Almost every study of failed AI projects lands on the same cause, and it is not the model. It is data that was never ready to carry a decision. This is a check you can score honestly before you approve the next build.
The short answer
- Data readiness is five questions, not one: who owns it, is it accurate, can the model reach it, is it lawful to use, and is it fit for this specific purpose.
- Score each of the twenty points as green, amber or red for the one use case in front of you, not for the estate in general.
- Any red in the ownership or lawfulness sections stops the project until it is fixed. The others are cost and time, not blockers.
- A high score on a curated pilot extract means little. The test is the live production feed the model will actually run on.
- The exercise takes an afternoon and prevents the most expensive failure, which is discovering the data was not ready after the money is spent.
Why a readiness check pays for itself
The failure studies converge on data. Poor data quality, inadequate data, and problems nobody owned appear across the Gartner and RAND findings on why AI projects fall short, while model selection appears in none of them. The pattern behind those numbers is examined in what the failure studies actually measured. The practical response is to test the data before the build, when a red flag costs a conversation, rather than after, when it costs the programme.
Score every point below for the specific use case you are considering. Data that is ready for a monthly management report is often not ready for a real-time automated decision, so readiness is always relative to the job.
1. Ownership and accountability
- A named person owns each dataset the use case depends on, and their name is written down.
- That owner can commit to a service level: how fresh the data is, and what happens when a feed breaks.
- Accountability for the data sits with the business function that uses it, rather than defaulting to IT by absence.
- Someone is answerable for the definitions, so "active customer" or "completed order" means one thing across the sources involved.
2. Quality and lineage
- You can state the current accuracy and completeness of the key fields, with a number rather than an impression.
- You can trace where each important number came from, through the systems it passed through.
- Duplicate and conflicting records are understood and quantified, not merely suspected.
- The rate at which the data goes stale is known, so the model's assumptions do not silently expire.
3. Access and integration
- The model can reach the data in production through a supported route, rather than a one-off extract someone built by hand.
- The data from different systems can be joined on a reliable key, so records about the same customer or asset line up.
- The live feed carries the same shape and meaning as the sample the pilot was built on.
- Latency is fit for the decision: real-time where the decision is real-time, batch where it is not.
4. Governance and lawfulness
- You have a lawful basis for using this data for this purpose, which is a separate question from holding it, as set out in the rules on training on your own data.
- Any special category data is identified, and the extra conditions for it are met.
- Retention and minimisation are handled, so the model is not trained on data you should have deleted.
- A data protection impact assessment exists where the processing is high-risk, which for AI on personal data it usually is.
5. Fitness for this use case
- The data actually contains the signal the model needs, rather than a proxy that correlates until it does not.
- The historical data is representative of the population the model will run on, so it is not learning a pattern that has since changed.
- Known bias in the data is documented, so a fairness problem is a decision rather than a surprise.
- There is enough of it, at the granularity the use case requires, to support the claim being made.
Reading the score
The twenty points are not equal. The ownership and lawfulness sections are gates: a red there stops the project, because no amount of engineering fixes an absent owner or a missing lawful basis, and both are management decisions rather than technical ones. The quality, access and fitness sections are effort: a red there is a cost and a timeline, and a plan can be built to clear it. A use case that is green on the gates and amber on the effort is a fundable project with known work in front of it. One that is red on a gate is not ready for capital, whatever the model can do.
The constraint
For most organisations the binding limit exposed by this check is ownership, in the first section, rather than the technical quality measured in the others. Funding data-cleaning tools and integration work while no named owner will commit to a service level spends money on the symptom, because the data will decay again the moment the project ends, and the decision that closes the gap is the appointment of an accountable owner.
How to use the result
Run the check as a scored conversation with the business owner and the data lead in the room together, because the disagreements between them are where the real state of the data shows. Record the reds, separate the gates from the effort, and put the total in front of whoever is asked to fund the build. A project that starts with an honest readiness score, and a plan to clear the ambers, is the kind that survives contact with production. A project that skips the check tends to discover its data was not ready at the least convenient moment, which is the failure the whole exercise exists to prevent.
Sources
- RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024, on data adequacy and problem definition as recurring causes.
- Gartner press releases on generative and agentic AI project abandonment, 2024 and 2025, naming poor data quality among the drivers.
- Information Commissioner's Office, guidance on AI and data protection, on lawful basis and impact assessments.