Most AI initiatives fail not because the models are wrong but because the data beneath them was never prepared for the job
PixeloneStocker / Getty Images
One of the world's largest PC manufacturers spent seven years collecting telemetry data from its devices. When Sachin Dharmapurikar's team at The Modern Data Company finally examined it, they discovered that two of the 70 fields had been recorded incorrectly the entire time. No one had checked. The data just sat there, accumulating volume and cost but not value.
That anecdote captures a pattern playing out across corporate America. Companies spent the last decade stockpiling data in the cloud, betting that more information would eventually yield better decisions. Then ChatGPT arrived, and executives assumed they could feed all that stored data into a large language model and watch insights appear. Dharmapurikar calls this the "ChatGPT curse."
The results have been bad. According to McKinsey's 2025 Global AI Survey, which included about 2,000 participants across 105 countries, 88% of organizations now use AI in at least one business function, but just 39% report any enterprise-level EBIT impact. A RAND Corporation study found that more than 80% of AI projects fail, roughly twice the failure rate of IT projects that do not involve AI. The technology is not usually the problem. The data is.
Dharmapurikar identifies four prerequisites that enterprises routinely skip: Data quality at scale, lineage traceability, governance, and semantic metadata. Each one is a distinct failure point.
The first prerequisite is the most obvious and the most neglected. Gartner estimates that poor data quality costs organizations an average of $12.9 million per year. According to a 2020 Gartner survey, 59% of organizations do not even measure data quality.
AI magnifies the consequences. A flawed report can be caught and corrected by a human analyst. A flawed model trained on the same data will produce errors at machine speed, at scale, with no one checking the math. Informatica's CDO Insights 2025 survey of 600 data leaders found that 43% cited data quality, completeness, and readiness among the top obstacles preventing generative AI initiatives from moving from pilot to production.
The costs of data quality failures are not hypothetical. In 2022, Unity Technologies suffered a $110 million revenue loss in part due to its ad-targeting platform ingesting corrupted data from a large client, compromising the machine learning algorithms powering its Audience Pinpointer tool and resulting in an 8% hit to annual earnings. Unity's own SEC filings described the problem as "the consequences of ingesting bad data from a large customer" and noted the company had to redirect resources to address data quality challenges.
The lesson is direct: AI models are only as reliable as the data they consume. Volume is not the same as quality.
When an AI model produces an answer, a business needs to know how it arrived at that answer. What data went in? Where did that data originate? What transformations did it undergo? Without this chain of custody, there is no way to debug errors, satisfy regulators, or build trust with stakeholders.
This is the problem of lineage traceability. A Capital One $COF-commissioned Forrester study of data decision-makers found that 73% cited transparency, traceability, and explainability of data flows as a top challenge for machine learning deployments.
Consider what happened with Zillow. Zillow's AI-driven iBuying program collapsed in 2021 when its Zestimate algorithm overvalued properties in volatile markets, resulting in more than $500 million in write-downs and the shutdown of the entire business unit. The company could build a model. It could not trace whether the inputs feeding that model still reflected reality.
Without lineage, there is no early warning system. Errors compound silently until they become financial crises.
Governance, in Dharmapurikar's framework, is not a compliance exercise. It is the set of rules and controls that determine which data an AI system can access, how that data is classified, and what happens when the model encounters gaps.
Without governance, AI models fill in blanks with plausible-sounding nonsense. A Gartner survey of data management leaders found that 63% of organizations either lack or are unsure whether they have the right data management practices for AI. Gartner predicts that, through 2026, organizations will abandon 60% of AI projects that lack AI-ready data.
BCG's October 2024 report, "Where's the Value in AI?," surveyed 1,000 senior executives across 59 countries and found that 74% of companies have yet to show tangible value from AI, with about 70% of implementation challenges stemming from people- and process-related issues rather than technology. Governance sits at the intersection of people and process. It requires someone to own data classification, set access policies, and enforce standards across departments. In most organizations, ownership does not exist.
Sales teams, manufacturing teams, and web teams each collect data in their own silos, and transferring information between departments requires bureaucracy. AI needs data from across an organization. The reality, as Dharmapurikar's work demonstrates, is fragmented systems that do not communicate with one another.
The fourth prerequisite is the least intuitive but may be the most consequential. Semantic metadata is the layer of business context that tells an AI model what a data field actually means.
A column labeled "customer_LTV" in a retail dataset means something different from the same label in an enterprise software database. The lifetime value of a consumer who buys running shoes is measured in hundreds of dollars over several years. The lifetime value of a Fortune 500 client on a multi-year software contract is measured in millions. Without metadata that captures this distinction, a model trained on both datasets will produce averages that neither customer accurately describes.
This problem extends to basic terms. Definitions of fundamental concepts such as "customer," "order," and "revenue" differ between departments in most organizations. Historical records contain gaps, inconsistencies, and formatting mismatches. "Revenue" in one system might mean booked orders; in another, shipped and invoiced; in a third, recognized revenue under ASC 606 rules.
Two-thirds of data leaders surveyed by Informatica revealed they have not been able to transition even half of their generative AI pilots to production. The semantic layer is a significant reason why. Pilot projects work on small, curated datasets where the business context is understood by the team running the experiment. Production deployments pull from enterprise-wide data stores where no such shared understanding exists.
"People are now more calculated, more rational and pragmatic about this stuff," Dharmapurikar told Quartz. "The reality is kicking in big time that there is no easy solution."
The data support his assessment. According to S&P Global $SPGI Market Intelligence's 2025 survey of more than 1,000 enterprises, 42% of companies abandoned most of their AI initiatives, up from 17% in 2024, and the average organization scrapped 46% of AI proofs of concept before they reached production.
The four prerequisites Dharmapurikar identifies are not a wish list. They are a diagnostic. Companies that skip data quality will train models on noise. Companies that skip lineage will be unable to explain or debug outputs. Companies that skip governance will watch models hallucinate. Companies that skip semantic metadata will generate conclusions that sound right but mean nothing.
BCG's research found that AI leaders allocate resources following a 10-20-70 rule: 10% to algorithms, 20% to technology and data, and 70% to people and processes. The companies that succeed at AI do not start with the model. They start with the data beneath it.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
