How to Strengthen AI Governance at Your Organization

Editor’s note: This is the second of a two-part series on Data Quality and AI Governance. The first article was published in July and examined how data quality is foundational because it determines whether an AI system can be trusted, assessed, or governed at all.

Part 1 of this series explained that data quality is the foundation of artificial intelligence (AI) governance in healthcare because it shapes whether AI systems can be trusted, assessed, explained, and governed. That foundation matters because the consequences of poor-quality data are not theoretical. They play out in safety, equity, compliance, trust, and operational performance.

This article turns from the foundation to the implications. It examines the concrete risks that emerge when AI is used with weak data, the themes that reinforced this lesson at the HIMSS 2026 Global Health Conference, what good data quality looks like in practice, and the actions AHIMA members, credential-holders, and others can take to strengthen governance in their organizations.

Risks of AI Use with Poor-Quality Data

When organizations use AI with poor-quality data, the risks are not abstract.

One risk is patient safety and quality. In healthcare environments, inaccurate or incomplete source data can lead to faulty AI outputs, misleading summaries, misclassification, or inappropriate prioritization. Even when a tool is used in operational or nonclinical contexts, flawed data can still affect access, communication, coordination, and documentation in ways that impact patient experience and downstream care.

A second risk is bias and inequity. If the data used to train, tune, or operate an AI system underrepresent certain populations, reflect historical disparities, or contain biased proxies, the resulting system may produce uneven performance across groups. Governance teams cannot meaningfully assess fairness unless they understand who is represented in the data, who is missing, what variables are being used, and how those variables were collected and transformed. The World Health Organization (WHO) and Organisation for Economic Co-operation and Development (OCED) both emphasize that trustworthy AI must account for fairness, human rights, and robust governance. These principles become difficult to operationalize when the data foundation is weak.

A third risk is distortion in documentation, medical coding, and the revenue cycle. Weak source data can propagate and compound errors in environments where AI is used to support summarization, medical coding review, autonomous medical coding, or operational decision-making. AI can scale inconsistency just as easily as it can scale efficiency. When the underlying data is unreliable, organizations may produce outputs faster, but not better.

A fourth risk is compliance and legal defensibility. As transparency expectations grow, organizations are increasingly expected to understand and document data sources, model logic, intended use, validation evidence, and limitations. Those who cannot explain the sources, attributes, and limitations of AI-enabled tools will find it difficult to explain their reliance on those tools during audits, investigations, or public scrutiny.  

Trust is difficult to build and easy to lose, leading to the fifth risk: operational mistrust. Staff quickly lose confidence in AI systems that produce inconsistent, obviously flawed, or contextually inappropriate outputs. Once clinicians, HI professionals, medical coding teams, or operations leaders perceive an AI tool as unreliable, governance becomes reactive rather than strategic.

Moving from Pilot Projects to Adoption

In March, the HIMSS 2026 Global Health Conference underscored several themes clearly reflected and reinforced this thesis in practical ways. Organizations are moving beyond pilot projects toward enterprise-scale adoption, but the path to scaling AI safely depends on trusted data, strong metadata, effective governance, semantic consistency, and workforce readiness.

Several HIMSS sessions were closely aligned with the argument that data quality is foundational. Across organizations, the maturity curve appears to be moving from pilot use cases to governance to trusted data and platforms, and only then to scaled operational deployment. One presentation on enterprise semantic models reinforced this point by showing that reusable definitions, metadata standards, governance structures, and stakeholder alignment are all prerequisites for trustworthy analytics and AI readiness. A session on multimodal AI in healthcare reinforced the same idea: data cleanliness and metadata are essential because AI can only use information it can interpret reliably. These conference themes underscore an important governance lesson that as AI capabilities advance, the need for data discipline does not diminish. It grows exponentially.

Another key takeaway was that governance is evolving into an enterprise operating model rather than remaining a narrow approval function. Sessions on scalable AI governance, academic health system review, and rural AI strategy all pointed back to the same foundational needs. Those include trusted data environments, clear oversight structures, effective risk triage, and deployment approaches tailored to context. For AHIMA members and credential-holders, this is vitally important. The profession’s longstanding strengths in documentation integrity, information governance, data quality, standardization, and stewardship are becoming more central in the AI era.

What Good Data Quality Looks Like

If data quality is foundational, then healthcare organizations need a practical understanding of what “good” looks like in the context of AI. The following list outlines the basics:

1) Completeness: Accuracy and completeness are tied to the intended use. Completeness should be evaluated in relation to what the AI system is expected to do and what decisions or actions may follow from its outputs.

2) Standards: A standard and consistent data ecosystem is vital. Variation in definitions, metadata, or sources can introduce issues for AI systems. Enterprise semantic alignment and disciplined use of standards remain essential.

3) Provenance and Timeliness: AI is not auditable if the data is not also auditable, which requires elements of timeliness and traceability. Governance requires these aspects along with any transformations documented.

4) Fit: Data must be representative, and context determines fit. Without context and recognizing what purpose data provides, AI systems are prone to failures that are difficult to identify.

5) Stewardship: Data can only be truly governed through clear accountability and responsibility. Organizations need named owners, review processes, escalation pathways, and lifecycle controls for the data assets that support AI.

These are familiar principles to health information (HI) professionals, but there is now an urgency to apply them to AI governance.

Recommendations for HI Professionals

The path forward should begin with recognizing that AI governance is a natural extension of HI governance.

Organizations should start by creating an inventory of AI-related data assets and data flows. That includes identifying the structured and unstructured data sources that feed AI tools, understanding how that data is transformed, and documenting stewardship responsibilities.

They should define fitness-for-use criteria for each use case. It is not enough to say that the data is available. Governance requires asking whether the data is suitable for the intended operational, administrative, or clinical context.

They should incorporate data quality review into AI governance workflows. Procurement, implementation, validation, and post-deployment monitoring should all include explicit review of data quality, provenance, representativeness, and known limitations.

They should involve HI expertise early and visibly. HI professionals are well-positioned to lead in areas such as documentation integrity, source data reliability, metadata stewardship, information lifecycle management, coding and terminology consistency, audit readiness, and transparency documentation.

Finally, organizations should treat workforce development as part of governance. AHIMA provides AI resources and upskilling materials, recognizing that professionals need both AI literacy and governance literacy. Staff need to know not only how to use AI tools, but how to question outputs, understand limitations, and identify when data quality issues may be affecting performance.

AI governance in healthcare is often treated as a question of oversight structures, policies, or regulatory readiness. Those things matter without a doubt. However, the stronger and more foundational starting point is data quality. Poor data quality is like fueling a diesel vehicle with gasoline.

If organizations cannot trust the data that fuels AI systems, then they cannot claim those systems are trustworthy, fair, explainable, or well-governed with any degree of authority. The organizations that invest in strong data governance, documentation integrity, semantic consistency, metadata stewardship, and accountable information practices establish the conditions for AI governance to become real rather than rhetorical.

For HI professionals, this is a significant opportunity. The future of AI governance in healthcare cannot be built by technical teams alone. It will also be built by professionals who understand that trustworthy AI begins with trustworthy information.

Anthony E. Roscoe, MSL, RHIA, FACHDM, is Education Director, Applied AI in Health Information at AHIMA.