An AI use case inherits the data it consumes. Their conditions of use, their defects, their origin and any absence of an owner follow it right into its results, whether the model is trained on this data or merely reads it to answer.
Data governance for AI applies to the datasets and documents that an AI system consumes the roles and rules through which an organisation knows who is accountable for a piece of data, who may use it and on what conditions. It is checked before a use case is authorised, on five subjects that the committee has documented: the right to use the data, its quality in relation to the use, its traceability, personal data and responsibilities. This grid is a proposed method, which none of the texts cited requires in this form. The roles themselves are described on our data governance page.
Do we have the right to use this data for this purpose?
A piece of data was collected, bought or received for a use. The AI use case proposes another, and the answer depends on the origin of the data.
For data bought or received from a partner, the answer is read in the contract: permitted scope of use, right to derive a model from it, duration. For data entered into a market AI service, it is read in the provider’s terms, in particular on the reuse of inputs.
For personal data, the GDPR, Regulation (EU) 2016/679, lays down in its Article 5(1)(b) the principle of purpose limitation: data is collected for “specified, explicit and legitimate purposes” and not further processed in a manner incompatible with those purposes. Article 6(4) lists what the controller takes into account in assessing this compatibility, including the link between the purposes, the context of collection and the possible consequences for individuals.
For each dataset, the committee expects its source, the contract governing its use and, for personal data, the legal basis for the processing and the initial purpose. Our article on data provenance and lineage describes the record that brings them together.
Is this data sufficient for what the system must do with it?
Quality is assessed in relation to the use, and use by a model adds a criterion that completeness or uniqueness checks do not measure: representativeness. A complete file with no duplicates may contain almost no cases from a population that the system will nevertheless have to handle.
For high-risk systems, Regulation (EU) 2024/1689, known as the AI Act, makes this a requirement. Its Article 10(3) requires training, validation and testing datasets that are “relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose”. This requirement falls to the provider. The deployer has its own: under Article 26(4), to the extent that it exercises control over the input data, it ensures that this data is relevant and sufficiently representative in view of the intended purpose of the system. The application dates are given in our article on the AI Act’s obligations for businesses.
For an assistant that reads a document base, the question concerns the documents: outdated versions, duplicates, contradictory texts, which the system may reproduce without flagging them. Our entry on RAG sets out the mechanism.
The document expected is a measured status of the data, compared with an expectation written by its owner. The corresponding rules and checks fall within our data quality offer.
Will we be able to say, in a year’s time, what went into the system?
Traceability is prepared before go-live, because it is hard to reconstruct after the fact.
It begins with the list of datasets used to train and evaluate the model, each with its version. Without this list, a disputed result cannot be reproduced, and a new training run cannot be compared with the previous one.
It continues with the path of these datasets: their origin, then the processing that prepared them. This path is also used to respond to a request for rectification or erasure, by indicating which datasets contain the data concerned.
What remains is what the system receives and produces in operation. For high-risk systems, Article 12(1) of Regulation (EU) 2024/1689 requires that they technically allow for the automatic recording of events over their lifetime.
The file states where these traces are kept, by whom, and for how long.
Can individuals be identified in this data?
If the answer is yes, the GDPR applies to the use case, whether or not the AI regulation classifies it as high-risk. The committee then hears the data protection officer.
Article 5(1)(c) requires data that is “adequate, relevant and limited to what is necessary”: each column entrusted to the model is justified by the purpose. Article 35(1) requires an impact assessment before processing when it, “in particular using new technologies”, is likely to result in a high risk to the rights and freedoms of individuals.
Article 22(1) gives individuals the right not to be subject to a decision based solely on automated processing which produces legal effects concerning them or similarly significantly affects them. The file therefore describes the part the system plays in the decision and the part that falls to a person.
A dataset presented as anonymised is checked before being treated as such, as our article on data anonymisation explains. The qualification belongs to the controller.
Who will be informed on the day the data changes?
When both forms of governance are in place, an AI system has an owner, recorded in the register described on our AI governance page, and each dataset has an owner. The committee’s question concerns the link between the two.
A source can change without its consumers being warned: a field takes on another meaning, an application is replaced, a supply contract comes to an end, a retention period expires. The data owner must know that an AI system depends on it, and the system owner must know who then decides to retrain, restrict or stop.
The file names the owner of each dataset, carries their written agreement for this use and describes the channel through which a change is reported to them.
Data governance for AI is read in the file submitted to the committee
| Subject | Who is accountable | Document attached to the file |
|---|---|---|
| Right to use the data | The data owner, with the legal function | Source, contract or legal basis, initial purpose |
| Quality in relation to the use | The data owner and the project team | Measured status, compared with a written expectation |
| Traceability | The team that operates the platform | Versioned datasets, path, planned logs |
| Personal data | The controller, after the opinion of the data protection officer | Opinion, and impact assessment where required |
| Responsibilities | The system owner | Named owners, written agreements, reporting channel |
In this method, a missing document does not carry the same weight depending on the row. On the right to use the data and on personal data, it blocks authorisation, since the answer depends on the law or a contract. On the other three rows, and outside any requirement specific to a high-risk system, the committee may authorise a reduced scope, writing down the condition set and the date on which it will be checked.
The authorisation applies to the data described in the file: an added dataset, a replaced source or a broadened purpose bring the use case back before the committee.