Most AI Projects Are Failed by the Organisation, Not the Model

In 2021 Zillow's Homes segment, which ran its Zillow Offers home-buying business, lost $881 million before income taxes, according to its 2021 annual report. In the third quarter alone Zillow bought 9,680 homes. Its shareholder letter said it had been "unintentionally purchasing homes at higher prices than our current estimates of future selling prices", and its results release took a write-down of about $304 million. Its chief executive, Rich Barton, said "the unpredictability in forecasting home prices far exceeds what we anticipated". Zillow announced the wind-down on 2 November 2021.

Read quickly, that is a model failure, and Barton's own words say the forecasts were off. But Stanford's Amit Seru has described Zillow bidding more for some homes than its own model predicted, to compete. What should have stopped it sooner is the subject of a companion piece on when to stop an AI deployment.

My argument is that most AI projects that fail are failed by the organisation around the model. Every model is wrong some of the time. Whether a wrong answer becomes a loss is decided by the organisation, and organisations get it wrong in seven repeatable patterns.

An AI failure pattern is a repeatable way an organisation turns a model's errors into a loss, through the problem it picks, the data it trusts, the owner it never names or the limits it never sets. Each pattern leaves a sign a leader can check before the loss.

How Often Do AI Projects Fail, and Why?

Nobody has a clean failure rate, and not every abandoned pilot is a failure. One stopped early and cheaply is the system working. RAND, citing outside estimates, puts AI project failure above 80 percent, twice the rate for IT projects without AI. Surveys are more specific. In S&P Global Market Intelligence's 2025 survey of 1,006 IT and business professionals in North America and Europe, 42% of companies had abandoned the majority of their AI initiatives before production, up from 17% a year earlier, and on average 46% of projects were scrapped between proof of concept and broad adoption.

The causes matter more than the rate. RAND interviewed 65 experienced data scientists and engineers and found five leading root causes: the problem misunderstood or miscommunicated; too little data to train the model; a focus on the latest technology over the user's problem; too little infrastructure to manage data and deploy models; and problems too hard for AI. On my reading four of the five are organisational, and 84% of the interviewees cited a leadership-driven root cause as the main reason projects fail. RAND's own caveat applies: most interviewees were engineers, not executives, so the answers may lean toward blaming leadership. Executives point the same way. BCG's 2024 survey of 1,000 senior executives attributes around 70% of AI implementation challenges to people and process, 20% to technology and only 10% to the algorithms.

What Are the Seven Patterns?

1. The Shiny Object

Leadership starts with the technology and looks for a use. Pilots get attention and succeed. Nothing reaches production, because nobody defined what working means for the business. Two of RAND's root causes describe it: the problem misunderstood or miscommunicated, and a focus on "the latest and greatest technology" over solving real problems for users. The sign: pilots with no business owner, no production date and no measure of success agreed before they began.

2. The Data Mirage

The data looks ready and is not. By mid-2016 IBM's Watson Health had acquired four health-data companies for about $4 billion, IEEE Spectrum reported. Internal IBM documents reported by STAT in 2018 showed Watson for Oncology had been trained on a small number of "synthetic" cancer cases, hypothetical patients. MD Anderson's own, separate Watson-based project had paid about $62.1 million to outside firms by August 2016, a University of Texas procurement audit found, before it was put on hold. Amazon's recruiting model, trained on ten years of résumés mostly from men, learned to penalise the word "women's". Gartner expects organisations to abandon 60% of AI projects that lack AI-ready data through 2026. The sign: training data nobody outside the build team has checked against the cases the system will meet.

3. The Talent Trap

Organisations hire specialists before they build what those specialists need. RAND's fourth root cause is that organisations "might not have adequate infrastructure to manage their data and deploy completed AI models". Skilled staff spend their weeks moving data by hand, and when the few who know how leave, the project stalls. I know of no reliable figure for how often this happens, so treat this pattern as my judgment built on RAND's finding. The sign: a system that would stall if one engineer left.

4. The Governance Ghost

A system goes live with nobody answerable for what it says. Air Canada's website chatbot told Jake Moffatt he could apply for a bereavement fare retroactively, which the airline did not allow. Before the British Columbia Civil Resolution Tribunal, Air Canada argued it could not be held liable for information from its agents, including a chatbot; in effect, the tribunal said, it suggested the chatbot was "a separate legal entity that is responsible for its own actions". The tribunal rejected that on 14 February 2024: "It should be obvious to Air Canada that it is responsible for all the information on its website." It ordered the airline to pay CAD $812.02. The sum is small. The precedent, that a company answers for what its chatbot says, is what counts. The sign: a live AI system whose owner nobody can name.

5. The Autonomy Illusion

Automation is handed more of the job and judged on cost. In February 2024 Klarna said its AI assistant was handling two-thirds of its customer service chats, "the equivalent work of 700 full-time agents". By May 2025 it was recruiting human agents again. Sebastian Siemiatkowski, the chief executive, told Bloomberg, as Fortune reported: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." Klarna told Fortune it was "very much still AI-first". The sign: an automation case measured on cost per contact with no quality measure beside it.

6. The Compliance Cliff

Regulation is left until the deadline, and the work done then looks compliant without changing anything. The EU AI Act's rules for the high-risk uses listed in Annex III, which include hiring and credit scoring, apply from 2 December 2027, moved from 2 August 2026 by the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026. Breaching the obligations listed in Article 99(4) can cost up to EUR 15 million or 3% of worldwide annual turnover. The prohibited practices in Article 5 carry up to EUR 35 million or 7%. This pattern has no enforcement case under the Act yet, so treat it as a forecast rather than a record. It is on the list because the date is set in law, though it has already moved once. The sign: no list of which of your systems the high-risk rules cover.

7. The Change Resistance Spiral

A sound framework meets an organisation that never changed how it works. Engineering routes around it and the board stops funding what it cannot see working. BCG says the companies getting value from AI follow a rule of putting 10% of their resources into algorithms, 20% into technology and data, and 70% into people and processes. The sign: governance policies that exist, and gate reviews that have never met.

How Fast Do Failures Spread?

Failures also travel. Charles Perrow's Normal Accidents (1984) described systems whose parts interact in ways their designers did not expect and are tightly coupled, so one failure spreads faster than people can follow it. AI deployments have both traits. Google's Gemini image feature shows how short the chain can be.

Google's own account gives two causes: the tuning above, and a model that "became way more cautious than we intended". From the pause it took four days to reach the share price and five to reach the chief executive's memo to staff. No single step was large. Together they ended in a one-day 4.5% fall that Forbes tied to the controversy.

Seven Questions for Your Next Portfolio Review

1. Does every pilot have a business owner, a production date and a success measure agreed before it started? (Shiny Object)
2. Has anyone outside the build team checked the training data against the people and cases the system will meet? (Data Mirage)
3. Which AI system would stall if one engineer left? (Talent Trap)
4. Can you name the owner of every AI system in production? (Governance Ghost)
5. Is every automation case measured on quality for the people it serves, as well as on cost? (Autonomy Illusion)
6. Do you have a list of systems the EU AI Act's high-risk rules will cover from December 2027? (Compliance Cliff)
7. When did your last gate review actually meet, and what did it change? (Change Resistance Spiral)

A no, a blank or a vague answer on any of these is a reason to look closer, not a verdict. When a closer look confirms a pattern in a live system, someone needs written stop criteria and the authority to pause it. That is the subject of Nobody Tells You When to Stop an AI Deployment.

“Before they begin any AI project, leaders should be prepared to commit each product team to solving a specific problem for at least a year.”

Sources and Changes

• Zillow Group, Form 10-K for 2021, filed 10 February 2022; Q3 2021 results and shareholder letter, Form 8-K, 2 November 2021.
• Stanford Graduate School of Business, "Flip Flop: Why Zillow's Algorithmic Home Buying Venture Imploded", 9 December 2021.
• RAND, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (RR-A2680-1), 13 August 2024.
• S&P Global Market Intelligence, Voice of the Enterprise: AI and Machine Learning, Use Cases 2025, 30 May 2025.
• Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk", 26 February 2025.
• BCG, "AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value", 24 October 2024.
• IEEE Spectrum, "How IBM Watson Overpromised and Underdelivered on AI Health Care", 2 April 2019; STAT, "IBM's Watson supercomputer recommended 'unsafe and incorrect' cancer treatments", 25 July 2018; University of Texas System, special review of the MD Anderson Oncology Expert Advisor project, November 2016; Forbes, "MD Anderson Benches IBM Watson", 19 February 2017.
• Reuters, "Amazon scraps secret AI recruiting tool that showed bias against women", 10 October 2018.
• British Columbia Civil Resolution Tribunal, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024.
• Klarna, "Klarna AI assistant handles two-thirds of customer service chats in its first month", 27 February 2024; Fortune, "As Klarna flips from AI-first to hiring people again", 9 May 2025.
• Regulation (EU) 2024/1689 (the AI Act), Article 99, and Regulation (EU) 2026/1744 (the Digital Omnibus), Official Journal texts, read 6 October 2026.
• Charles Perrow, Normal Accidents: Living with High-Risk Technologies, Basic Books, 1984.
• Google, "Gemini image generation got it wrong. We'll do better.", 23 February 2024; Forbes, "Google's Gemini Headaches Spur $90 Billion Selloff", 26 February 2024; Semafor, Pichai memo, 27 February 2024.

Changes on 6 October 2026. Cut from about 4,900 words, with a new headline. Zillow's loss is now the $881 million segment figure from its 2021 annual report, not the earlier $569 million estimate. Corrected: Klarna's 700-agent claim, a quotation from its chief executive, and Air Canada's "separate legal entity". Removed for want of a primary source: per-project loss figures, a multi-agent "87%" claim, two composed examples, and the forecast and startup sections. Seven custom diagrams and a dead download link went too.