Does the data have to be in order before an AI investment?
Data difficulties were cited in 2025 by 43.5 per cent of EU enterprises that considered AI without using it. Data quality for AI is still not settled for the data as a whole but for the fields a decision reads, and part of the check can be done by the AI step itself.

Key insights
- Among EU enterprises with at least ten employees that considered AI but do not use it, 43.5 per cent cited data availability or quality difficulties in 2025, according to Eurostat. Lack of expertise: 70.3 per cent.
- Among Swedish enterprises that considered AI without using it, 44.3 per cent cited data difficulties in 2025, close to the EU level. Counted across all enterprises with at least ten employees, it is 3.2 per cent.
- Gartner defines AI-ready data as data that is representative of a specific use case. That means whether the data is in order cannot be answered until the use case is named.
- In experiments, language models find errors within a row but miss errors that only show across many rows. On a hospital table, repair of unfamiliar names fell far more than detection for three of four commercial models.
- The AI step can itself be the data quality check for fields where an error shows against the rest of the record. Fields where errors only show across many records, and keys other systems read, need checking first.
Data quality for AI comes up early in almost every AI investment, and it tends to get an answer that pushes everything else back: first the data has to be in order. That answer assumes data quality is a property of the data as a whole. It is not. It is a property of the fields a particular decision reads, and part of checking them can be done by the AI step itself, while another part should not be left to it.
Data difficulties are not the most common reason to hold back from AI
Eurostat's survey of ICT usage in enterprises asks the companies that considered AI but do not use it why they held back. In the EU compilation for 2025, 43.5 per cent of those enterprises cited difficulties with the availability or quality of the necessary data. The year before, the share was 44.4 per cent. The share is therefore roughly the same in both years, but it is not the most common reason. A lack of relevant expertise was cited by 70.3 per cent, and two further reasons ranked above data.
All eight answer options. Share of enterprises with at least ten employees, excluding the financial sector, that considered AI but do not use it. Several answers allowed, so the bars do not sum.
Source: Eurostat, The use of artificial intelligence (AI) technologies in the European Union, 2026 edition
For Sweden, Eurostat's table shows the same question. Among Swedish enterprises with at least ten employees that considered AI without using it, in the sectors the statistics cover and excluding the financial sector, 44.3 per cent cited data difficulties as a reason in 2025, against 37.8 per cent the year before. Sweden is therefore close to the EU level. Counted across all enterprises it is 3.2 per cent, because only 7.2 per cent of Swedish enterprises had considered AI without using it. The survey's list of AI technologies was extended in 2025, so the comparison between the years is not fully like for like.
Both figures are the companies' own answers. The question does not distinguish which data or which use case the answer concerns, and that is precisely the distinction that decides whether the answer can be acted on.
Gartner's definition of AI-ready data makes data quality a question per use case
Gartner wrote in February 2025 that 63 per cent of organisations either do not have, or are unsure whether they have, the right data management practices for AI, and predicted that through 2026 organisations will abandon 60 per cent of AI projects unsupported by AI-ready data. The forecast runs through 2026 and cannot yet be checked. The basis is also unclearly reported: the body of the press release cites a survey of 1,203 data management leaders in July 2024, while the page's meta description cites 248 respondents in the third quarter of the same year.
The useful part of the text is the definition. Data is AI-ready when it is representative of the use case, with every pattern, error, outlier and unexpected emergence needed to train or run the model for that specific use. Gartner also writes that AI-ready data is not "one and done", but a practice that follows the use cases that exist and those still to come.
The consequence is that the question of whether the data is in order has no answer until the use case is named. The same customer register can be good enough to summarise case history and not good enough to decide which customers to prioritise, because the two uses read different fields.
Data is never in order in general, only in order for a particular decision.
That also moves the question of who decides. Which fields a decision reads, and how wrong they may be, is a question about the decision and not about the system. The person who owns the decision is the one who can answer it.
The language model finds the errors in the row, not in the pattern
In a study from March 2025, presented at a workshop at the ICLR conference, researchers at ETH Zurich and CSEM had four language models with access to Python clean three open datasets into which the researchers had planted errors. The models found and corrected errors visible in a single value or a single row, by testing the value against other fields in the same row, but the researchers also show cases where a detected error was corrected arbitrarily. They missed errors that only show across many rows: shifts in distribution, trends and biases. No model detected that ten had been added to a field for booking lead time, but only for the year 2016.
A preprint from November 2026, by researchers at Minzu University of China and Renmin University of China, tests a question that decides what such experiments say about an organisation's own data: how much of the models' accuracy on open benchmark tables comes from recognising the hospitals and films in them? The researchers replaced the names with invented ones and kept the errors in the same places. On the hospital table, DeepSeek-V4-Pro's ability to repair errors in name and location fields fell from an F1 score of 0.88 to 0.16, where 1 is the best possible score. Its ability to find the errors held up better, from 1.00 to 0.73, and its repair of the other fields stayed between 0.91 and 0.95.
One model, DeepSeek-V4-Pro. The errors sit in the same places in all three variants. F1 combines precision and recall, and 1 is the best possible score. Preprint.
Source: Yang et al., Cleaning or Recalling?, preprint arXiv:2611.08106, November 2026, table 8
An organisation's own operational data resembles the invented names in that respect. Customer numbers, article codes and internal project names are generally not in what a model was trained on. What the model can do with them is see that something does not fit the rest of the row, and even that ability falls when the names are unfamiliar. It repairs them considerably worse. One of the ETH study's examples of a detected error, implausibly high fish consumption in landlocked countries, also relied on the model's knowledge of the countries and not only on the row.
The AI step can be the data quality check
That is where the argument lies, and it is a judgement: the experiments measured models asked to clean data, not models that flag anomalies while a use case runs. When a use case runs, the model reads every record it touches, which no manual review does. If it is allowed to flag what does not fit the rest of the record, and the error is then corrected in the source system by whoever owns the task, the run itself becomes the data quality work that would otherwise have preceded it. The work also lands on the fields the use case actually reads, rather than on everything that might one day be used.
The same movement appears in maintenance logs in manufacturing, where the free text can test the code that was set. It applies to operational data in general: case management systems, customer registers, purchase lines, project accounting. The condition is that the flag goes to a person and not straight back into the register.
What argues against letting the AI step do the work
Gartner's press release puts it plainly: if the data has issues, then the data is not ready for AI. The experiments give that objection substance. In the Minzu and Renmin study, the researchers gave well-known films a different real director in the table and tested whether the models kept it. Despite explicit instructions to rely on the table, commercial models changed it back to the film's real director in between 0 and 40 per cent of cases, depending on the model. Without that instruction they did so in 68 to 100 per cent of cases. A model allowed to repair freely can therefore replace a value that deliberately differs in an organisation's own data with what the model itself believes is right.
A field that changed meaning in a system migration, a unit that changed in a particular year, or a code that started being used in a new way are, moreover, errors of exactly the kind the models missed in the ETH study. They only show across many rows, and they can pass untouched through the whole run.
The evidence also has limits. The ETH study is a workshop paper and the Minzu study a preprint. Both were run on open datasets with errors the researchers constructed or transferred themselves, and neither measures Swedish or Nordic operational data. In the ETH study the models also received feedback scored against a clean test set, and in some runs hints about where the errors were, which is absent when a use case runs for real. Eurostat's figures are self-reported.
An error that only shows across many records is a reason to clean first, not to wait with everything.
The objection therefore concerns certain fields and not the data as a whole. It is a reason to sort the fields before the use case starts, not a reason to postpone it.
The criterion: three questions for every field the decision reads
Take the use case the leadership team most wants in operation and list the fields the decision actually reads. Then put three questions to each field, in this order.
Would an error in the field change the decision? If not, the field needs no action before the start, however incomplete it is.
Does an error in the field show against other fields in the same record? A delivery date before the order date, a line amount that does not equal quantity times unit price, a free-text note describing a complaint on a line coded as a new sale. Then the AI step can be the check: the model flags, and whoever owns the task corrects it at source.
Does an error only show across many records, or is the field a key that other systems look up against? Then it should be checked before the start by someone who knows the field's history, and the model should not be allowed to write to it.
The answers differ between organisations, and that is the point. One whose decision rests on free text and fields that can be tested against each other can start now and let use carry the cleaning. One whose decision rests on a code that changed meaning at the last system migration starts with that code, and only with that code. In both cases it is the owner of the decision who settles which fields must be right, because the owner knows what an error costs. An AI readiness assessment goes through which data the prioritised use cases need in its requirements and design step.
Common questions
Not for the data as a whole. Data quality for AI is settled per use case, because each decision reads only certain fields. Fields where an error shows against other fields in the same record can be checked by the AI step itself, by flagging anomalies that are then corrected in the source system. Fields where errors only show across many records, and keys that other systems look up against, need checking before the use case starts.
Gartner defines AI-ready data as data that is representative of the use case, including every pattern, error, outlier and unexpected emergence needed to train or run the model for that specific use. The definition makes readiness a question per use case rather than a property of all the data. Gartner also describes the work as an ongoing practice rather than something done once.
According to Eurostat, 43.5 per cent of EU enterprises with at least ten employees that considered AI but do not use it cited difficulties with the availability or quality of data in 2025, and 44.4 per cent the year before. It is one of the more common reasons but not the most common: a lack of relevant expertise was cited by 70.3 per cent. The figures are the companies' own answers.
Among Swedish enterprises with at least ten employees that considered AI without using it, 44.3 per cent cited difficulties with the availability or quality of data as a reason in 2025, against 37.8 per cent in 2024, according to Eurostat statistics that exclude the financial sector. Counted across all enterprises it is 3.2 per cent, because only 7.2 per cent of Swedish enterprises had considered AI without using it.
Partly. In a study from ETH Zurich and CSEM, presented at a workshop in 2025, language models found and corrected errors visible in a single value or row by testing the value against other fields in the same row, but the researchers also show cases where a detected error was corrected arbitrarily. They missed errors that only show across many rows, such as a shift that applied to one year only. The model therefore works best as a check that flags, with the correction made in the source system.
Errors that only show across many records: shifts in distribution, trends and biases, according to a 2025 study from ETH Zurich and CSEM. None of the models tested noticed that a field had been increased by ten for one year only. A 2026 preprint from Minzu and Renmin universities also shows that models repair unfamiliar names far worse than familiar ones, and that their ability to detect the errors falls too.
On the fields the decision in the use case actually reads, not on the data as a whole. For each field, ask whether an error would change the decision, whether an error would show against other fields in the same record, and whether it would only show across many records. Only the last group, together with keys that other systems depend on, needs cleaning before the start.
The owner of the decision the use case supports. Which fields a decision reads, and how wrong they may be, depends on what an error costs in that decision, and the decision owner knows that better than whoever runs the system. The practical test is three questions per field: does an error change the decision, does it show within the record, and does it only show across many records, or is it a key other systems look up against.
If this lands on your desk, we should talk.
Ampliro Insights
New analysis, roughly weekly.
We write when the rules change and when something turns out to work in practice. One piece at a time, no sequences, and you can leave from any issue.
We store your address to send Ampliro Insights, and for nothing else. More in the privacy policy.