Map data by its role in the service
Begin with the path through the application. Identify data used for development, evaluation, retrieval and live operation, along with outputs that are retained or reused. These roles have different risks. A dataset suitable for a controlled test may not be permitted for production use, and production feedback may not be reliable enough to become training material without review.
Annex A.7 of ISO 42001 addresses data-related controls, including acquisition, quality, provenance and preparation. The relevant controls are selected through the treatment process. The data records and checks in this article are practical suggestions for implementing a suitable approach, not a reproduction of the standard’s requirements. Reference: ISO/IEC 42001:2023.
Distinguish source data from the representation the AI system uses. Documents may be cleaned, split, indexed or transformed before retrieval. Sensor readings may be filtered or aggregated. Each step can remove context or introduce error. Record the transformations that matter to the use, rather than assuming that access to the original source explains the processed result.
Name an owner for each important data decision. The technical team may run a pipeline while another function owns the source and its permitted use. Clarify who can approve a new source, correct an error, change a preparation rule and decide whether a limitation is acceptable. Shared ownership should not mean that nobody can make the decision.
Establish origin and permission before use
Record the source, acquisition route, date or period, relevant version and known restrictions. For supplied data, retain the agreement or other basis on which it is used. For internal data, identify the source owner and approved purpose. Do not infer unrestricted use merely because the data is accessible to an employee or available online.
Separate technical availability from permission. A connector may be able to read a document collection while the intended service is not approved to expose all of it. A dataset may contain material subject to contractual, privacy or intellectual-property constraints. Involve the appropriate specialists where those issues arise and retain the decision relevant to the deployment.
Track permissions through transformations and copies. A restriction can be lost when data moves into a new store or is combined with another source. Include derived datasets, indexes and cached content in the handling design. Deleting the original file may not remove every representation used by the service.
Ask suppliers precise questions about data handling. Establish what is transmitted, retained, used for service improvement or shared with subprocessors under the actual arrangement. Verify the applicable service configuration and terms. Broad statements about a vendor’s products may not describe the endpoint or account settings your organization uses.
Define quality for the task
Quality is fitness for a particular use, not a universal label attached to a dataset. Identify the characteristics that affect the service: completeness, accuracy, timeliness, consistency, representativeness or the reliability of labels. A large dataset can still be unsuitable if it omits the operating conditions that matter most.
Set acceptance criteria with domain input. A customer-support assistant may need current policy documents and clear ownership of conflicting guidance. A forecasting tool may need stable units, time alignment and knowledge of missing periods. Explain why each criterion matters and how a failure affects the deployment decision.
Use checks at the point where they can prevent a problem. Validate structure and required fields at ingestion, detect unusual distributions where appropriate and review content conflicts before indexing. Not every issue can be resolved automatically. Define a route for uncertain or disputed data rather than forcing all inputs through a binary pass or fail rule.
Retain quality findings and their treatment. If a gap is accepted for a limited pilot, record the conditions and the evidence to be collected. Avoid turning a known limitation into a permanent assumption without review. Quality decisions should remain connected to the service’s approved audience and operating range.
Preserve provenance through preparation
Provenance should help a reviewer understand the path from source to use. Record relevant transformations, selection decisions, versions and responsible people or processes. The level of detail should fit the risk and the need to investigate. A simple source label may be enough for some low-consequence material and insufficient for a consequential dataset with several preparation stages.
Document choices that change meaning. Removing outliers, merging categories, translating text or excluding incomplete records can alter what the system learns or retrieves. Explain the reason and evaluate the effect. A cleaning step should not silently remove difficult cases that the live service will still encounter.
Keep label definitions and review methods clear. Human labels can reflect disagreement, missing context or inconsistent interpretation. Record how labels were created, what instructions were used and how material disputes were resolved. A label called ground truth should not imply certainty that the process does not support.
Version preparation logic alongside data snapshots where practical. If a pipeline changes, identify which evaluations and deployments depend on the old result. This supports targeted retesting and investigation. Without that connection, teams may know that a dataset changed but not which decisions relied on it.
Protect evaluation integrity
Keep evaluation data sufficiently independent from development and tuning to support the intended claim. Repeatedly adjusting a system to a small known test set can make the reported result less representative of future use. Explain how test cases were selected and what protections exist against inappropriate reuse or leakage.
Include important operating variations. Language, document quality, rare events, equipment state and user behaviour may affect performance. The relevant variations depend on the service. Do not claim broad suitability from a convenient sample that excludes difficult conditions. Where coverage is limited, make the release restriction explicit.
Control access to sensitive evaluation material. Tests can contain confidential examples, security scenarios or personal information. Use synthetic or minimized data where it can answer the question, and govern representative data carefully where it is needed. Evaluation evidence should not create an uncontrolled copy of production information.
Review expected outputs and scoring rules. A test can be reproducible yet measure the wrong thing. Domain reviewers should confirm that the expected answer or outcome fits the task. Keep disagreements and uncertainty visible where they affect interpretation. The evaluation should support a decision, not merely produce an attractive score.
Example: a knowledge assistant with conflicting policies
Imagine a software provider building an assistant over internal support material. The retrieval pipeline ingests documents from several teams. Some articles are current, others describe retired features, and a few contain customer-specific workarounds. The initial quality check confirms that files can be read, but it does not establish that the content is suitable for general support advice.
The data owner introduces source categories, validity periods and audience restrictions. Documents without a responsible owner or clear status are held from the general index pending review. The pipeline preserves the source link and relevant version so that staff can inspect the basis of a response. Customer-specific material is handled through a separate access-controlled route.
Evaluation includes questions where old and new guidance conflict. The team checks whether the assistant retrieves the intended source and whether the response makes limitations clear. It also tests what happens when a document is withdrawn: the source, derived index and cache must reflect the change according to the service design. A successful document deletion alone is not sufficient evidence.
The operating process assigns responsibility for content changes and monitors recurring corrections. The governance improvement is not a claim that every answer will be correct. It gives the service a defensible data boundary and a way to identify and correct a known source of unreliable advice.
Example: sensor data for industrial prediction
Consider a manufacturer using sensor data to predict equipment maintenance needs. Historical records combine readings from devices with different calibration histories and sampling rates. A data scientist can train a model on the combined table, but production engineers explain that gaps and changes in units have operational meaning.
The team records sensor origin, units, sampling intervals and known maintenance events. Preparation rules align time series and flag missing periods rather than silently filling every gap. Domain experts review whether the chosen treatment preserves the patterns needed for the task. The dataset’s provenance links back to equipment and relevant configuration changes.
Evaluation includes equipment states not well represented in the initial data. The team discovers that performance is uncertain after a particular maintenance procedure. The deployment is therefore limited while additional evidence is collected, and operators receive instructions for that condition. The risk record reflects the data limitation and the practical fallback.
When a supplier replaces a sensor model, the data change triggers review before the previous evaluation is assumed valid. This is the value of provenance: it connects a physical change to the evidence supporting the AI service. The organization can investigate why results changed without relying only on the model provider’s explanations.
Monitor data changes during operation
Choose signals that relate to the service’s assumptions. Changes in missing values, input length, language mix, source freshness or sensor distribution may indicate that existing evidence no longer fits. A signal should have an owner and a response route. Collecting statistics without deciding what they mean will not improve control.
Distinguish data drift from harmful performance change. A changed input distribution can be a reason to investigate, but it does not automatically prove the service has become unsuitable. Conversely, stable distributions do not guarantee correct outcomes. Combine data monitoring with performance, complaints and operating observations.
Manage feedback before reuse. User corrections can be valuable, but they may contain mistakes, sensitive data or attempts to manipulate the system. Define validation and approval before feedback changes a dataset or behaviour. A feedback loop should not become an unreviewed path from arbitrary input to production changes.
Plan retention and deletion across the lifecycle. Identify what must remain to explain decisions and what should be removed when no longer needed. Include supplier copies and derived representations where relevant and contractually controllable. Verify the available mechanisms instead of promising deletion outcomes that the architecture cannot support.
Test whether the records are usable
Select one output and ask what source information and transformations contributed to it, within the technical limits of the service. Then select one source change and identify the deployments that may be affected. These two tests reveal whether provenance works in both directions. Perfect reconstruction may be impossible, but the limitation should be understood and managed.
Review recurring quality failures for process causes. Conflicting documents may indicate unclear content ownership. Repeated missing fields may indicate an unsuitable intake interface. Inconsistent labels may indicate poor instructions or insufficient domain review. Correct the source of the problem rather than repeatedly cleaning the same defect downstream.
For a scoped support discussion, bring a data-flow map, a sample provenance record and one difficult quality decision. Ask how the proposed approach will improve release evidence and operational response. A data catalogue is valuable when it answers those questions, not simply when it contains many fields.
Frequently asked questions
Do data controls matter when we do not train a model?
Yes. Retrieved content, prompts, evaluation material and production inputs can affect the service. Assess your actual data activities and responsibilities, including what a supplier handles and what remains under your control.
Is provenance the same as a source URL?
A source URL can be one element, but provenance may also need version, acquisition, preparation and use history. Choose detail that supports the relevant decisions and investigation needs. A live link alone may not identify what was used earlier.
Can we use synthetic data for all testing?
Synthetic data can answer some questions, especially boundary and workflow tests. It may not represent all real operating conditions. Explain what the test can establish and where appropriately governed representative evidence is still needed.
Should every data change block release?
Use a screening method based on the possible effect on assumptions and risk. Minor corrections and material new sources need different review. Record the decision and repeat relevant evaluation when existing evidence no longer supports the use.