When Bold Becomes Reckless

What IBM’s $4B Lesson Teaches Us About Overspin

I’ve been interviewing CEOs for my book on trust as a competitive advantage. A few conversations in, I started noticing a pattern: IBM’s Watson kept emerging in our AI-focused discussions. Some executives referenced it as a cautionary tale. Others treated it like a mirror, quickly adding, “We’re not making that mistake.”

Here’s what most people miss about Watson Health: IBM wasn’t reckless. They were pioneering. After Deep Blue beat Kasparov in 1997—and became the first computer to defeat a reigning world chess champion under tournament conditions—IBM had earned the credibility to aim high.

Watson launched between 2007 and 2011, before modern deep learning existed, before anyone understood what AI would actually need to work at enterprise scale. They chose bold because that’s what innovators do when they’re first to the frontier.

But somewhere between Jeopardy victory and clinical reality, bold became reckless. And the gap between those two destroyed $4 billion in value.

The Pattern I Recognize Because I’ve Lived It

Early in my career, I worked at a startup that built a point-of-sale solution to bring health insurance and providers together at the point of care. The vision addressed a pain point that still exists: eliminate administrative friction, reduce claim denials, and improve patient experience. IBM acquired us right as the dot-com bubble burst.

We had vision, but the infrastructure wasn’t ready.

I watched our valuation evaporate faster than technology could catch up. Not because the idea was wrong—knowing what your healthcare bill will be is still a huge annoyance for both patients and providers. The issue was that we promised transformation on infrastructure that couldn’t support it. The technology we needed didn’t exist yet. The integration challenges were far more difficult than anyone anticipated. The market timing was off by decades.

Sound familiar?

What Actually Happened to Watson

Let’s get specific about where IBM went wrong, because the details matter. (Check out Haverin Substack for a deep dive on Watson. Stuart Miller’s analysis is gold!)

After Watson’s Jeopardy victory in 2011, IBM moved aggressively into healthcare. They announced partnerships with Memorial Sloan Kettering Cancer Center and MD Anderson Cancer Center, promising AI that would revolutionize cancer treatment. The marketing was ambitious: Watson would help cure cancer by analyzing millions of pages of medical literature and providing evidence-based treatment recommendations in seconds.

MD Anderson invested $62 million over four years trying to build an “Oncology Expert Advisor” using Watson. By 2017, they shelved the project entirely. The system couldn’t reliably ingest medical records. It struggled with unstructured physician notes. In one internal test at Memorial Sloan Kettering, Watson suggested treatment regimens that doctors flagged as “unsafe and incorrect.”

The problem wasn’t that Watson couldn’t process language. It could. The problem was that clinical decision-making requires understanding context, nuance, and edge cases that no amount of training data from Manhattan’s Upper East Side could generalize to patients in China, India, or even rural Texas. IBM had built a system optimized for trivia questions and tried to retrofit it for life-or-death medical decisions.

Between 2015 and 2021, IBM spent approximately $4 billion building Watson Health through acquisitions like Merge Healthcare, Phytel, Truven Health Analytics, and Explorys. By 2022, they sold the entire division to private equity firm Francisco Partners for an undisclosed sum—reportedly around $1 billion. The business was generating about $1 billion in annual revenue but had never turned a profit.

The Uncomfortable Truth About Funding Innovation

Here’s what nobody wants to say publicly: if IBM had been honest in 2011 about Watson’s limitations, it wouldn’t have gotten the funding.

Imagine the pitch: “Watson can answer Jeopardy questions, but it’ll need another decade of development, different data architectures, and fundamental advances in natural language understanding before it can safely recommend cancer treatments.”

Markets reward boldness. Boards reward vision. Investors fund transformation, not incremental improvement. So companies shoot for the moon even when they know—or should know—that the infrastructure isn’t ready.

And then you’re trapped. You’ve made promises you can’t walk back without destroying the trust you need to actually build the thing. Walking back bold claims destroys more value than admitting uncertainty from the start.

We’ve Seen This Pattern Before

Watson isn’t unique. This pattern repeats across technology cycles:

The Dot-Com Era (1999-2000): My startup was part of this wave. Hundreds of companies promised to revolutionize industries with internet technology. Many visions were directionally correct—we do have telehealth, digital payments, and e-commerce infrastructure today. But the technology stack, network speeds, consumer adoption curves, and integration capabilities weren’t ready in 1999. Companies burned through billions of dollars trying to force the future to align with investor timelines rather than technological readiness.

Virtual Reality (2014-2016): Facebook acquired Oculus for $2 billion in 2014, triggering an investment frenzy. By 2016—dubbed “The Year of Virtual Reality”—Oculus Rift, HTC Vive, and PlayStation VR all launched to significant fanfare. Mark Zuckerberg promised a billion people in VR. By 2018, the industry hit what analysts called the “Trough of Disillusionment.” Consumer adoption lagged. Content libraries remained thin. The infrastructure (affordable headsets, comfortable hardware, compelling use cases beyond gaming) wasn’t there yet.

Current AI Wave (2023-Present): MIT‑linked research and subsequent analyses suggest that roughly 95% of enterprise generative‑AI pilots fail to deliver clearly measurable ROI, leaving only about 1 in 20 that show a meaningful business impact. The issue is less about model accuracy and more about what the researchers describe as a “learning gap.” Organizations struggle to integrate AI into real workflows and connect it to the data and processes that actually drive value. In parallel, a recent industry survey found that only about 9% of organizations report that all their data is accessible and usable for AI initiatives, and 37% cite data integration as their top technical challenge.

The pattern is consistent: Promise transformation → Discover infrastructure gaps → Choose between walking back claims (destroying trust) or doubling down (risking capital).

How IBM Is Approaching AI Differently Now

What’s interesting about IBM’s recent acquisitions—DataStax (announced February 2025) and Seek AI (acquired June 2025)—isn’t that IBM is back in AI. It’s how they’re back.

No moonshots. No “cure cancer” promises. Just infrastructure.

DataStax provides NoSQL and vector database capabilities powered by Apache Cassandra, as well as Langflow for low-code AI application development. The pitch: We’ll help you access the unstructured data your AI can’t reach. It’s technical, specific, and boring.

Seek AI converts natural-language questions into database queries to generate business insights. The pitch: Allow business users to query their databases in natural language, without writing SQL. Again, boring yet accurate.

It’s so boring, it’s almost brave.

These acquisitions address the actual bottlenecks companies face: data accessibility, query capabilities, and integration challenges. They’re solving infrastructure problems that don’t make headlines but determine whether AI initiatives succeed or fail.

When a company with IBM’s history pivots from “cure cancer with AI” to “help you access your unstructured data,” that’s not retreat. It’s a pivot to a credible strategy. They learned that trust compounds faster through delivered infrastructure than promised transformation.

The Real Question for Your Organization

The CEOs I’m talking to now face the same pressure IBM faced in 2011: Promise big to win board support, or stay within what you know you can deliver and risk looking timid.

There’s a specific kind of loneliness when you’re the person in the room who knows the timeline is wrong. Someone always knows. The question is whether anyone listens—or whether the pressure to capture headlines and “heat” silences honest assessment.

IBM was bold in 2011 because it was pioneering. We have more data now—dot-com in 1999, virtual reality in 2016—but we won’t know about today’s AI wave for years.

So here are the questions worth asking:

1. Is there infrastructure that can carry the weight of your promises? Not “could it theoretically?” but “does it currently?” If the answer is “not yet,” how long will it actually take? Double your estimate, then add six months for integration challenges you haven’t anticipated.

2. How does your solution rely on capabilities you don’t control? Watson depended on medical data standardization, clinical workflow integration, and regulatory acceptance—none of which IBM could force into existence on its timeline.

3. If you had to be brutally honest about limitations, would you still get funded? If the answer is no, you’re probably overpromising. The question becomes: Is it worth the trust you’ll destroy when reality hits?

4. Are you solving infrastructure problems or selling transformation? Today, the companies making AI work are solving boring technical problems: data accessibility, integration challenges, and query optimization. The ones struggling are still selling transformation without infrastructure to deliver it.

The Infrastructure Comes First

I learned this watching my startup’s valuation evaporate. IBM learned it by spending $4 billion on Watson Health. The lesson is consistent across technology cycles: Infrastructure and delivered capability before promised transformation. Boring reliability before sexy innovation.

The gap between what technology can do and what we promise it will do is where value gets destroyed—not just financial value, but the trust needed to build ambitious things in the first place.

The question isn’t whether AI will transform business. It will. The question is whether your organization’s promises are outpacing its infrastructure.

Because once you’ve sold the dream, walking it back destroys more value than admitting you don’t have all the answers yet.

---

What landmines are hiding beneath your boldest promises? I’m interviewing CEOs about trust as a competitive advantage. If you’re navigating this tension, I’d like to hear your story. Reply to this email or leave a comment below.

---