The Hard Truth: Some Applications Should Never Use AI

[00:00:00]Using AI on the wrong problem is like creating a rocket ship to deliver your groceries. Impressive, expensive, and completely unnecessary. Let's talk about it.

[00:00:11]So, welcome to Dave Lenticum is not AI.

[00:00:14]I'm Dave, an AI SME, and an AI skeptic looking at AI with a balanced perspective. Let's get started. I made this video because a lot of people asked me to make these this video. So, many enterprises are using AI when it's clearly not needed. So, this is costing enterprises billions of dollars daily as simple applications are over-engineered and of course overpriced. So, let's look at the type of applications that are best for AI. In other words, looking at the problem patterns where AI is going to be a solution to that problem pattern. Because I think it's very important because lots of people are making this mistake now. So, I'm not going to re-mention it here, but there's a lot of AI failures going on. You know, according to the MIT report and other surveys I'm seeing as well, you know, as many as 95% of the AI systems out there are failing to return return on investment, which means they're failing.

[00:00:59]And largely, that's because people don't know how to use the technology, but also they're picking the wrong problems to solve when using AI and mismatching the technology, in this case AI, with the kind of problem or solution you're trying to bring to bear to solve the problem is just going to kill it. So, [clears throat] first we need to do is start with the core decision. Is the problem deterministic or probabilistic?

[00:01:23]So, ultimately, a deterministic problem has a clear exact answer. So, if you give the system the same input, it should always produce the same results.

[00:01:34]That's the difference. Probabilistic problem involves uncertainty. So, systems estimates what's most likely to happen rather than guaranteeing one exact correct answer each and every time. So, if the system needs to calculate an exact answer from clear rules, traditional software is usually a better fit. If the system needs to predict, classify, rank, and interpret, estimate under uncertainty, then AI becomes a much more attractive solution to make that happen. So, this is core to it. In other words, it look at the problems you're looking to solve. In other words, is it going to go, you know, deterministic or probabilistic? And should you look at the problem or solve the problem in a way where AI is going to be a fit or typically traditional application development will be a much better fit. Keep in mind that they're not equal in cost. Now, obviously, some of the estimates are all over the place and your mileage is going to vary, but normally generally speaking, AI systems are going to cost five to 10 times that of their traditional analog. In other words, if I'm solving an inventory control problem, you know, I'm going to pay a million dollars to build that system using a traditional approach. I'm going to pay 10 million dollars if I'm going to use AI. That's a big difference. Also, the operations are different as well. AI systems are going to cost 20 to 30 times the amount of money to operate over time. So, it's not free. If we're using AI to solve a problem where traditional software is perfectly capable of solving the problem, we're going to be wasting a lot of money.

[00:03:05]So, AI is strongest when input is messy or unstructured. So, ultimately, traditional systems do well with clean fields, you know, like in fixed formats, but struggle with emails, PDFs, images, audio, chat messages, and free text.

[00:03:19]AI is able to add value when software has to understand meaning instead of just reading predefined fields. Good examples are support tickets, claim descriptions, contracts, and you know, visual inspection. We already we already know this from using LLMs, ChatGPT, you know, Grok, for example. They're they're very good at looking at unstructured information and making finding patterns within that. That could could include a video or an image or a PDF or a spreadsheet, whatever. And if you're doing stuff like that, then AI is going to have some value because we're not having to program the specific ways in which we're going to call the information out of that unstructured format, unstructured data. So, keep in mind the more unstructured it is, the more AI is going to be a good fit.

[00:04:05]However, most applications that we're building out there are going to be extremely structured. They're dealing with structured information, structured content, user interfaces that are very structured, where AI is not necessarily going to be indicated. And I think what's happening now is people are taking problems where that are highly structured and they're throwing AI at it and they're overspending on solving that problem.

[00:04:29]So, next AI works well when businesses have patterns but not clear rules. So, a strong point, you know, in that I'm trying to make here is that many companies have years of data and experienced staff but cannot write stable rules engines. So, that often is the sign that the logic lives in patterns rather than procedures. So, examples include fraud detection, demand forecasting, predictive maintenance, and churn prediction.

[00:04:58]So, very much like the last point we made, you know, where you're dealing with something like fraud detection, where you're looking for patterns within data.

[00:05:07]AI is really really good at looking for that. So, in other words, we can figure out what these patterns, find the patterns, and figure out what they mean.

[00:05:12]However, the majority of applications we're going to build for businesses out there don't need to do that. So, if we're using AI, then and you know, that requirement is not there, we're just going to be overspending and over-engineering the application, which is a fail, by the way. We're not going to get ROI back from that application, even if we're able to make it work and solve the issue because it's going to be too expensive to build and too expensive to operate.

[00:05:36]So, the economics matter. You know, AI should solve a problem, you know, worth paying extra for the technology to solve the problem.

[00:05:45]So, ultimately, AI is not just a technical choice. It's a business investment. It makes sense when it reduces costly errors, saves labor at scale, improves revenue, or detects things humans would normally miss. If the business problem is small, stable, and easily automated with rules, AI may just add cost to complexity.

[00:06:07]So, and this is kind of core to everything. You got to remember, I can make anything work with enough money and time, and most technologists can. And so, I can certainly build a system where AI is going to be a force fit and it works perfectly. I can build a system where I'm using traditional, you know, development tools and technology, and it's going to work perfectly as well. And both of them are going to solve the problem. However, one is going to cost a million dollars, the other one's going to cost 10 million dollars. And that matters because we don't have an unlimited pile of money within these enterprises out there.

[00:06:37]There's budgets that have to be adhered to, and what we're trying to do is build these applications with the minimum viable technology that's able to return value to the business. That's the goal there, and that's what the architect's role is to really, you know, pick the right technological configuration that's going to bring the most value back to the business. That's not always AI. In fact, it's a small minority of problem domains out there where AI is going to be a fit, and we need to understand how to recognize those things. Right now, everybody is just approaching every problem they're looking to solve in the business right now and then force-fitting AI, and that's why they're failing.

[00:07:11]So, human in the loop is often the smartest architecture. So, you know, ultimately, this makes the discussion more realistic. Many of the best AI systems do not replace humans completely. They assist them by drafting, prioritizing, flagging risks, extracting information, or recommending actions. This helps organizations get AI's benefits without controlling error, bias, and compliance risk. So, ultimately, AI is best used to augment human productivity. It's not designed to replace human productivity.

[00:07:47]I always tell people, you're not going to replace humans with AI. You're typically going to replace tasks. Now, you may be able to make the few humans that you have on staff much more efficient and maybe able to do stuff with less humans and move those humans to other aspects of running your business where they can provide more value. But this idea that we're going to have a genetic people, you know, you know, like McKinsey's been talking about where they're in essence equate to a human being. That's not necessarily realistic.

[00:08:14]We are going to see productivity benefits, and we're seeing that now, but the idea should be that we're building AI to augment human genius and innovation, not replace them.

[00:08:26]So, AI is useful when the environment changes faster than rules can keep up.

[00:08:31]So, ultimately, rules-based systems are brittle when the world shifts consistently, and AI is better in environments where fraud patterns evolve, customer behavior changes, you know, equip equipment performance over time drifts, or language usage changes over time. So, that's a strong reason to consider AI even when the rules engines already exist. That's absolutely a core thing to this. You got to remember, AI is good at providing dynamic behavior.

[00:08:57]It's able to learn as it goes, and it's able to become better at what it's able to do. That's why we use it. So, if we have a non-static environment, we're constantly re-evaluating the way in which we approach a problem based on the changes in the data, based on the patterns in the data, AI is a great fit.

[00:09:12]You know, fraud detection, you know, customer demographics, you know, all these sorts of things that are going to shift things over time in very real-time way. So, if you're using a traditional development technology approach, you're going to have to change your applications every time you want to change the behavior within that application. However, AI is capable of changing its behavior ongoing, and that provides a real business benefit if you're dealing with systems that need it. If you're not, it's going to be overkill.

[00:09:39]So, not everything should use AI, and knowing when and not to use it, you know, is basically part of being a good architect. And we don't have a lot of good architects out there, certainly not a lot of good AI architects. And so, people are making mistakes and overusing the technology right now. I find them all all time. I'm auditing projects for a living. You know, look at something it's failed and what happened? Well, they picked the wrong use case. They used AI, ended up spending 10 to 20 times the amount of money on solving the same problem, and ultimately was a weaker application or use of the technology, and that's going to impact the business. You do too many of those, you're going to take the business down because you're going to be wasting wasting your resources in trying to solve problems using the wrong technologies. So, you know, ultimately if the problem is based on a clear policy, exact calculations, strict compliance, and high, you know, explainability, traditional software is usually a better choice. And that's going to be a good portion of the applications you're solving, and your mileage is going to vary based on the industry that you're in, the kind of company you work for, the age of your technology, all that stuff matters. So, good examples are billing engines, you know, eligibility checks, tax calculations, approval workflows, and policy enforcement. The real skill is not adding AI everywhere, but knowing where AI generally creates value, and that's the core point I'm trying to make here.

[00:11:01]Well, anyway, let me know what you think of the comments, um whether you think people are making lots of mistakes and overusing AI, which I believe, or you think that uh you know, AI should be a part of every application development project out there, which I don't believe, and that's a good way to go bankrupt. Anyway, don't forget to like and subscribe and check out my other videos on this channel.

[00:11:18]Also, check out my InfoWorld Cloud Computing blog, my 100+ LinkedIn Learning courses, and of course, my Generative AI Architecture course at Don Google Cloud Careers. And finally, my latest books, Unlocking the Power of Cloud and An Insider's Guide to Cloud Computing. Finally, check out my other YouTube channel, The Cloud Computing Insider. Link is in the description. So, until next week, stay very, very safe.

[00:11:37]The magical YouTube algorithm thinks this is another one of my videos that you should watch next.