In June 2012, MD Anderson Cancer Center signed a contract with IBM to build the Oncology Expert Advisor (OEA), a system that would use Watson's cognitive computing capabilities to help oncologists analyze patient records, identify treatment options, and match patients to clinical trials. The original contract was scoped at six months and $2.4 million. By the time the project was terminated, the contract had been extended twelve times, IBM had been paid $39.2 million, and PricewaterhouseCoopers had drawn another $23 million in consulting and project support. The total cost was approximately $62 million. The system never entered clinical use.

What it was supposed to do

The OEA was designed to serve as an AI-powered decision support tool for cancer treatment. Watson would ingest a patient's medical records - lab results, imaging reports, pathology findings, treatment history - and cross-reference them against medical literature, clinical trial databases, and MD Anderson's own treatment protocols to generate ranked treatment recommendations. The vision was ambitious: a system that could synthesize more medical literature than any human oncologist could read, applied to individual patient cases in minutes.

MD Anderson is one of the world's premier cancer treatment centers. If Watson could prove itself there, the commercial potential was enormous. IBM marketed Watson Health aggressively during this period, positioning the technology as a transformative force in medical decision-making. The MD Anderson partnership was a flagship initiative.

Internal pilots for leukemia and lung cancer were conducted. Physicians involved in the pilots gave positive feedback. In a June 2015 Washington Post article, MD Anderson assistant professor of leukemia Koichi Takahashi described Watson's ability to synthesize a patient's history as "amazing. He beats me." Another physician, Candida Vitale, said she was surprised by the system's capacity to compile information: "Even if you work all night, it would be impossible to be able to put this much information together like that."

Fifteen months after those positive reviews, IBM ended support for the OEA.

What went wrong

The University of Texas System audit, released in February 2017, identified a convergence of procurement failures, scope management problems, and a critical technical gap that collectively sank the project.

Procurement. The audit's primary focus was on how the $62 million was spent, not on Watson's technical performance. The original contract was extended twelve times without competitive rebidding. PricewaterhouseCoopers was brought in to create a "Business Plan for a Flagship Informatics Tool" and to lead capability assessments. The audit found that standard procurement processes were bypassed, and the project was overseen primarily by Dr. Lydia Chin, the program director, without adequate institutional oversight.

The financial findings were stark. Gift funds designated for the project had a deficit balance of $11.59 million as of August 31, 2016. MD Anderson had spent donor money it had not yet received. For a cancer center already dealing with financial pressures that would lead to significant job cuts, this deficit was a serious governance problem.

Scope creep. The project started with a six-month timeline and a focused scope. Over four years of extensions, the scope expanded beyond what the original design could support. The system was piloted for leukemia and lung cancer, but the broader goal of covering multiple cancer types, integrating with clinical trial matching, and producing treatment recommendations ready for clinical use proved far more complex than the initial contract anticipated.

Epic integration. The most concrete technical failure was the simplest. MD Anderson's internal pilots of the OEA were built on ClinicStation, the hospital's previous electronic medical records system. During the course of the Watson project, MD Anderson migrated to Epic, one of the most widely used EHR platforms in American healthcare. The OEA was never updated to work with Epic. A system that cannot read patient records from the hospital's actual records system cannot function as a clinical decision support tool. The audit noted: "OEA has not been updated to integrate with the current system (Epic)."

The Epic migration was not a surprise event. It was a known institutional transition that the Watson project should have accounted for. The failure to plan for it - or to execute the integration work when it became necessary - suggests either a fundamental project management failure or a technical barrier that proved more difficult than expected. Epic's interoperability limitations have been criticized elsewhere, including by MD Anderson itself, which attributed $405 million in costs to the Epic implementation. Whether Epic's architecture made integration difficult or whether the Watson team simply did not prioritize it, the result was the same: the OEA was stranded on a deprecated system.

IBM's response

IBM's official response to the audit was careful. The company argued that "the report assessed procurement practices, not the value or functionality of the OEA system." IBM noted that Watson had achieved 90% accuracy in lung cancer decision support during the pilot, framing the technology as sound even if the procurement and management around it were not.

This was a fair reading of the audit's literal scope. The UT System auditors were investigating how money was spent, not whether Watson's medical recommendations were accurate. But the framing obscured a relevant point: accuracy in a pilot environment using a deprecated records system is not the same as clinical utility. A system that achieves 90% accuracy on historical cases pulled from ClinicStation but cannot read records from Epic is not a viable clinical tool. The technical achievement did not translate into practical value because the infrastructure required to deploy it was never built.

IBM also had commercial reasons to avoid linking the MD Anderson outcome to Watson's capabilities. At the time of the audit, IBM was actively expanding Watson Health partnerships elsewhere. Memorial Sloan Kettering was training Watson for Oncology. Jupiter Medical Center was preparing to deploy Watson for Oncology in clinical settings. The Watson Medical Imaging collaborative had expanded to 24 organizations. IBM did not want one failed deployment to undermine the broader strategy.

The broader Watson Health trajectory

The MD Anderson failure was not an isolated incident. It was an early indicator of a pattern that would repeat across IBM's Watson Health division. The technology consistently showed promise in controlled demonstrations but struggled with the messy realities of clinical deployment: fragmented data systems, inconsistent data quality, the gap between academic accuracy and clinical reliability, and the difficulty of integrating AI recommendations into existing clinical workflows.

IBM's Watson Health division was eventually sold to Francisco Partners, a private equity firm, in 2022 for a reported $1 billion - a fraction of what IBM had invested in it. The sale effectively acknowledged that Watson's healthcare ambitions had not delivered the returns IBM expected.

At MD Anderson, the aftermath followed a familiar institutional pattern. The hospital issued an RFP to restart its AI decision support program and did not exclude IBM from the bidding, leaving open the possibility of a renewed partnership with the company that had just failed to deliver on the first attempt.

What the $62 million bought

The audit's findings map cleanly onto well-known enterprise IT project failure modes. A small initial contract expanded through repeated extensions. Scope grew without proportional controls. A single point of oversight concentrated authority in one program director. Financial commitments outpaced institutional approval. And a fundamental technical dependency - the EHR migration - was either ignored or failed to be addressed.

The Watson-specific element is the gap between what cognitive computing was promised to do and what it could actually deliver in a production healthcare environment. Watson's natural language processing could ingest and synthesize medical literature. It could cross-reference patient data against treatment databases. These are genuinely useful capabilities. But they are not, by themselves, a clinical decision support system. That requires integration with the hospital's actual records, validation against the institution's treatment protocols, regulatory compliance, clinician training, and workflow integration. The technology was one component of a much larger system that was never built.

The $62 million produced internal pilots on a deprecated records system, positive quotes from physicians who tested early versions, and an audit report documenting how the money was spent without adequate oversight. It did not produce a system that could help a single patient at MD Anderson.