'm using because leaders who invest time and energy to understand where and how to apply AI early will not be distracted by the hype cycle. They will be able to focus on use cases that can be scalable and they can get ahead of the competition. So my question to you how you are using agents, which is the new emerging theme and many believe that we are reaching a point where the traditional scaling and training, these models are slowing down because of the significant increase in compute requirement and just some improvement in performance. So probably next year is the big shift from just better content generation to really action. But that carries a lot of risk. Very curious to hear, maybe in the comments, maybe as a feedback how you are using agents where you are focusing on your experimentations. And with that looking forward for connecting and um, hearing from you.

Speaker A: Excellent. Talk to you next time.