Only Seven Percent Say Their Data Is Completely Ready for AI

Data governance for AI is the work of knowing, for each AI use, what data it runs on, whether that data measures what the use needs, who may reach it, where it came from and who answers for it. It has to be done before the feature is funded, because no model is more reliable than the data under it.

My argument is one plenty of executives will push back on: no AI feature that runs on your organization’s data should be funded until that data has been through five tests a team can answer, or fail to answer, in a day. Where it fails, the data work is funded first, in the same decision. Most organizations work the other way round. They fund the feature, build the demo, and meet the data problem in production, where it costs more to fix. I wrote this for the executive who approves AI budgets and keeps hearing that the data will be sorted out later.

What the Surveys Say

In October 2025 Harvard Business Review Analytic Services surveyed 231 people involved in their organizations’ data decisions, for a report sponsored by Cloudera. Their answer to one question sets the scene.

Gartner’s figures point the same way from a larger sample. In a July 2024 survey of 1,203 data management leaders, 63 percent said their organization either did not have, or was not sure it had, the right data management practices for AI. In the same release, in February 2025, Gartner predicted that "through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data." A year earlier it had predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, and named poor data quality among its four reasons, with risk controls, cost and unclear value.

None of this is new. In 2017 Tadhg Nagle, Thomas Redman and David Sammon asked 75 executives to check the last 100 records their departments had created. On average 47 percent of new records had at least one critical error, and only 3 percent of the scores counted as acceptable even by the loosest standard. AI made an old problem expensive, because a model learns from the errors and repeats them in every decision it touches.

The Model Can Be Right and the Data Wrong

The clearest case I know comes from health care. In 2019 Ziad Obermeyer and colleagues studied a commercial algorithm that US health systems used to pick patients for extra care, one they describe as affecting millions of patients. It predicted health care costs, and by some measures of predictive accuracy cost looked like a good proxy for health. But cost stood in for illness. Because less money is spent caring for Black patients than for White patients with the same needs, Black patients at a given risk score were considerably sicker. Removing the disparity would have raised the share of Black patients receiving extra help from 17.7 percent to 46.5 percent. The bias arose, the authors found, because the algorithm predicted costs rather than illness, and they warn that "the choice of convenient, seemingly effective proxies for ground truth can be an important source of algorithmic bias in many contexts."

The model’s arithmetic was not the problem. Cost was the wrong thing to measure for this use. No model upgrade fixes that, and a review that checks only accuracy will pass it. The question that catches it is a data governance question: does this data measure what this use needs?

The EU now asks that question in law. Article 10 of the AI Act requires providers of high-risk systems to apply data governance practices to their training, validation and testing data, covering, among other things, "the formulation of assumptions, in particular with respect to the information that the data are supposed to measure and represent." The data must also be "relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose." A company that buys rather than builds is not off the hook: under Article 26(4), a deployer that controls the input data "shall ensure that input data is relevant and sufficiently representative." For high-risk uses listed in Annex III, such as hiring and credit scoring, those duties apply from 2 December 2027, after the Digital Omnibus (Regulation (EU) 2026/1744) moved the date from 2 August 2026.

Five Layers, Built in Order

I use a five-layer model, the Data Readiness Pyramid, to find where the data under an AI use breaks down. Each layer depends on the one below it.

1. Inventory. You know every source the use draws on, and each one has a named owner.
2. Quality and fitness. The data is accurate, complete and current, and it measures what the use needs.
3. Access. The people and systems that need the data can reach it quickly, and nobody else can.
4. Context. Metadata, lineage and agreed business meanings let you trace any output back to its sources.
5. Governance. Owners, classification, retention rules and change control keep the first four layers true over time.

A use that needs layer four cannot run on data that has not cleared layer one, yet that is what a feature-first budget asks of it. The pyramid works one use at a time. The whole estate can wait. Gartner’s own advice points the same way: AI-ready data must be "representative of the use case," and it is "not ‘one and done.’"

Five Tests Before You Fund the Next Feature

The pyramid says what to build. Each test says where one use stands today. Treat the pass marks as my own thresholds rather than an industry standard, and tighten them for high-risk uses.

If a use fails inventory or fitness, the two the others rest on, or any two tests, fund the data work first and the feature second, in the same decision. A single other failure can be fixed alongside the build. Fitness is the test I would never waive, because it asks the question the health care algorithm’s builders got wrong.

Agents Write Back

Agents change the stakes. A forecasting model reads data and produces a number a person can check before acting on it. An agent reads data, acts, and writes the result back into the systems that the next decision reads. A bad record becomes a bad action, and the output of that action becomes the next input.

For agents, tests four and five stop being good practice and become the control itself, and a trace that takes a day is too slow. You need to know what the agent read, what it wrote and what it was allowed to touch, at the moment it acts. Reading the logs afterwards tells you what went wrong. It does not stop the second write. Who answers for those writes is covered in You Cannot Outsource Accountability to an Agent.

Where This Argument Is Weakest

The strongest evidence against me comes from MIT’s Project NANDA. Its July 2025 report, built on a review of over 300 public AI initiatives, interviews with 52 organizations and a survey of 153 senior leaders, found that "95% of organizations are getting zero return" from generative AI. Its diagnosis was not data: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning," meaning tools that do not retain feedback or fit the work. I take that seriously. Clean data under a tool nobody uses is still a failed project. My claim is narrower. For a use that runs on your own data, readiness is necessary, not sufficient, and it is the one condition you can check before you spend.

The second objection is that this becomes a reason never to ship. That risk is why the tests apply to one use at a time and are built to be answered in about a day; a question nobody can answer in that time counts as a fail. And the tests do not ask for data that is completely ready, the bar only 7 percent cleared. They ask whether the data is fit for one use.

The surveys are also weaker than they look. That 7 percent comes from 231 readers of one publication in a sponsored study, and Gartner’s 60 percent is a prediction whose window closes at the end of this year. I read them as signals that agree with each other, not as a measure of your organization. The five tests measure that.

What to Do Before the Next Funding Decision

Take the next AI feature that reaches you for approval. Ask its owner to run the five tests on its data and bring the results to the meeting. Fund any gaps in the same decision as the feature, with the data work scheduled first. If nobody can answer the fitness question, what the data actually stands for, the feature is not ready, however good the demo looked. These tests sit inside a baseline such as Minimum Viable Governance, and Pillar 2 of the five-pillar readiness assessment takes the organization-wide view.

Sources and Changes

• Harvard Business Review Analytic Services, "Taming the Complexity of AI Data Readiness", sponsored by Cloudera, March 2026; online survey of 231 members of the HBR audience, October 2025.
• Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk", Q&A with Roxane Edjlali, 26 February 2025.
• Gartner, "Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025", 29 July 2024.
• Tadhg Nagle, Thomas C. Redman and David Sammon, "Only 3% of Companies’ Data Meets Basic Quality Standards", Harvard Business Review, 11 September 2017.
• Ziad Obermeyer and others, "Dissecting racial bias in an algorithm used to manage the health of populations", Science 366 (6464), 447 to 453, 25 October 2019.
• Regulation (EU) 2024/1689 (the AI Act), Article 10, Article 26(4) and Annex III, as amended by Regulation (EU) 2026/1744 (in force 27 July 2026), which sets 2 December 2027 for Annex III high-risk systems in Article 113.
• Aditya Challapally and others, "The GenAI Divide: State of AI in Business 2025", MIT Project NANDA, July 2025 (preliminary findings).

All sources checked on 6 October 2026.

Changes on 6 October 2026. Cut from about 5,300 words under a new headline. The 7 percent and Gartner’s 60 percent now say what their sources measured. Secondary statistics, an ROI chart and a vendor case study were removed; the 2019 health care algorithm study and the AI Act’s data rules were added, and the second test now asks whether the data measures what the use needs. A download box pointing to a missing file was removed.