An artificial intelligence assistant can retrieve a document quickly and still give an employee the wrong answer. The document may contain an expired policy, an incomplete customer record or a draft that should never have been treated as authoritative. Improving the model will not automatically resolve these weaknesses in the information around it.
For businesses deploying AI, data ownership determines who can explain a record, approve its use and correct it when circumstances change. Clear responsibility helps organisations assess whether information is suitable for a particular application. It also connects technical performance with the business consequences of an inaccurate output. The objective is dependable use within a defined purpose, supported by evidence and a practical route for resolving problems.
Defining the decision before collecting the data
A project needs a specific description of what the system will help someone do. Summarising internal guidance differs from recommending a customer action or automatically changing an account. Each purpose creates different requirements for accuracy, access and review.
The NIST AI Risk Management Framework provides a voluntary approach to managing AI risks across design, development, use and evaluation. Its relevance for business leaders is that risk assessment belongs alongside the intended application, rather than being postponed until a model has been selected.
A useful project description identifies the users, the information they need and the consequences of an incorrect response. This makes it easier to distinguish data that is necessary from material collected because it happens to be available. An unrestricted document collection may add complexity without improving the intended decision.
Responsibility should also extend beyond the technology team. A department that creates a policy understands its commercial meaning and effective date. Engineers can implement retrieval and access controls, but they need that business context to determine which version should be available and how changes should be reflected.
Distinguishing source quality from model accuracy
Data quality and system performance are related, but they are different assessments. A model can produce the wrong output from accurate records. It can also accurately reproduce a statement that is outdated or unsuitable for the question being asked.
The UK Information Commissioner's guidance on accuracy and statistical accuracy distinguishes the accuracy of personal information from statistical measures of model performance. The distinction remains useful when evaluating business systems, although organisations must assess their own applicable data protection obligations separately.
Consider a hypothetical distributor using an assistant to answer questions about returns. Its archive contains an old policy offering a longer return period. If the assistant retrieves that document and describes it clearly, the response can still mislead the employee. The source needs correction or exclusion, while the retrieval process needs a way to recognise current authority.
Testing should therefore examine both the underlying records and the resulting answers. A high average performance score does not establish that the system handles rare exceptions correctly. Some mistakes matter far more than others because they affect payments, contractual commitments or the treatment of a customer.
Knowing where information came from
Provenance means knowing the origin and relevant history of information. In a business application, that can include who created a document, when it became effective and whether someone approved it. These details help users understand the basis of an answer and investigate discrepancies.
The NIST Generative AI Profile includes suggested documentation covering data provenance, quality, limitations and evaluation. It is guidance for risk management, not a certification that an application is safe or suitable for every task.
Businesses should distinguish how data enters their particular system. Information used to train or adjust a model differs from documents retrieved when answering a question. A supplier's statement about one process may not describe the other. Contracts and technical assessments need to address the actual arrangement rather than relying on the broad label of AI.
For an internal assistant, a visible source reference can support checking, but its presence does not prove the answer is correct. The cited document must actually support the statement, and the system must use the relevant passage accurately. Users need enough time and access to perform that review when it matters.
Assigning authority over changes
A named owner needs a defined scope of responsibility. Someone may maintain a customer database without being authorised to approve every new use of its contents. Ownership should identify who can resolve quality issues, who approves access and who assesses whether the proposed purpose is appropriate.
The OECD AI Principles include accountability and transparency, recognising responsibilities across the AI lifecycle. For an operating company, these principles support a practical question: can management identify who made a material decision and the information on which it depended?
A change process can link the business owner with the technical team. When a product is withdrawn or a policy amended, the owner identifies the change and the system team updates the relevant information source. A subsequent check confirms that the application no longer presents the superseded instruction as current guidance.
This responsibility continues after launch. A system that passed an initial test may deteriorate as documents accumulate or business rules change. Periodic reviews should focus on the information most likely to affect important decisions, rather than treating every record as equally consequential.
Controlling access to sensitive material
Combining information can change what a user is able to discover. An employee might previously have needed access to several separate systems to assemble a sensitive account history. An assistant could make that history easier to retrieve unless permissions are enforced throughout the process.
Access assessment should consider what the user can request and what information the system can retrieve. It should also consider logs, exports and other places where the resulting content might appear. Restricting one interface is insufficient if the same information becomes available through another route.
Testing can include questions designed to expose confidential material or cross boundaries between teams. The organisation needs a response when the system retrieves something it should not reveal. Depending on the circumstances, that may involve restricting a source, changing permissions or suspending the affected function while the issue is investigated.
Useful information should remain accessible to authorised users. Excessively broad restrictions can make a service ineffective and encourage informal workarounds. The aim is to align access with the task and record the reasons for any material exception, supported by appropriate security and privacy assessment.
Evaluating realistic questions and exceptions
An evaluation set should resemble the work people will actually perform. Straightforward questions can establish basic functionality, but ambiguous requests and unusual cases reveal different weaknesses. Reviewers need reliable expected answers and an explanation of what would make a response unacceptable.
A hypothetical procurement assistant might correctly summarise standard purchasing limits while mishandling a temporary exception for a particular contract. That case requires more than fluent wording. It requires the relevant exception, its validity period and the approval conditions to be interpreted together.
Domain specialists should help design these tests because they understand where mistakes have practical consequences. Their involvement also helps establish when the system should decline to answer or direct the employee to an authorised colleague. An explicit limit can be more useful than an unsupported attempt to satisfy every request.
Monitoring needs to distinguish technical failures from content problems. Slow retrieval, missing documents and an inaccurate interpretation call for different remedies. Recording the failure category and its consequence helps management decide whether the source, the application or the surrounding process needs attention.
Retiring information when its purpose ends
Information can remain available after the business reason for using it has ended. A completed project, expired licence or withdrawn product can leave documents in a shared collection. Ownership needs to include decisions about retirement, subject to applicable retention requirements and any continuing operational need.
Removing a source also requires coordination. The organisation should understand whether copies, indexes or derived records remain accessible elsewhere in the application. A change record can identify what was removed, why the decision was made and which checks confirmed the result. This gives future reviewers a way to distinguish deliberate limitations from an accidental gap in coverage, and helps users understand why an answer may no longer be available.
Budgeting for continuing stewardship
Data ownership creates work that needs resources. Employees must review records, investigate errors and test changes. A project budget that covers software alone can understate the continuing cost of maintaining a useful service. Responsibility without time or authority is unlikely to resolve recurring issues.
Management can assess the value through specific outcomes, such as fewer incorrect internal answers or less time locating approved guidance. These measures should include the effort required for review. Faster first responses are not an improvement if colleagues spend more time correcting them afterwards.
Clear data ownership supports a more realistic decision about where AI belongs. It helps a business understand which information is dependable, which uses need further controls and which tasks still require direct expertise. As the application evolves, that understanding gives managers a practical basis for improving the service and limiting its use where the evidence remains insufficient.
References
1. National Institute of Standards and Technology — AI Risk Management Framework.
https://www.nist.gov/itl/ai-risk-management-framework
2. Information Commissioner's Office — What do we need to know about accuracy and statistical accuracy.
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/what-do-we-need-to-know-about-accuracy-and-statistical-accuracy/
3. National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework Generative Artificial Intelligence Profile, July 2024.
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
4. OECD — AI Principles.
https://oecd.ai/en/ai-principles
