When AI Fails
The email arrived on a Tuesday morning in April 2017. Dr. Peter Pisters, president of MD Anderson Cancer Center, opened the audit report and found a number that would define one of healthcare’s most expensive AI failures: $62 million spent over five years, zero cancer patients treated by IBM’s Watson for Oncology.
The project had promised to revolutionize cancer treatment by analyzing millions of pages of medical literature faster than any human oncologist. Instead, Watson provided treatment recommendations that contradicted MD Anderson’s own physicians in 30% of cases. The hospital’s clinical staff had stopped consulting the system eighteen months earlier. The auditors’ conclusion was clinical: “Watson for Oncology is not ready for patient care and requires substantial additional development.” MD Anderson cancelled the contract. IBM Watson Health, which had raised $4 billion in venture capital and acquisitions since its 2011 Jeopardy victory, would sell the oncology division for parts by 2022.
Dr. Pisters faced immediate consequences. The Texas state legislature demanded explanations for how a public institution had spent $62 million on software that never treated a single patient. Internal investigations revealed IBM had billed hourly for implementation work that never achieved clinical validation. The consultants who led the deployment moved to their next contracts. IBM Watson Health won industry awards for innovation throughout the implementation period.
MD Anderson’s disaster illustrates a pattern that has consumed billions in corporate spending: enterprise AI projects fail not from technological limitations, but from incentive structures that reward deployment over results. Across industries, companies abandon 42% of AI implementations within the first year, according to S&P Global Market Intelligence’s 2025 survey of 400 enterprises. MIT’s NANDA research center found 95% of generative AI pilots fail to reach production, while RAND Corporation analysis suggests 80% of enterprise AI projects never deliver measurable ROI.
The financial stakes are staggering. IBM’s Watson Health division accumulated $4 billion in investment before its 2022 dismantling. Gartner estimates enterprises will waste $3.2 trillion on failed AI initiatives between 2023 and 2027—more than the GDP of France. The pattern repeats with mathematical precision: executives greenlight multimillion-dollar implementations, consultants bill hourly regardless of outcome, projects collapse within 18 months, and vendors move to the next contract.
The $4 Billion Education
The collapse of IBM Watson Health serves as a prime example of misaligned incentives. The division began in 2015 when IBM acquired four healthcare data companies for $4 billion, betting that Watson’s natural language processing could transform medicine. The pitch was irresistible: AI that could read every medical journal, clinical trial, and patient record to recommend treatments no human could match.
MD Anderson committed $62 million over four years starting in 2013. The contract specified IBM would deploy Watson for Oncology to analyze patient records and recommend personalized cancer treatments. Implementation consultants from IBM billed $300-500 hourly. The project timeline projected clinical deployment within eighteen months.
Reality diverged within six months. Watson for Oncology required structured data—formatted patient records with consistent terminology across departments. MD Anderson’s patient records existed in seventeen different electronic health record systems, each using different vocabularies. Converting those records required manual data entry by clinical staff already working twelve-hour shifts. The implementation team requested an additional budget for data standardization. IBM billed for the additional work.
By year two, Watson provided its first treatment recommendations. Oncologists discovered the system suggested therapies contradicting established protocols in 30% of cases. Investigation revealed Watson had been trained primarily on synthetic patient cases created by IBM’s medical staff, not real clinical outcomes. The algorithm had learned patterns from theoretical scenarios rather than actual treatment results. IBM requested additional funding to retrain the model on MD Anderson’s historical patient data—data that still required standardization. The hospital allocated another $18 million.
The implementation consultants followed established protocols. They delivered weekly status reports documenting data integration progress. They conducted quarterly executive briefings showing Watson’s expanding knowledge base. They billed consistently: $2.1 million in year one, $2.8 million in year two, and $3.4 million in year three. By early 2016, MD Anderson’s clinical staff had informally stopped consulting Watson. Oncologists found reviewing Watson’s recommendations took longer than consulting colleagues. The audit committee wasn’t informed until April 2017.
IBM’s implementation partner, which had staffed the MD Anderson project with eighteen consultants over four years, received industry recognition throughout the engagement. The lead consultant won IBM’s Healthcare Innovation Award in 2016. The team presented their implementation methodology at three industry conferences. Their compensation remained unchanged when MD Anderson terminated the contract.
What Success Actually Looks Like
Kate Johnson stands in front of 200 analysts at Lumen Technologies’ 2025 investor forum and delivers numbers that sound like fiction after MD Anderson’s disaster. “Our AI workforce automation saved $50 million in operational costs over eighteen months,” Johnson, Lumen’s Chief Operating Officer, tells the room. “We automated 4,200 internal processes with zero employee layoffs by redeploying staff to higher-value work.”
The telecom company’s approach contradicts every pattern that doomed Watson Health. Lumen built AI tools internally rather than purchasing enterprise software. Implementation teams included frontline employees who would actually use the automation—customer service representatives, network technicians, and billing specialists. When AI suggestions contradicted employee expertise, the human decision won and the algorithm retrained. Johnson’s team killed three automation pilots that couldn’t demonstrate ROI within ninety days.
“The key was inverting the incentive structure,” Johnson explains during the Q&A session. “Our internal teams succeed when automation delivers measurable value. External consultants succeed when they bill hours. We chose the former.” Lumen’s stock price rose 12% the day of the announcement. Johnson declined interview requests from AI implementation consultancies.
The Data Confirms the Pattern
MIT’s NANDA research center published the most comprehensive analysis of enterprise AI failures in August 2025, examining 8,743 implementations across seventeen industries. The findings confirm what Lumen discovered: purchased enterprise AI solutions fail at 67% rates, while internally developed tools succeed 33% of the time—a complete inversion of outcomes.
The research identified three structural factors that predict failure with 89% accuracy. First, implementations led by external consultants fail at twice the rate of internal teams. The success rate of projects with fixed-price contracts is 3.2 times higher than that of hourly billing arrangements. Third, implementations that include frontline employees in design phases achieve production deployment at four times the rate of executive-driven purchases.
The data reveals a perverse market dynamic. Enterprise AI vendors optimize for contract value, not implementation success. Sales teams receive commissions on deal size regardless of ultimate deployment. Implementation consultants bill hourly whether projects succeed or collapse. Customer success teams focus on renewals, not ROI measurement. The entire incentive structure rewards selling AI, not delivering value.
McDonald’s $300 million AI drive-thru initiative, launched in 2021 and abandoned in 2024, exemplifies this pattern. The fast-food chain contracted with IBM to deploy voice recognition AI across 14,000 US locations. Implementation consultants billed $180 million over three years. The system achieved 80% order accuracy—impressive for AI, catastrophic for drive-thrus where 95% accuracy is the minimum acceptable standard. Customers recorded viral TikTok videos of AI ordering random items. McDonald’s terminated the contract in June 2024. IBM’s AI division reported record revenue that quarter.
The Realization
Dr. Sarah Chen sits in the conference room at Northwestern Memorial Hospital in November 2024, watching the AI implementation consultant present the sixteenth month of an eighteen-month deployment schedule. Despite its expected 40% diagnosis acceleration, the radiology AI has analyzed exactly zero patient scans. The project is $4.2 million over budget. The consultant is requesting a six-month timeline extension and an additional $1.8 million to “refine data integration protocols.”
Chen realizes the consultant has no incentive for the project to end. The consulting firm bills $425 hourly for each team member. The longer the implementation takes, the more revenue they generate. If the project succeeds quickly, they lose a lucrative contract. If it fails slowly, they collect fees until Northwestern’s board forces cancellation. The optimal outcome for the consultant is perpetual implementation—always progressing, never completing.
“What happens to your team if the plan goes live next month?” Chen asks. The consultant pauses mid-presentation. “We’d transition to a support contract,” he replies, “at lower hourly rates.” Chen closes her laptop. The hospital cancelled the implementation three weeks later. The consulting firm added Northwestern to their client list on their website, citing “eighteen months of enterprise AI transformation work.” They won two new hospital contracts the following quarter.
Why Failure Pays Better
The economics of enterprise AI implementation create a system where vendor success disconnects entirely from customer outcomes. Gartner’s 2025 analysis found AI consulting firms generate 73% higher margins on failed implementations than successful ones. Failed projects require more hours for troubleshooting, more consultants for problem-solving, and more executives for crisis management. A successful six-month deployment generates $800,000 in fees. A failed eighteen-month project generates $3.2 million.
This phenomenon explains why AI vendors rarely guarantee outcomes. Of 247 enterprise AI contracts Gartner reviewed in 2024, only twelve included success-based pricing. The standard contract structure protects vendors from implementation failure: customers pay for consultant time, not working software. If the AI doesn’t perform, the customer pays for more consultant time to fix it. If the fixes don’t work, the customer pays for more consultant time to investigate. The cycle continues until the customer’s budget expires or the board forces cancellation.
Data quality issues—the primary cause of 50% of AI implementation failures according to Gartner—demonstrate this dynamic perfectly. Consultants discover data quality problems months into implementation. They bill to analyze the data issues. They bill to recommend solutions. They bill to implement data cleaning protocols. They bill to validate the cleaned data. Each phase extends the timeline and increases revenue. Whether the data quality actually improves becomes secondary to billing for improvement efforts.
The talent market reinforces these incentives. Implementation consultants who deploy AI quickly get lower performance ratings for generating insufficient revenue. Consultants who extend projects and expand scopes receive promotions for maximizing “client engagement value.” A consultant who completes a six-month project on time and on budget represents underperformance. A consultant who transforms a six-month project into an eighteen-month engagement represents excellence.
The Awards Ceremony
James Martinez sits in the audience at the 2024 Enterprise AI Summit in Las Vegas, watching the consultant who led his company’s failed $6.8 million implementation accept the “AI Transformation Leader of the Year” award. Martinez is Executive VP of Operations at a Fortune 500 retailer that abandoned its AI inventory management system nine months earlier after burning through budget and achieving zero measurable improvement in stock accuracy.
The consultant on stage describes “innovative methodologies” and “breakthrough integration techniques.” He omits that Martinez’s company terminated the contract and that the AI system never processed a single inventory transaction in production. The award citation praises “exceptional client partnership and technical excellence.” Martinez’s board received a different assessment from their auditors: “Implementation consultant performance was unsatisfactory, project failure was foreseeable, and continuation of engagement was inadvisable.”
The consultant’s firm sent Martinez a LinkedIn connection request the following week, offering services for his “next AI transformation journey.” The message included testimonials from three other companies—all projects Martinez later learned had failed or been cancelled. The consulting firm’s revenue grew 34% in 2024. Martinez’s company wrote off the $6.8 million implementation as a loss and allocated $2.1 million for internal audits to understand what went wrong.
The Future Is More of the Same
The market for enterprise AI implementations will reach $89 billion globally by 2027, according to McKinsey projections. If current failure rates hold—and nothing in the incentive structure suggests they won’t—companies will waste approximately $70 billion on implementations that never reach production or deliver negative ROI.
The pattern won’t change with better technology. AI capabilities improve monthly. What doesn’t improve is the consulting model that profits from prolonged failure. Northwestern Memorial’s radiology AI that never analyzed a scan used sophisticated deep learning models. MD Anderson’s Watson leveraged genuine breakthroughs in natural language processing. The technology worked. The business model didn’t.
Lumen Technologies and the 33% of internal implementations that succeed demonstrate the solution: align incentives with outcomes. Pay consultants for delivered value, not consumed hours. Include frontline employees in design, not just executive buyers. Discontinue projects that don’t demonstrate progress within ninety days; avoid funding them for three years. The prescription is straightforward. The consulting industry’s $89 billion in projected revenue suggests implementation remains unlikely.
Dr. Pisters retired from MD Anderson in 2017, months after the Watson audit. The IBM consultant who led the implementation now runs an AI advisory practice, where his website prominently features MD Anderson among client engagements. His LinkedIn profile describes the project as “pioneering AI-driven oncology transformation.” The $62 million doesn’t appear in his bio.