Rewiring healthcare data strategy for AI

IT leaders have long focused on building the infrastructure to generate and manage data by building core systems, establishing data quality standards, and synchronizing data across their enterprise. But the priority is shifting now from building a perfectly integrated data environment to making the right data accessible, governed, and usable for the outcomes leaders want to achieve.  

AI changes the sequence of this work. Data doesn’t have to be perfect or housed in one place before organizations can begin creating value. Instead, data can remain in its native locations while generative and agentic AI workflows access, connect, and analyze information across disparate sources.  
 
That’s driving leaders to ask a different set of questions: 

Do I have the data I need to use AI effectively?

Healthcare organizations need to first confirm that they’re collecting the right data to facilitate the experiences and workflows they’re hoping to achieve with AI implementation. The best way to do this? Start with the end result and work backwards. 

Take, for example, a health system that wants to use AI to identify patients with abnormal lab results who haven’t received appropriate follow-up care. The desired end result isn’t simply “find abnormal results.” It’s a care coordination workflow that flags the right patients, creates the right next step, and helps the care team reach them at the right time. Working backward from that outcome, leaders would need to verify that several data elements are available, reliable, and connected, including:

• Clinical protocols: Which abnormal results require follow-up, how quickly that follow-up should happen, and which type of test, appointment, or intervention is needed
• Lab results: Whether abnormal results from internal and external labs have been captured consistently and clearly marked
• Care history: Whether the patient has already completed the recommended follow-up, received treatment, or had a related encounter
• Scheduling data: Whether a follow-up appointment has been scheduled, completed, missed, or canceled
• Outreach history: Whether the patient has already been contacted and through which channel
• Contact preferences: Whether the organization has current phone, portal, language, and communication preference data 

With those pieces in place, AI can do more than surface a potential care gap. It can help prioritize outreach, create tasks for care coordinators, and support a workflow designed around the organization’s actual patients, protocols, and operations. 

Is my data accessible to AI?

Healthcare organizations often operate with fragmented information spread across core systems, departmental applications, data warehouses, cloud environments, and third-party vendor solutions. This fragmentation creates inconsistencies, duplicate records, conflicting reports, and a lack of confidence in decision-making.

Traditionally, organizations have built API integrations that enable data replication to provide visibility. This can create data latency and accuracy issues. Even as integrations have become more advanced and data replication has become less necessary, many organizations are still spending time and resources aligning data standards and building structures to replicate data across applications. 

AI challenges the notion that healthcare organizations need to build separate systems to house and normalize every data set before they can create value. The goal is to create a governed, trusted view of which sources AI should access and use, not to move every data set into one place. Technology leaders should establish governance over which data sources are authoritative, how information is defined, and how data should be accessed, interpreted, and applied across the enterprise. 

An agentic mesh can operationalize this, connecting data across native systems and providing context for AI workflows. This approach allows AI to retrieve information from original sources rather than relying on replicated or outdated copies, improving quality and confidence in the data used. 

Is my data clean enough for AI? Do I have the expertise I need to manage AI-related data?

Many healthcare leaders assume they need to resolve every data quality issue before pursuing AI. John Young, Global Managing Director at Databricks, debunked this myth on a recent episode of The GuideLine podcast. Organizations can create meaningful value with imperfect data, while AI can help identify quality gaps, uncover inconsistencies, and prioritize remediation efforts as initiatives move forward. The goal isn’t perfect data—it's understanding which data issues matter most to the outcome and addressing those first. 

Most organizations are capable of identifying and organizing the majority of the data needed for AI initiatives. The greater challenge lies in the remaining edge cases—the incomplete records, inconsistent data definitions, unique workflows, and exceptions that occur in real-world healthcare environments. While the first 80% of data readiness can be addressed through technology and governance, the final 20% often requires deeper domain expertise, strong stewardship, and careful validation so that AI tools produce reliable results.  

This may require a shift in mindset. Organizations need to take a holistic approach to their AI and data strategy, approach, and management. Guidehouse helped a U.S. federal agency do this by aligning its AI initiatives with the agency’s broader data strategy, developing an AI metadata catalog, and establishing governance processes for managing AI-related data and use cases. This improved visibility, consistency, and oversight of AI activities across the organization. 

Once I know I have the data I need, how do I make sure AI is using it correctly?

Pilot programs inform development of data governance, validation standards, and AI workflow oversight. As organizations test out these workflows, they begin to understand where data gaps lie. 

Pilots offer healthcare organizations a controlled environment to validate their data, test guardrails, and confirm that AI can be used responsibly before scaling more broadly. By defining what agents can access, how responses are sourced, and where human oversight is required, leaders can limit risk, validate performance, and build confidence for wider deployment. An effective pilot confirms that an AI tool is able to achieve its goals and positively impact outcomes without harming patients or creating unnecessary risk. 

But testing with real data can be challenging, and deriving synthetic data that fully accounts for high-impact and high-volume use cases isn’t simple. Rare, high-impact use cases need to be tested, but real-world production data often contains too few examples to validate them reliably. Whether using synthetic data representative of real populations or production data, the goal is a realistic set of use cases in a testing environment that supports predictable agent deployment. 

We recently helped a major biopharmaceutical manufacturer translate medical, legal, regulatory, compliance, business, and engineering requirements into practical controls for an AI system that could answer medication questions and connect patients with support services. As part of this project, we validated 167 complex prompts related to the AI system’s capabilities and tested 43 governance validation scenarios.  

Through that work, we established clear, compliant data-related guardrails, including:  

• Information governance: Limits on the internal and external sources that agents could use to answer questions, paired with controls on how information could be retrieved and used 
• Source prioritization and traceability: Guidance on how agents should prioritize sources and track where answers were sourced from
• Knowledge controls: Boundaries for AI-generated content, with agents prohibited from offering medical advice, treatment recommendations, or clinical interpretation

Healthcare organizations do not need to wait for perfect data or a fully centralized environment to move forward with AI. By starting with a clear outcome, governing access to the right data, and testing workflows against real-world scenarios, leaders can build confidence and scale AI responsibly