IBM Watson Health
The $4B lesson in what happens when you sell AI before you solve data
What They Said
IBM positioned Watson as the future of healthcare AI. The marketing was breathtaking in its ambition: Watson would “revolutionize cancer treatment,” “transform drug discovery,” and “make precision medicine accessible to every doctor.” CEO Ginni Rometty called it the company’s “moonshot.”
Between 2015 and 2017, IBM spent over $4B acquiring healthcare data companies — Truven Health Analytics ($2.6B), Merge Healthcare ($1B), and Phytel — to feed Watson’s appetite for medical data. The acquisitions signaled a clear strategy: buy the data, build the AI, sell the insights.
What Actually Happened
The gap between Watson’s marketing and its capability was staggering. Multiple hospitals that deployed Watson for Oncology reported that it produced treatment recommendations that were “unsafe and incorrect.” MD Anderson Cancer Center’s $62M Watson project was shelved after an internal audit found the system was trained on a small number of synthetic cases rather than real patient data.
The core problem wasn’t the AI models — it was the data. Healthcare data is fragmented across thousands of incompatible electronic health record systems, riddled with inconsistencies, missing critical context that only exists in doctors’ handwritten notes, and governed by privacy regulations that make aggregation enormously difficult. IBM’s acquisitions bought data assets, but those assets required years of cleaning, standardizing, and contextualizing before they could train reliable models.
IBM tried to skip that step. They sold Watson to hospitals before the data foundation was ready, creating a devastating cycle: hospitals expected results, Watson couldn’t deliver them, negative press accumulated, and the product’s reputation cratered before the underlying technology had a chance to mature.
The Root Cause
Data debt is invisible until you try to build on it. IBM’s healthcare data acquisitions looked transformative on paper — millions of patient records, decades of clinical history. But the data was spread across incompatible formats, inconsistent coding schemes, and varying quality standards. Cleaning and standardizing this data was a multi-year engineering project that IBM underestimated by at least 3x.
The second failure was organizational: the sales team was incentivized to close hospital deals, not to ensure deployment readiness. Watson was sold into environments where the data infrastructure couldn’t support it, the clinical staff wasn’t trained to use it, and the expectations were set by marketing materials rather than technical reality.
The Pattern to Watch For
If your AI vendor can’t show you the data pipeline — not the model, the pipeline — walk away. Every enterprise AI failure in healthcare traces back to the same root: someone assumed the data was ready when it wasn’t. The model is the easy part. The data quality, standardization, and governance infrastructure is where healthcare AI projects actually live or die.
What You Should Steal
MD Anderson’s internal audit process — the one that killed the Watson project — is actually a best practice worth replicating. They asked three questions: What data was this model actually trained on? How does that data compare to our patient population? What happens when the model encounters a case outside its training distribution? Those three questions would have saved IBM $4B if they’d asked them first.