Much of the first wave of AI hype centered on scale: bigger models, more compute, and more training data. Today, attention is expanding beyond ever-larger language models and toward the harder work of making AI useful across workflows.
The shift is happening while AI adoption is still relatively early. In 2026, roughly 17% to 20% of U.S. businesses reported using AI, while 20% to 23% expected to begin using it within the next six months, according to the U.S. Census Bureau’s Business Trends and Outlook Survey.
The core of that usability is building context, connecting data cleanly across disjointed sources, and establishing enough trust for its outputs to inform real-world decisions, especially as AI is applied across more functions and use cases.
In 2027, that will put pressure on the foundations beneath healthcare’s emerging AI operating layer. Organizations that back AI systems with strong data governance, including structured, domain-specific context, will be poised to extend AI into more meaningful, scalable applications.
Context becomes a discipline
Early AI work placed a lot of emphasis on prompt engineering. Now, more of the work is moving toward building context, or the background information a model uses to “understand” what’s happening and generate a relevant, accurate response for a specific application. That context can include the data being queried, the history of an interaction, the rules governing a workflow, and other signals that help the system determine what information matters for a particular request.
“We have enormous amounts of healthcare data across EHRs, claims, pharmacy, and other sources. The harder problem is building the context around that data and making it trustworthy enough to support a particular application or decision.” - Srinivas Attipalli, Head of AI, Definitive Healthcare
Curation is part of the equation
Building that context starts with deciding what information the AI system should have access to in the first place. An enormous data lake does not automatically provide an AI model with better context. In some cases, it can make retrieval harder by giving the system more competing or irrelevant information to sort through. Curation helps address that problem by creating a foundation from which the information most relevant to a particular use case can be assembled.
Connecting data to build context
Data also needs to be connected. Healthcare data remains notoriously scattered across datasets with mismatched structures and definitions. For example, the same physician might appear under slightly different names across clinical records. Entity resolution helps determine when those records refer to the same person, allowing the information to be linked into a more complete profile. This gives an AI model a more comprehensive view of entities and their relationships, allowing it to reason over connected, contextualized information rather than isolated records.
AI raises the stakes for trust
We’ve all seen a large language model (LLM) invent information or misstate a “fact” with total conviction. As AI becomes embedded into more workflows, inaccuracies can travel further and become harder to spot. Scale turns a manageable error into a repeated one.
“Gaps or limitations in the data are no longer necessarily contained to a single report or analysis. When an AI model ingests flawed data to generate recommendations, it can exponentially amplify that information across downstream workflows.” - Srinivas Attipalli, Head of AI, Definitive Healthcare
For AI-generated intelligence to support consequential decisions, organizations need confidence that the underlying information is complete, current, appropriate for the question or task, and supported by systems in which:
- The source is clear: Users can understand where the underlying information came from and what it represents.
- The result can be checked: Outputs can be evaluated against verified information or an established methodology.
- The process is consistent: The same analysis can be performed again using the same inputs and methodology.
- Use is controlled: Organizations can determine who can access the information, how it can be used, and what actions the system can take.
Trust requires built-in governance
Governance provides the guardrails that make this level of trust possible. This is especially important with healthcare data, where state-specific privacy laws overlay federal HIPAA standards, and data sources can carry unique, vendor-specific usage restrictions. Organizations need to understand these constraints and apply guardrails around what information can enter an AI system, how it can be used, and what the system can send or act on. In practice, this means implementing role-based access, automated policy enforcement, immutable audit logs, and strict controls around automated actions, alongside human oversight.
From AI point solutions to AI operating layer
The industry is rapidly hitting an AI point-solution plateau. While speeding up an individual task by 20% certainly has value, adding another standalone AI tool to an already bloated technology stack eventually hits diminishing returns. As these isolated wins dry up, the industry is moving from micro-tasks to macro-systems.
NVIDIA’s State of AI in Healthcare report offers a glimpse of that shift: generative AI and LLMs are the most widely used AI workloads among survey respondents, followed by data analytics and predictive analytics, while agentic AI is gaining traction for multistep tasks within defined workflows.
