AI-Ready Data: Assessing Your Data – The Data Quality Dimensions That Determine AI Success

Many AI initiatives do not fail because of the model, but one level deeper: because the underlying data is incomplete, inconsistent, or not legally usable. This may only become apparent after significant investments have already been made. This article explains how financial institutions can assess whether they have AI-ready data before AI models go into production — across nine data quality dimensions with specific assessment questions, placed within the regulatory framework ranging from the EU AI Act to BCBS 239. It also explains why the four traditional data quality dimensions alone are not sufficient for AI initiatives.

Data readiness for AI is one important aspect of AI readiness: an organization’s ability not merely to experiment with AI on an isolated basis, but to transition it into operations reliably, at scale, and in compliance with applicable regulations. A suitable data foundation, a viable strategy with prioritized use cases, the right technical infrastructure, and robust governance (including the necessary skills and clearly defined roles) are the elements that need to work together.

What Is AI-Ready Data? The Data Foundation in the Context of AI Readiness

The data foundation comprises the entirety of the data an AI model accesses. This means not only the individual training dataset, but the complete body of data across the entire lifecycle: data used for pre-training, fine-tuning, or a RAG architecture; validation and test data; as well as the data the model subsequently processes in production. The data foundation encompasses not only the data values themselves, but also their structure and meaning (metadata and definitions) to enable semantic understanding.

The provenance of the data and its processing path, from the source to the model, must also be taken into account, as must the governance framework that defines who is permitted to use which data and for what purpose.

This also establishes the conceptual distinction: the data foundation is the subject of the assessment; whether it qualifies as AI-ready data is the judgment resulting from a systematic, criteria-based evaluation.

AI-Ready Data Is Not the Same as BI-Ready Data

It is worth separating AI-ready data from data that is merely fit for conventional business intelligence, because the two are sometimes conflated. BI systems are built for retrospective analysis: they report what happened, aggregated along predefined dimensions and structured according to a fixed schema. For example, a BI system can show revenue by product category, but it cannot explain what that figure means for the business.

AI applications need more than the aggregate. They need the relationships between data domains, the semantic context that gives a value its meaning. A data foundation that has served periodic risk reporting reliably for years is therefore not automatically AI-ready data.

Why the Data Determines the Success of AI Projects

AI readiness is often discussed primarily as a question of infrastructure: Which platform? Which model? What budget? These are legitimate questions, but they do not go far enough. No AI application can be better than the data on which it is based. Everyone is familiar with this statement, but not everyone draws the operational conclusion from it: the need for a systematic, criteria-based assessment of the data before an AI application goes into production.

When AI applications are based on incomplete, inconsistent, or outdated data and consequently produce inaccurate results, even excellent model technology can do little to resolve the underlying problem.

Assessing data readiness for AI should therefore not be treated merely as part of a technical preliminary project. It is a strategic prerequisite for any successful AI readiness initiative. Organizations that take this seriously do not start with model selection; they start with a structured assessment of whether their data foundation is fit for purpose.

There is also a cost dimension that only becomes visible across several projects. Where data problems are addressed reactively for each AI initiative the same work is repeated every time. Each new use case carries the full cost of sourcing, cleaning, reconciling, and documenting its data again. Over time this makes further AI projects progressively harder to justify. Treating data readiness for AI as a shared foundation rather than a per-project task is what turns that curve around.

Third-Party AI Models: Who Is Responsible for the Data Foundation?

If a company does not train an AI model itself but instead purchases one, it has no influence over the training data.

First, this means that the company needs assessment criteria, precisely because, as the deployer, it does not control the training data itself. An organization that trains a model itself can make improvements. An organization that buys one in can only assess it, establish contractual safeguards, and monitor it. This does not make the suitability assessment unnecessary; rather, it makes it the only remaining means of control.

Second, it must be taken into account that if the data foundation encompasses more than the provider’s training dataset, the other data is controlled by the deployer. These data can themselves cause an AI application to produce erroneous output, for example, an outdated document repository in a RAG architecture.

Furthermore, it should be noted that under Article 25 of the AI Act, a deployer can become a provider if it markets a system under its own name, substantially modifies it, or changes its intended purpose.

The Regulatory Framework: Is Data Quality for AI a Regulatory Requirement?

When a financial institution assesses data readiness for AI as part of its risk management, it is no longer operating in a regulatory vacuum. Data quality is addressed in several overlapping regulatory frameworks.

For significant European institutions, the principles of BCBS 239 on risk data aggregation apply. The ECB has further specified these supervisory expectations through its RDARR Guide, published in May 2024. For less-significant institutions in Germany, the MaRisk requires institutions to ensure adequate data quality for risk management and risk reporting. The data integrity requirements under DORA point in the same direction: robust, reliable, and traceable data are a regulatory necessity.

The GDPR also establishes an important framework. Its principle of accuracy makes data quality a direct legal requirement for personal data, while at the same time providing the basis for the legally sensitive handling of bias and personal data.

EU AI Act Article 10: Data Governance and Data Quality Requirements

In addition, an AI-specific regulatory layer is emerging. Article 10 of the EU AI Act sets out the data requirements for high-risk AI systems: training, validation, and test data must, to the extent possible, meet defined quality and data governance requirements. These high-risk obligations have been postponed to December 2, 2027, by the “Digital Omnibus.”

A structured AI readiness assessment of the data is therefore also an instrument for minimizing regulatory and legal risks.

The Four Classic Data Quality Dimensions for AI Models

The established data quality literature identifies, in particular, the following four dimensions for assessing data quality. They provide the necessary starting point.

1. Accuracy: Correctness and Precision of the Data

Data must accurately reflect reality. Correctness refers to the degree to which the data corresponds to the actual facts, that is, whether the value is correct. Precision concerns variations that may occur when measurements required to determine data points are repeated.

Assessment question: Are defined tolerance ranges for deviations in data values in place, and are they actively monitored?

2. Completeness of the Data

A dataset is complete when all relevant data points are present, that is, there are no missing values. For AI applications, however, completeness is not a binary property. It is not only the number of missing values that matters, but also why those values are missing.

Missing values may occur purely at random, for example due to unreliable data collection processes.

More critical is when the missingness itself follows a pattern — when values are systematically missing more frequently for certain groups, periods, or characteristics than for others. If, for example, income information is disproportionately missing for a particular customer group, the model learns from a biased subset of reality. In this case, the problem is not the number of gaps, but their systematic nature.

Assessment question: Are missing values randomly distributed, or do they follow a pattern that could bias the model?

3. Timeliness of the Data

Data must be up to date. Timeliness is a relative measure: it has to be assessed against what the specific AI use case actually requires.

Assessment question: How frequently is the data updated, and does this frequency correspond to the time horizon of the AI application?

4. Consistency of the Data

In the context of data quality, consistency refers to the uniform and matching representation of the same facts across different data sources, systems, and points in time.

In the AI context, consistency extends beyond the technical alignment of values to the business meaning attached to them. The same field label can carry different definitions in different divisions: one system’s “customer” is the contract holder, another’s is the beneficial owner. As long as these datasets are used separately, the divergence remains invisible and causes no harm. Once they are linked across the organization to give an AI model the cross-domain view it requires, that divergence becomes a defect in the model’s understanding of reality.

Assessment question: Is there a central definition and metadata structure applied consistently across all systems, and is the business meaning of each relevant field documented so that data can be linked reliably across domains?

Why Structured Data Is No Longer a Requirement for AI

The four dimensions of accuracy, completeness, timeliness, and consistency are relevant to any data-processing application. One requirement, however, becomes considerably less significant in the AI context, and it is one that some data readiness checklists still carry over unexamined: the traditional structuring of data.

Rule-based systems and conventional reporting require data to be available in fixed tables and schemas. In those contexts, structuring genuinely was a hard requirement. AI models, by contrast, process precisely the types of data that were previously difficult or impractical to use: contracts, emails, images, and audio. For AI, data structuring is therefore no longer a requirement. The ability to work with unstructured data is part of AI’s value proposition, and treating structure as a precondition would exclude much of the data that makes AI worth deploying in the first place.

What replaces structure is context: not a fixed schema, but documented meaning and reliable linkage.

Five Additional Data Quality Dimensions for AI-Ready Data

What does remain relevant for AI are five additional dimensions that determine whether a data foundation qualifies as AI-ready data.

5. Data Availability and Usability

Data can be of high quality and still be unusable for model operations if it is inaccessible, cannot be processed, or has not been authorized for use.

Data availability in the narrower sense is also one of the three objectives of information security, alongside confidentiality and integrity. It means that the required data can be reached at all, and reached reliably — protected against outage, loss, or blockage. In the AI context, this security objective is necessary but not sufficient; it needs to be complemented by three further aspects.

Data must first be discoverable. In grown system landscapes, the practical obstacle is frequently not that data is locked down, but that no one knows it exists, or how reliable it is. A data catalogue answering those four questions is the prerequisite for everything else in this dimension: access cannot be authorized for data that has never been identified, and a use case cannot be scoped against a data foundation that has never been inventoried. Discoverability is also what makes the difference between a one-off data sourcing exercise and a foundation that the next AI use case can build on.

Data must also be available in a form that machines can actually process which, in the case of AI applications, is not the same thing as structured data.

Access to the data must be legally and organizationally authorized. This dimension also touches on the security objective of confidentiality — sufficient access rights are its counterpart. Availability is therefore not merely a question of IT implementation, but also a matter of data governance for AI: Who is permitted to use which data, and for what purpose?

Accordingly, authorization applies to the purpose, not to the dataset itself. An existing authorization for a conventional system covers a particular processing purpose. Using the same data in an AI system will regularly constitute a change of purpose that must be reassessed. In conventional systems, “authorized” is a persistent status; in the AI context, it can only be a statement made for a specific use case.

Assessment question: Is there a catalogue documenting which data exists, and are data accesses legally, organizationally, and technically authorized for the AI use case?

6. Data Representativeness and Data Bias in AI

A dataset can meet the four fundamental data quality dimensions and still be unsuitable for training an AI model if it structurally underrepresents or overrepresents certain groups, periods, or scenarios. A lack of data representativeness may be associated with bias in the data. Data bias in AI, in turn, can lead to bias in the model itself or even to discrimination when the AI is used.

Assessment question: Does the data reflect the actual composition of the target population?

7. Data Lineage Requirements Under the EU AI Act

For every data point that feeds into a high-risk AI system, it must be possible to trace where it originated, how it was transformed, and which quality checks it underwent. Data lineage, that is, the documented origin and complete transformation path of a data point, is the central instrument for establishing this traceability. It maps the entire chain of data preparation. Data lineage requirements for AI arise not only from internal quality objectives, but also from regulatory requirements such as the EU AI Act and BCBS 239.

Particularly with regard to data lineage, it is advisable not to wait until 2027 to begin implementation, despite the postponement of the AI Act’s high-risk obligations. Missing documentation of this kind cannot be reconstructed retrospectively, and even a generous budget later cannot remedy such a deficiency.

Assessment question: Is the complete transformation path of every data point from its source to the model documented?

8. Uniqueness of the Data

Each real-world entity — a customer, a contract, a transaction — should be represented exactly once in the data. Duplicate entries are problematic in any data-processing system; in AI training, a model interprets duplicate data points as more frequent events and consequently assigns them disproportionate weight. The model therefore learns a version of reality in which certain cases appear to occur more frequently than they actually do. To prevent this, consistent entity-level deduplication is required.

Assessment question: Is there a defined entity-resolution strategy that systematically identifies and resolves duplicates?

9. Fitness for Purpose

The overarching dimension: various criteria have been discussed, but the appropriateness of relying on them also depends on assessing them in relation to the specific use case. Data that is sufficient for periodic reporting may be entirely unsuitable for real-time risk monitoring. To account for such differences, fitness for purpose requires a separate set of data foundation requirements to be defined for each AI use case.

Assessment question: Has a requirements profile for the data foundation been defined for each AI use case against which its suitability can be assessed?

Key Takeaways: A Data Readiness Checklist for AI Projects

When assessing whether you have AI-ready data, the finer nuances determine both operational success and compliance with regulatory requirements. Failure to meet the data quality dimensions above creates risks in the operation of AI:

Limited generalization capability: Incomplete data leads to poor AI output for previously unseen cases.

Undetected distortion: Inaccurate data can distort AI output without this being immediately apparent.

Inconsistencies: Inconsistent data from different sources introduces processing errors into the AI pipeline.

Statistical overweighting: Undetected duplicates in the data distort AI output.

Risk of discrimination: Lack of data representativeness can lead to data bias in AI and to discrimination.

Drift from reality: Outdated datasets fail to reflect current developments.

Project termination before go-live: Insufficient data availability prevents the AI application from being put into operation.

Regulatory sanctions: Incomplete data lineage puts compliance with the EU AI Act and BCBS 239 at risk.

Strategic misalignment: Without documented fitness for purpose, the AI system may fail to meet its intended use case.

Conclusion and Practical Recommendation

Organizations that do not want to leave the success of their AI projects to chance or risk regulatory sanctions should establish a standardized set of criteria for assessing their data. A robust AI readiness assessment is the strategic foundation of any AI strategy that is intended to deliver genuine value. Only once the organization has AI-ready data should it move toward AI development and scaling.

Frequently Asked Questions About AI-Ready Data and Data Quality for AI

What is AI-ready data?

AI-ready data is data that meets the quality, governance, and documentation requirements of a specific AI use case across its entire lifecycle. Data readiness for AI is assessed across nine dimensions: accuracy, completeness, timeliness, consistency, availability and usability, representativeness, data lineage, uniqueness, and fitness for purpose.

What are the data quality dimensions for AI?

The four classic data quality dimensions provide the necessary starting point: accuracy, completeness, timeliness, and consistency. For AI applications, however, these are not sufficient. At least five additional dimensions need to be considered: availability and usability, representativeness, data lineage, uniqueness, and fitness for purpose.

Does AI require structured data?

No.Structuring was a hard requirement for rule-based systems and conventional reporting, which depend on fixed tables and schemas, but AI models process precisely the data types that were previously impractical to use: contracts, emails, images, and audio. Structure is not what makes data AI-ready; documented meaning and reliable linkage across data domains are. Criteria lists that still name “structured” as a requirement for AI-ready data are carrying over an assumption from the business intelligence era.

Is data quality for AI a regulatory requirement?

For AI applications that are subject to regulatory requirements in the financial sector, data quality is a regulatory obligation. Data quality for AI is therefore not merely a best practice; it is addressed in several regulatory frameworks. For financial institutions, these include MaRisk, BCBS 239, and DORA, with the GDPR applying across the board. The EU AI Act adds an AI-specific regulatory layer. A structured AI readiness assessment of the data therefore directly contributes to minimizing regulatory risks.

What does EU AI Act Article 10 require on data governance and data quality?

Article 10 of the EU AI Act requires training, validation, and test data for high-risk AI systems to meet stringent quality and data governance requirements. This will make requirements for AI training data — including representativeness, data governance, and documented provenance — binding for the first time.

What requirements apply to AI training data?

AI training data should adequately represent the target population in order to reduce data bias in AI and the risk of discrimination, and it must be deduplicated. If an organization does not train AI models itself but instead uses pre-trained or third-party models, data governance considerations shift to the supply chain. In this case, it is critical to assess what documentation the provider makes available regarding the provenance, composition, and lawfulness of the training data.

What data lineage requirements apply under the EU AI Act?

For every data point that feeds into a high-risk AI system, it should be possible to trace where it originated, how it was transformed, and which quality checks it underwent. These data lineage requirements for AI arise from internal quality objectives and, depending on the context of the application, also from regulatory requirements such as the EU AI Act and BCBS 239.

What does data availability mean in the AI context?

Data availability in the AI context means more than simply being able to access data from a technical perspective. It encompasses four aspects: ensuring that the required data can be accessed reliably and is protected against outages (the availability objective of information security); ensuring that the data is discoverable, meaning a catalogue records what exists, where it is held, who owns it, and how reliable it is; ensuring that the data is available in a machine-processable form; and ensuring that access has been legally and organizationally authorized. The last two points in particular make availability a matter of data governance for AI, not merely of IT implementation.

What is data governance for AI?

Data governance refers to the framework of rules, roles, responsibilities, and processes that defines who may use which data for what purpose and how the quality, provenance, and protection of data are ensured. In the context of AI readiness, data governance is not an end in itself. Rather, it is the organizational prerequisite for reliably enforcing the criteria for AI-ready data — from availability and data lineage to fitness for purpose.