There is a line you hear in almost every first meeting about AI: “Before we think about that, we need to sort out our data.” It sounds prudent. In most cases, it is the reason nothing happens for two years.
What Gartner says, and what it doesn’t
In February 2025, Gartner published two figures that have been doing the rounds ever since. In a survey of data management leaders carried out in July 2024, 63% of organisations said they lacked confidence in their data practices for AI. And the firm predicted that 60% of AI projects without AI-ready data would be abandoned by the end of 2026 (Gartner, 2025).
At first glance, this seems to vindicate those who want to tidy everything up first. But Gartner’s own notion of “AI-ready data” is defined by use case: data is ready when it serves the specific problem you want to solve.
AI-ready data is not perfect data. It is data that is fit for a specific question.
The trouble with never-ending projects
“Sort the data out first” projects have a fundamental flaw: they have no stopping point. Without a specific use, there is no way to decide what is good enough. Everything gets normalised, everything gets documented, and by the time you finally reach the AI, the company’s priorities have moved on.
Data is better sorted the other way round, starting from a question. When you know you want to predict stock-outs by store, you know exactly which fields have to be right and which can wait.
What the best forecasting competition teaches
The M5 competition, organised by researchers at the University of Nicosia using real Walmart data, is a good example. The winning methods used the sales history as it existed, with tens of thousands of product and store series (42,840 in total), and added information that was already available: prices, promotions and the events calendar (Makridakis, Spiliotis and Assimakopoulos, 2022). The gain came from making good use of the data that already existed.
What you really need to have
- Access. Someone with the authority to grant read access to the relevant systems. This is usually the slowest step, and there is nothing technical about it.
- Enough history. To forecast, examples from the past. To automate a task, examples of how it is done today.
- A common identifier. A customer, item or trip code that links the ERP to email and spreadsheets.
What gets sorted out along the way
Duplicates, names spelt in different ways, information buried in PDFs and emails, parallel spreadsheets: all of this can be handled while the project is under way. Today’s models cope well with these imperfections, and the project itself shows which of them actually matter. The data quality plan that comes out of it has priorities set by real use.
Sources
- Gartner (2025). Lack of AI-Ready Data Puts AI Projects at Risk (press release, 26 February 2025, summarised by Freevacy). Freevacy.
- Makridakis, S., Spiliotis, E., Assimakopoulos, V. (2022). M5 accuracy competition: Results, findings, and conclusions. International Journal of Forecasting, 38(4), 1346-1364.