Why are so many AI projects failing?

Keith Shaw: After all this work on artificial intelligence, it’s looking more and more likely that we’re all doing this wrong. On this episode of Today in Tech, we’ll find out why — and how we can still gain benefits from AI. Hi everybody, welcome to Today in Tech.

I’m Keith Shaw. I am here at Pega headquarters to speak with Alan Trefler, founder and CEO of Pega, and one of the leading voices in technology today. Thank you, Alan, for being on the show. Alan Trefler: Thank you, Keith. Keith: And thank you for hosting us here.

This is a wonderful studio. Let’s jump right in. You’ve been talking a lot about how companies misunderstand the role of AI systems, especially in design and architecture. Have you seen examples of hype-driven cycles that are leading to unscalable solutions?

Alan: Well, with all this hype — and with the fact that AI is incredibly powerful in certain settings — it’s unsurprising that people jump in and get swept away.

There are a couple of serious mistakes you see over and over again, and they often create what Gartner calls the “trough of disillusionment.” Gartner reported that at least 40% of all AI projects are failing.

I just saw another report saying almost all projects in the pilot stage are not working. The mistake is in deciding what to delegate to AI, what to curate, and what to let run free. Different settings require different approaches.

Too many companies think, “We’ll just let AI do all the mundane stuff” or even “AI will run the business at the front end.” But not enough people are using AI in planning and preparation.

Many just want to turn everything over to AI — especially if they buy into the hype around AGI. But that creates real issues, especially for customers who expect consistent and predictable behavior. Keith: So what’s the right approach if companies are doing this wrong? Alan: Very much so.

At Pega, we’ve focused on two distinct settings for AI: design time and run time. * Design time: This is where you should use AI for creativity and planning — envisioning new ways to serve customers, designing better workflows, and stimulating collaboration. * Run time: Here, reasoning can be dangerous.

Large language models are highly context-sensitive, and can give inconsistent answers. At run time, you need predictability. So you should use AI narrowly — for things like semantics, language interpretation, or summarization. The mistake is when companies move reasoning into run time. That’s where problems start.

Keith: So when companies want to add generative or predictive AI tools, what should they have in place first? Alan: They need more strategic thinking. Running your business with “prompt-and-check” is a disaster. Prompts are complex, confusing, and highly sensitive to context or model version.

If you’re dealing with, say, a collections operation, you need clear business rules on when to provide forbearance and when to enforce payment. You can’t leave that up to a language model. Reasoning belongs at design time, where you bake it into stable workflows.

At run time, AI should only handle language and mapping, not decision-making.

Keith Shaw: So does this mean that humans always need to stay in the loop, or could there come a time when they don’t? Alan Trefler: Humans don’t have to be in the loop if the workflows are properly curated.

Let’s say I’ve got 50 workflows that define how I open accounts or fulfill certain orders. If those workflows are vetted, approved, and reliable, they can execute without human involvement. In that case, the LLM is only connecting the person’s language to the right workflow.

But if I just delegate everything to the LLM and tell it to “figure it out,” I’m creating unpredictability. That will eventually lead to customer backlash. Keith: You’ve shared examples of AI being used incorrectly.

You’re also a chess master, and I think you used a chess puzzle example at PegaWorld. Can you talk about that? Alan: Absolutely. The most interesting thing about this story is how incredibly confident the LLM was while being completely wrong.

The night before PegaWorld, I took a chess puzzle from the Financial Times — a tough “mate in two” problem — and pasted it into ChatGPT. It responded confidently, saying it could solve it.

Then, in five successive iterations, it made illegal moves, declared false mates, and produced completely wrong answers. It was embarrassing because it was the wrong AI for the problem. LLMs are associative — they generate, but they’re not designed for accuracy.

When I gave the same problem to Stockfish, a proper chess engine, it solved it instantly and perfectly. The lesson is clear: you must use the right AI for the right task.

Keith: So does this mean we’re heading toward a world where companies and users will need to apply different AIs depending on the problem? Alan: Exactly. I don’t think you’ll need dozens, but you’ll need major categories of AI, each specialized for different tasks.

I often describe three types of AI that businesses should understand: * Statistical AI (Machine Learning): Old-fashioned pattern-finding from data. It’s incredibly valuable and still relevant, even if it’s less fashionable today. * Generative AI Features: Things like summarization, translation, or drafting a letter — capabilities built on language encoding/decoding.

These are useful features that businesses will mix and match. * Reasoning AI: How you want your business to respond to different customers or risks. This reasoning must be done at design time, not at run time.

If you confuse the second and third types — letting generative AI reason in real time — you’re going to make serious mistakes.

Keith: And here in 2025, we’re seeing the rise of agents and agentic AI. I recently spoke with Marc Benioff, who’s all-in on Salesforce’s Agentforce. Are we about to repeat the same mistakes with agents that we’ve made with generative AI? Alan: That’s the risk.

Agents should work by executing well-defined workflows. For example, if you’re a telco or a bank, you might have 200–300 defined services. You use AI at design time to create and refine those workflows.

Then, at run time, the agent uses AI to interpret customer language and connect it to the correct workflow. That approach is predictable, efficient, and scalable. But some people are just handing prompts to agents and telling them to run the business. That’s a recipe for chaos.

Keith: Since Pega has already built this design-and-architecture approach, have you seen success on the agentic side? Alan: Yes. Our customers are seeing cost savings, reduced complexity, and improved customer satisfaction. We’re still early, but the results are promising.

Keith: So in the debate between “design thinking” and “AI prompting,” you’re firmly on the design-thinking side. Alan: One hundred percent. Prompting has its uses — for creativity, brainstorming, or exploring new possibilities. But for operations — like onboarding a customer or fulfilling a service request — you want predictability.

That’s where design thinking and workflows matter.

Keith Shaw: When the technology first came out, a lot of companies debated accuracy versus creativity. Some wanted consistent answers every time — like pricing a product. Others wanted more creativity — like generating stories or jokes. How does that fit into your “design time versus run time” distinction?

Alan Trefler: That’s exactly the difference. Creativity belongs at design time. That’s when you can explore, brainstorm, and push boundaries. But when you’re in production — selling cars, running power plants, or serving customers — you need predictability. That’s why our model bakes reasoning into workflows.

Once those workflows are defined, the AI at run time can still interpret language — for example, recognizing that “next Thursday” is a specific date, or that a customer struggling financially may need a loan option. But the AI isn’t inventing new rules on the fly.

Keith: Otherwise, you could end up with an AI inventing a “buy now, pay later” plan without approval. Alan: Exactly. You’d want to design that product deliberately — with input from experts and regulators — and then bake it into workflows.

You don’t want AI improvising it at the point of sale.

Keith: That’s where agentic AI comes in, right? Agents can execute predictable workflows instead of making it up as they go. Alan: Yes. Think of a business catalog with a few hundred service options. AI helps design and refine those workflows up front.

Once they’re created, agents can reliably execute them. The key is that workflows are predictable, energy-efficient, and auditable. If two customers ask for the same service in slightly different words, the agent still delivers the same outcome. That’s what customers expect.

Keith: But some companies are still trying to run everything through big, open-ended prompts. Alan: And that’s the problem. Instructing an agent in plain English — “run my business this way” — and letting it decide what data to fetch or how to respond is dangerous.

It creates inconsistency, legal risks, and customer dissatisfaction.

Keith: Since Pega has already focused on design and architecture, how is that playing out with your customers? Alan: We’re seeing not just cost savings, but real reductions in business complexity. That leads directly to higher customer satisfaction.

And while we’re still early in this journey, the progress has been encouraging. Keith: It feels like you’re firmly on the side of design thinking versus AI prompting. Alan: One hundred percent.

Prompting can still be useful in creative domains — say, an investment banker analyzing a company in new ways, or a marketer brainstorming campaign ideas. But for transactional business processes — onboarding, service, fulfillment — you need predictability. Keith: So prompting is for brainstorming. Workflows are for execution. Alan: Exactly.

That’s the right use of AI for business problems.

Keith: When generative AI first came out, there was a fear of inaccuracy — like selling a car for $1. How do you address those concerns? Alan: In our model, that could never happen. The workflow determines pricing, with only limited flexibility built in.

At run time, the AI can interpret natural language — like mapping “next Thursday” to a date or linking financial hardship to a loan workflow — but it doesn’t invent new business rules. That’s the danger of running on prompts.

AI might decide to invent something new, which could be inconsistent, illegal, or simply wrong. Keith: And that’s where explainability matters. Customers need to know in advance how a system will respond. Alan: Exactly. After-the-fact explanations aren’t enough.

You need to be able to tell people what you will do before you do it. That’s what you’d expect from a physician, a lawyer, or a business — and AI systems should be held to the same standard.

Keith Shaw: A lot of companies seem to be plopping generative AI on top of their existing systems and declaring themselves “AI-powered.” I heard someone joke at RSA that you couldn’t find five companies not claiming to be AI-powered right now. Is that dangerous? Alan Trefler: Absolutely.

Just “plopping” AI on top of legacy systems without thoughtful integration is reckless. Every company has valuable back-end systems — some may even be modern and cloud-native. Those systems define the business.

The right approach is to put workflows on top of those systems to orchestrate how you want the business to run. Then, AI sits on top of the workflows to interpret customer language and connect it to the right processes. That’s predictable and reliable.

Running your business on raw prompts is not. Keith: So companies shouldn’t just throw AI on top of legacy systems, but they also don’t need to start completely from scratch either? Alan: Correct. You have to leverage what you already have.

Workflows allow you to do that while still introducing AI where it adds value. Prompts alone are “mushy” and dangerous. Workflows define clear outcomes. AI then helps match customer language to those outcomes.

Keith: It seems like there’s tension between doing this right — which might take longer — and the pressure from boards or executives who are demanding AI solutions immediately. How do companies balance those forces? Alan: First of all, doing it right doesn’t actually take longer.

In fact, it’s often faster. Crafting and debugging prompts takes a huge amount of time and testing. By contrast, using AI to challenge you during design and to help build workflows is faster and more reliable.

Once those workflows are in place, using AI for translation at run time comes almost for free. So not only is it better, it’s quicker, cheaper, and requires less testing.

Keith: Still, with so much hype, do you see a backlash coming? Are we nearing Gartner’s “trough of disillusionment”? Alan: I think we’re already seeing it. Many projects are disappointing or outright failing, not because the technology lacks potential, but because companies are misusing it.

Businesses are fundamentally collections of patterns. If you just throw AI at the problem, hoping it will figure out those patterns, you’re setting yourself up for failure.

Think about onboarding a new employee — you wouldn’t just tell them “figure it out.” You’d give them a manual, clear procedures, and expectations. AI needs the same kind of structure.

Companies that are relying on hundreds or thousands of loosely defined prompts — and then trying to build “control towers” to manage them — are heading toward chaos.

Keith: So are we listening to the wrong voices about AI? It feels like the loudest voices are the tech vendors who have something to sell, while academics or cautious experts get drowned out. Alan: That’s a fair concern.

Right now, it’s a cacophony — too many voices, too much noise. Ultimately, businesses need to think for themselves: “Do I want to run my business off a prompt, or off a set of reliable workflows?” The companies that succeed will choose workflows.

Keith Shaw: When you look at other companies deploying AI — Microsoft, Salesforce, ServiceNow — what stands out to you? Who’s getting it right? Alan Trefler: Honestly, not many. Let’s take those three leaders you mentioned.

All of them now have what they call “prompt studios.” The idea is to make it easier for businesses to author prompts — sometimes hundreds or thousands of words long — that explain the business to the LLM, which then tries to generate a digital worker or outcome.

For narrow jobs, that’s fine. We do it too — for example, writing a prompt to help parse a document. But if you’re trying to deliver end-to-end service for a business, relying on prompt studios is not predictable. Our approach is different.

You put workflows on top of your systems to define how you want your business to operate. Then the AI interacts with those workflows, not raw prompts. The advantage is that you don’t have to write or maintain massive prompt libraries.

If you want to change an outcome — for example, offer a different type of loan — you simply create a new workflow. It’s reliable, auditable, and consistent. Keith: Do you think this message is resonating outside your own customers? Are companies realizing they’ve been doing it wrong?

Alan: I think so. There’s definitely hope. Companies are beginning to see the limits of a pure prompt-driven approach. And interoperability is improving, so businesses can shift toward workflow-driven models without starting from scratch. But you’re right — hype adoption is often louder than actual adoption.

And many “AI features” being added today are useful, but they don’t deliver true end-to-end outcomes. For that, you need orchestration and workflows.

Keith: You’ve been working in AI and automation for decades. Are you optimistic about where these technologies are headed, or are you more cautious? Alan: I’m optimistic. These technologies can fundamentally improve the way we work, learn, and live.

But with any powerful new tool, there’s always a period of confusion. Right now, we are deep in that confusion. That said, I’ve seen enough to believe in the long-term opportunity.

For example, at PegaWorld in 2019, our head of AI showed a slide with four Rembrandt paintings, one of which was AI-generated. Even then, it was clear this was something extraordinary. He predicted it would take 10 years for such technology to matter.

Just three years later, ChatGPT was released — and it was already transformational. When we saw that, we completely reshaped our development agenda for 2023 and 2024, accelerating our work on using AI at design time to reimagine workflows.

That led to the launch of Blueprint, which has fundamentally changed how we and our customers build systems. And it’s free to try on our website.

Keith: Let’s talk about AGI. You’ve said you’re skeptical. Why? Alan: Because I think people have leapt too quickly from generative AI to AGI. We haven’t even mastered current AI capabilities, and yet there’s hype about a “super brain” that will run everything. That’s a mistake.

The power of generative AI comes from its creativity, its ability to generate variations and stimulate thinking. But when you’re executing business processes — running a power plant, managing financial systems — that’s when you need reliability, not creativity.

Humans separate those modes of thought naturally: sometimes we’re creative, other times we’re procedural. Expecting AGI to seamlessly do both is unrealistic. And, frankly, it fuels unnecessary fear — like the “paperclip apocalypse” scenario. I don’t spend time worrying about that.

Keith Shaw: Alan, this has been a fantastic discussion. Thank you for hosting us here at Pega and for sharing your insights. Alan Trefler: My pleasure, Keith. And I’m glad you’re not a deepfake. Keith: [Laughs] That’s good to hear. That’s going to do it for this week’s show.

If you’re watching us on YouTube, don’t forget to like the video, subscribe to the channel, and leave your comments below. Join us every week for new episodes of Today in Tech. I’m Keith Shaw — thanks for watching.