Guide
Data governance for AI: what changes, what the rules require and how to build it
AI data governance is the set of rules and controls for the data that trains, tests and feeds AI systems and the data they produce: where it came from, whether you may use it, whether it is good enough, and how long you keep it. It extends ordinary data governance to training data provenance, lawful basis, bias, unstructured data, and prompts and outputs as records, and the EU AI Act’s Article 10 now writes much of it into law for high-risk AI [4]. This guide covers what changes for AI, what the EU AI Act, GDPR, ISO/IEC 42001 and the NIST AI RMF require (compared in one table), a seven-step program, and the vendors whose own pages document this work.
Reviewed 5 Nov 2026 · next review by 4 Jan 2027 · 31 sources · report an error
What is AI data governance?
Data governance decides who owns data, what it may be used for, how good it must be and how long it is kept. AI data governance applies those decisions across the AI data life cycle, which ISO/IEC 8183 describes as running from obtaining data through to retiring it, across the creation, deployment and maintenance of AI systems [2]. Its quality side has its own ISO series: ISO/IEC 5259-1, published in July 2024, sets terms and a framework for assessing and improving data quality for analytics and machine learning [3].
The phrase “data governance and AI” is used two ways. This guide is about governing data for AI. The other sense, using AI for data governance, means tools that classify data automatically; several vendors below do both [21][23].
- 63%
- of organizations lack, or are unsure they have, the right data management practices for AI[1]Gartner survey of 1,203 data management leaders, July 2024
- 60%
- of AI projects without AI-ready data will be abandoned, Gartner predicts[1]Gartner prediction through 2026
- 34%
- of low-maturity AI organizations name data availability and quality a top challenge (29% of high-maturity)[29]Gartner survey of 432 organizations, Q4 2024
- 10%
- of organizations put data governance teams in charge of AI governance[28]IAPP survey of 670+ respondents, spring 2024
How does data governance change for AI?
Most data governance programs were built for structured data in warehouses and reports. AI adds new questions at each stage, and some answers can’t be changed after training.
| Area | Traditional data governance | Data governance for AI |
|---|---|---|
| Provenance | Which system a table came from | Where every training, validation and test set came from, and the original purpose of any personal data in it [4] |
| Lineage | Source to report | Source dataset through training, inference and deployment to output [19] |
| Lawful basis and consent | Checked when data is collected | Needs its own legal basis for training; unlawful training can taint later deployment [9][12] |
| Quality | Accuracy and completeness | Also relevance, representativeness and statistical properties for the people the system affects [4] |
| Bias | Rarely in scope | Datasets examined for biases that could harm health, safety or rights, with measures to detect and mitigate them [4] |
| Retention and deletion | Delete the record | Deleting source data may not remove it from a trained model; erasure or retraining can be ordered [9] |
| Prompts and outputs | Not data assets | Logs that must be kept for high-risk AI and can be held for litigation [7][13] |
| Unstructured data | Often out of scope | Documents, files and vector stores that feed retrieval and agents [20][21] |
Examples. In December 2024 Italy’s data protection authority fined OpenAI €15 million, finding that it had used personal data to train ChatGPT without an adequate legal basis; OpenAI said it would appeal [12]. In May 2025, in the New York Times copyright case, OpenAI was ordered to retain all of its ChatGPT chat logs; a federal judge ended that duty as of September 26, 2025, but logs already saved stayed accessible and data tied to accounts the Times flagged still had to be kept [13]. Both cases turn on data decisions made long before anyone asked about them.
What does EU AI Act Article 10 require?
Article 10 applies to high-risk AI systems. Systems that train models must be developed on training, validation and testing datasets that meet its quality criteria; systems that don’t train models must meet them for testing datasets [4]. The obligations apply from December 2, 2027 for Annex III systems and August 2, 2028 for AI in regulated products, after the Digital Omnibus on AI deferred them [6].
- Governance practices (Art. 10(2)). Datasets need data governance and management practices covering design choices; data collection and origin, including the original purpose of personal data; preparation such as labelling, cleaning and enrichment; the assumptions behind the data; availability, quantity and suitability; examination of possible biases; measures to detect, prevent and mitigate them; and data gaps and how to fix them [4].
- Quality (Art. 10(3)). Datasets must be relevant and sufficiently representative and, as far as possible, free of errors and complete, with appropriate statistical properties for the people the system will be used on [4].
- Context (Art. 10(4)). Datasets must reflect the geographical, contextual, behavioral or functional setting of use, as far as the intended purpose requires [4].
- Sensitive data for bias (new Art. 4a). The Omnibus deleted Article 10(5) and inserted Article 4a. Providers of high-risk systems may exceptionally process special categories of personal data where strictly necessary for bias detection and correction; providers and deployers of other AI systems and models, and deployers of high-risk systems, may do so under the same conditions where bias could affect health or safety, harm fundamental rights or lead to prohibited discrimination [4][5].
Article 4a sets conditions: other data, including synthetic or anonymized data, must not be enough; re-use must be technically limited, with measures such as pseudonymization; access must be strictly controlled and documented; the data must not be passed to other parties; it must be deleted once the bias is corrected or the retention period ends; and records must explain why the processing was strictly necessary. GDPR still applies alongside it, and Article 4a creates no duty to run bias detection [5].
Deployers have data duties too. Where a deployer controls the input data of a high-risk system, it must make sure that data is relevant and sufficiently representative, and it must keep the logs the system generates for at least six months, unless other law, such as data protection law, says otherwise [7]. Deployers that must carry out a fundamental rights impact assessment may cross-reference a GDPR data protection impact assessment that already covers the same ground [8].
What does GDPR require for AI training data?
The main guidance is the European Data Protection Board’s Opinion 28/2024 of December 17, 2024 [9][31]. Its four points shape most AI data programs in Europe:
- Models are not automatically anonymous. Whether a model trained on personal data is anonymous is decided case by case. It must be very unlikely to identify the people whose data trained it, or to give up their data through queries [31]; CMS summarizes the bar as a negligible likelihood of extraction [9].
- Legitimate interest is possible, not automatic. It cannot be the default legal basis for training or using AI models; the controller has to pass a three-step test, and mitigations such as pseudonymization, synthetic data, transparency and an opt-out offered from the start count in its favor [9].
- Unlawful development can follow the model. If a model was developed with unlawfully processed personal data, that can affect the lawfulness of its deployment unless the model has been duly anonymized [31], and authorities may order fines, limits on processing, erasure of all or part of the dataset, or retraining [9].
- Document the assessment. In our reading, each of these points requires written evidence: an anonymity assessment, a legitimate interest assessment, and records of the mitigations applied.
The rules may change. The Commission’s pending Digital Omnibus proposal includes GDPR changes, among them codifying legitimate interest as a basis for AI training [11]. On October 22, 2026 the IAPP reported that EU member states had agreed their negotiating position, keeping the existing definition of personal data and adding safeguards for sensitive data used in AI training; trilogue talks can begin only once the European Parliament agrees its own position, which the report did not say had happened [10]. Until a text is adopted, the EDPB opinion and current GDPR apply.
What do ISO/IEC 42001, the NIST AI RMF and US laws expect?
ISO/IEC 42001 has a group of controls on data for AI systems. In the numbering of the NIST-hosted crosswalk they are B.7.2 data for development and enhancement, B.7.3 acquisition of data, B.7.4 quality of data, B.7.5 data provenance and B.7.6 data preparation [14]. In outline, they ask you to define data management processes, document what data you acquire and from where, set quality criteria, record provenance from collection through transformation and storage, and document preparation methods such as labeling [15]. The crosswalk maps them to NIST MAP 2.3, MEASURE 2.3, 2.7, 2.8, 2.9 and 2.10, and MANAGE 2.2 [14].
The NIST AI RMF is voluntary, but its subcategories make a usable checklist for data: document data collection and scientific integrity (MAP 2.3), map the legal risks of components including third-party data (MAP 4.1), examine privacy risk (MEASURE 2.10), evaluate fairness and bias (MEASURE 2.11), set policies for third-party risks including intellectual property (GOVERN 6.1), and plan safe decommissioning (GOVERN 1.7) [16].
US state laws. California’s AB 2013, in force since January 1, 2026, requires developers of generative AI made available in California to post a high-level summary of their training data, covering sources and owners, volume, collection methods, intellectual property status, whether it includes personal information, collection dates and whether synthetic data was used [17]. Colorado’s SB 26-189 will require developers of automated decision-making technology to give deployers technical documentation that includes the categories of training data, from January 1, 2027 [18]. See our Colorado and CCPA/CPRA hubs.
AI data governance requirements in one table
Read across a row to see what each source asks for on the same topic. Build the control once and file the evidence against every column that applies to you.
| Topic | EU AI Act | GDPR (EDPB) | ISO/IEC 42001 | NIST AI RMF | US state law |
|---|---|---|---|---|---|
| Provenance and origin | Collection processes, origin and original purpose of personal data (Art. 10(2)(b)) [4] | Know whether personal data is in the training set and the model [9] | Acquisition (B.7.3), provenance (B.7.5) [14] | MAP 2.3, MAP 4.1 [16] | AB 2013: sources, owners, collection dates [17] |
| Lawful basis and purpose | GDPR applies alongside (Art. 4a) [5] | Legitimate interest only after a three-step test [9] | Privacy considerations in acquisition (B.7.3) [15] | MEASURE 2.10 privacy [16] | AB 2013: whether personal information is included [17] |
| Quality and representativeness | Relevant, representative, error-free as far as possible (Art. 10(3)); deployer input data (Art. 26(4)) [4][7] | Accuracy principle (our reading) | Quality of data (B.7.4) [14] | MAP 2.3; MEASURE 2.3 via crosswalk [14][16] | Not addressed |
| Bias and sensitive data | Examine and mitigate bias (Art. 10(2)(f)–(g)); sensitive data only under Art. 4a [4][5] | GDPR still applies alongside Art. 4a’s conditions [5] | Impact on individuals and groups (B.5.4) [14] | MEASURE 2.11 [16] | Colorado: developer documentation of training data categories [18] |
| Preparation and labeling | Annotation, labeling, cleaning, enrichment (Art. 10(2)(c)) [4] | Pseudonymization and synthetic data count as mitigations [9] | Data preparation (B.7.6) [14] | MAP 2.3 [16] | AB 2013: processing methods [17] |
| Logs, retention and deletion | Deployers keep logs at least six months (Art. 26(6)); Art. 4a data deleted after use [7][5] | Erasure or retraining possible after unlawful training [9] | Data management up to disposal (B.7.2) [15] | GOVERN 1.7 decommissioning [16] | Colorado: developers and deployers keep compliance records at least three years [18] |
| Documentation and disclosure | Records explaining any Art. 4a processing [5] | Evidence of the anonymity and legitimate interest tests [9] | Documented processes and provenance (B.7.2, B.7.5) [15] | MEASURE 2.1 test sets documented [16] | AB 2013 public summary; Colorado deployer documentation [17][18] |
How should you govern prompts, outputs and unstructured data?
Generative AI turns every conversation into data. Prompts can contain personal or confidential information, outputs can be relied on in decisions, and both may be logs you must keep or a court may order you to keep [7][13]. Treat them as a data asset with an owner, a classification and a retention period.
- Classify at the point of entry. Tools now classify sensitive data in prompts, files and vector stores as well as in training sets [20].
- Set retention by purpose. High-risk AI logs: at least six months for deployers in the EU [7]. Everything else: as short as the purpose allows, with a documented legal-hold process [13].
- Govern the retrieval corpus. Retrieval-augmented generation (RAG) systems answer from your documents, so access rights on the source files carry into answers. Inventory the vector stores and the folders that feed them [20][21].
- Keep security separate but linked. Leakage, poisoning and access control are covered in our AI data security guide; employee use of unapproved tools in our shadow AI guide.
How to build data governance for AI: seven steps
- 01
- 02
- 03
Decide the lawful basis before training
For personal data, document the legal basis; if relying on legitimate interest, complete the three-step test and apply mitigations such as pseudonymization and an opt-out [9].
- 04
- 05
- 06
- 07
Monitor and review
Re-check data quality and drift in production, review the map when a model is retrained or a vendor changes, and report coverage as a governance metric. Our AI governance maturity model shows what measured looks like [16].
Which vendors do data governance for AI?
We checked six directory vendors’ own product pages for data governance for AI specifically: dataset lineage, provenance, quality, classification or policy enforcement on data used by AI. Five describe it; descriptions are the vendors’ own claims and we have not tested them. Dataiku Govern’s page describes project and model registries, sign-offs and audit timelines, but not dataset-level lineage or quality [27]. Vendors did not pay to be listed, and the order is alphabetical.
- 01
Best for: Finding sensitive data across training sets, vector stores and prompts
Discovers and classifies sensitive and regulated data across training sets, vector stores, files and prompts, traces lineage from source to output, runs compliance checks before training, and maps evidence to the EU AI Act, NIST AI RMF, ISO/IEC 42001, GDPR and CCPA. Also inventories models, agents and MCP servers.[20]
- 02
Best for: Linking AI use cases to governed datasets and lineage
Registers AI use cases, models and agents and connects each to the datasets it relies on, with lineage from source datasets through training, inference and deployment. Teams can check quality rules and privacy requirements before using data, and use EU AI Act and NIST AI RMF assessment templates.[19]
- 03
Best for: Model documentation in IBM and third-party estates
Model-first: factsheets record a model’s creation information, the data used and its lifecycle stage, and monitoring covers drift, quality and safety. Its model governance page does not describe dataset lineage, so check that in a demo.[25][26]
- 04
Best for: Enforcing consent and purpose rules on data used by AI
Its Data Use Governance product, announced in May 2025, turns privacy, consent and regulatory policies into controls such as column masking and row filtering, pushed down to platforms including Snowflake and Databricks. OneTrust says its own teams use its platform to document AI development processes, data sources, inputs and governance decisions.[23][24][30]
- 05
Best for: Data discovery, lineage and LLM firewalls in one platform
Now part of Veeam Software (Dec 2025)
Now part of Veeam, which completed the $1.725 billion acquisition in December 2025. Its platform covers data discovery and classification, lineage, catalog, quality and access governance, plus curating and sanitizing data for model training and context-aware firewalls for prompts, retrieval and responses.[21][22]
Compare the wider field in our AI governance platforms ranking and EU AI Act compliance tools.
Frequently asked questions
+What is AI data governance?
It is the set of rules and controls for data used to train, test and run AI and the data AI produces: provenance, lawful use, quality, bias, retention and documentation. For high-risk AI in the EU, Article 10 of the AI Act sets many of these requirements [4].
+What is the difference between data governance and AI governance?
Data governance controls data wherever it is used. AI governance controls AI systems: whether to use them, how, and who is accountable. Data governance for AI is where they overlap, and ISO/IEC 42001 has its own data controls for it [14]. See what AI governance is.
+What does Article 10 of the EU AI Act require?
High-risk AI systems must be developed on datasets subject to documented governance practices covering origin, preparation, assumptions, suitability, bias examination and mitigation, and data gaps, and the data must be relevant, sufficiently representative and as error-free and complete as possible [4]. It applies from December 2, 2027 for Annex III systems [6].
+Can you use personal data to train AI under GDPR?
Yes, if you have a lawful basis and meet the other GDPR principles. The EDPB says legitimate interest cannot be the default basis and needs a three-step test, and that unlawful training can lead to orders to erase data or retrain the model [9]. Proposed GDPR changes on AI training were still in negotiation in October 2026 [10].
+Can you process sensitive data to detect AI bias?
Under the EU AI Act’s new Article 4a, providers of high-risk systems, and providers and deployers of other AI where strictly necessary, may process special categories of personal data for bias detection and correction, but only if other data won’t do and with strict security, access, deletion and record-keeping conditions [5].
+What is data provenance in AI?
Data provenance is the record of where data came from and how it was collected, transformed and stored, so you can trace what a model was trained on. ISO/IEC 42001 has a provenance control [14], and the EU AI Act asks for the origin of data and the original purpose of personal data [4].
+Do prompts and AI outputs need data governance?
Yes. EU deployers of high-risk AI must keep system logs for at least six months [7], and in 2025 OpenAI was ordered to retain ChatGPT chat logs during copyright litigation [13]. Prompts can also contain personal or confidential data, so they need classification and retention rules.
+How does AI help data governance?
AI-driven classification can label structured and unstructured data with business, regulatory and consent context at scale, as OneTrust describes [23]; Securiti describes automated discovery and classification of data across cloud and shadow assets [21]. Treat vendor accuracy claims as unverified until tested on your own data.
Sources
- [1]Gartner, Lack of AI-ready data puts AI projects at risk (2025-02-26).
- [2]Standards Council of Canada, ISO/IEC 8183:2023 — Artificial intelligence — Data life cycle framework (2023-07-25).
- [3]ISO, ISO/IEC 5259-1:2024 — Data quality for analytics and machine learning — Part 1: Overview, terminology and examples (2024-07-02).
- [4]European Commission (AI Act Service Desk), Article 10: Data and data governance (consolidated text as at 27 July 2026).
- [5]European Commission (AI Act Service Desk), Article 4a: Processing of special categories of personal data for bias detection and correction.
- [6]White & Case, EU AI Omnibus enters into force, amending the AI Act (2026-08-04).
- [7]European Commission (AI Act Service Desk), Article 26: Obligations of deployers of high-risk AI systems.
- [8]European Commission (AI Act Service Desk), Article 27: Fundamental rights impact assessment for high-risk AI systems.
- [9]CMS, EDPB Opinion 28/2024: key takeaways on processing personal data in the context of AI models (2024-12).
- [10]IAPP, EU member states finalize Digital Omnibus negotiating stance (2026-10-22).
- [11]IAPP, A view from Brussels: Simplification, barely; uncertainty, for sure (2026-04-23).
- [12]Euronews, Italy’s privacy watchdog fines OpenAI €15 million after probe into ChatGPT data collection (2024-12-20).
- [13]Engadget, OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions (2025-10-11).
- [14]Microsoft, hosted by NIST AI Resource Center, NIST AI RMF to ISO/IEC FDIS 42001 AI Management System Crosswalk.
- [15]ISMS.online, ISO 42001 Annex A control A.7: Data for AI systems.
- [16]NIST, AI RMF Core (GOVERN, MAP, MEASURE, MANAGE).
- [17]Crowell & Moring, California’s AB 2013 requires generative AI data disclosure by January 1, 2026 (2025-12-08).
- [18]Colorado General Assembly, SB26-189: Automated Decision-Making Technology (bill page and summary) (2026).
- [19]Collibra, Collibra AI Governance.
- [20]BigID, BigID AI governance.
- [21]Securiti, Securiti (DataAI Command Platform overview).
- [22]Veeam (Business Wire), Veeam completes acquisition of Securiti AI (2025-12-11).
- [23]OneTrust, Data Use Governance.
- [24]OneTrust, OneTrust unveils new data governance solution to close the enforcement gap for AI-ready data (2025-05-08).
- [25]IBM, IBM watsonx.governance: model governance.
- [26]IBM, Governing assets with watsonx.governance (documentation).
- [27]Dataiku, Dataiku Govern.
- [28]IAPP, AI Governance Profession Report 2025 (2025-04-16).
- [29]Gartner, Gartner survey finds 45% of organizations with high AI maturity keep AI projects operational for at least three years (2025-06-30).
- [30]OneTrust, OneTrust using OneTrust: managing AI (webinar).
- [31]European Data Protection Board, EDPB opinion on AI models: GDPR principles support responsible AI (2024-12-18).
How we keep this page current
Every claim links to its source. We re-check this page at least every 60 days, and sooner when a standard, law or vendor changes; vendor names, acquisitions and shutdowns update automatically from our directory. Vendors cannot pay to be included or ranked.