Avoid the pitfalls of healthcare AI

Artificial intelligence holds significant promise to improve healthcare efficiency, reduce costs, and enhance patient outcomes. Yet many healthcare AI initiatives will fall short of their declared goals, not because the technology is insufficient, but because it is deployed in ways that do not align with how healthcare organizations operate.

For healthcare leaders, the challenge is no longer whether AI will matter—it’s how to ensure it delivers meaningful, sustainable value. Achieving that goal requires integrating AI into the data, workflows, and governance that underpin care delivery, rather than funding it as a series of isolated investments. When AI is operationalized this way, it can reduce operational friction, improve clinical and business decision-making, and deliver solutions to longstanding problems in healthcare.

Understanding how AI works

Traditional software follows explicit rules written by programmers. By contrast, AI systems are trained on data to recognize patterns and generate predictions, classifications, or content. This means AI is only as good as the data it is trained on and uses to draw conclusions.

Learning from large, multimodal datasets (which can include images, genomic information, and unstructured as well as structured data) allows AI systems to perform tasks that were previously difficult for computers, such as interpreting natural language commands, identifying patterns in medical imaging, and detecting meaningful correlations across large clinical datasets—for example, early signs of disease from subtle changes in vitals and lab results or tumors in radiology images.

There are different types of AI. Predictive AI identifies patterns and forecasts likely outcomes, as in identifies abnormalities in radiology images. Generative AI creates new content such as text, code, or images. In healthcare, generative AI can help clinicians summarize patient records, draft discharge summaries and referral letters, and support research through rapid literature review. A newer form of AI, called agentic AI, goes a step further. It can plan, make decisions, and take self-directed actions within boundaries set by human rules.

Together all these forms of AI can reduce people’s manual, repetitive work, while providing data and information to support human decision-making.

Computing infrastructure is another foundational component of enterprise AI. AI models require very large amounts of computational power to train and operate across large datasets. These workloads are typically handled by massive numbers of high-performance processers deployed in data with the power, cooling, and networking required for AI at scale. Many organizations access this infrastructure through cloud services, allowing them to use advanced computing capabilities without building their own facilities and to pay only for the compute resources they use. This combination of specialized hardware and cloud infrastructure has made enterprise AI more practical to deploy.

A common misconception among healthcare organizations is that success with AI requires building proprietary models. In practice, most organizations will not train foundation models from scratch. Instead, they can deploy and adapt existing AI models trained on healthcare- and life sciences-specific information such as drug and formulary data, clinical research, practice guidelines, epidemiology data, payer rules, and real-world evidence.

Bolt-ons vs. built in

While AI models can be powerful, they need to be fully integrated with data across the healthcare organization and broader ecosystem. Standalone AI models outside core applications are insufficient and can introduce significant risk. For example, standalone AI analytics tools may identify high-risk patients, but often they cannot take action to ensure they receive timely care or follow up.

Bolt-on approaches can solve specific workflow issues or inefficiencies, but they rarely deliver sustained enterprise impact because their insights remain disconnected from the rest of the organization. A model operating in isolation has no inherent awareness of real-time patient data, local clinical workflows, organizational policies, or continuously evolving medical knowledge unless it is explicitly connected to those systems.

As a result, its outputs may be generic, outdated, or misaligned with the specific clinical context in which decisions are being made. Standalone models are also susceptible to “hallucinations,” generating responses that appear plausible but are not factually grounded.

This limitation is especially important in healthcare, where incomplete or incorrect information can cause patient harm. A bolted-on model may generate a well-structured clinical summary that omits a critical lab abnormality, suggest a treatment approach that conflicts with a patient’s allergies or current medications, or provide recommendations that do not reflect the latest clinical guidelines or the organization’s care protocols. Because these outputs are often delivered in authoritative language, users may not immediately detect the problem without careful review.

Embedded AI

More effective than bolting on point AI solutions to existing processes and workflows is to embed AI directly within the enterprise applications where work already occurs. When AI is natively integrated into electronic health records (EHRs), revenue cycle platforms, supply chain systems, and HR systems, it can operate with access to real-time, context-rich data inside established workflows, so the AI model is using both the external data it was trained on as well as internal data from the organization.

Embedded AI can surface insights at the moment decisions are made—during clinical documentation, discharge planning, medication ordering, or claims adjudication—rather than requiring users to leave their workflow or interpret disconnected outputs. Integrated AI agents can also trigger downstream actions, such as scheduling follow-up care, initiating referrals, or updating care plans. Such tight integration reduces friction, improves user adoption, and increases the likelihood that AI will deliver a return on investment.

A clear example is AI-enabled clinical documentation embedded directly in the EHR. When ambient listening, note generation, coding support, and follow-up tasks are built into the existing EHR workflow, the AI solution reduces friction by eliminating extra logins, duplicate data entry, and workflow disruption. This makes adoption more likely because clinicians can use these capabilities inside the system they already know and use every day. A bolted-on solution may offer similar features, but if it requires users to step outside their usual workflow, usage tends to drop, and with that, the likelihood of achieving a meaningful return on the technology investment.

High-impact AI requires a full technology stack

A strong AI strategy requires cloud infrastructure for scale and security, a unified data layer to integrate and standardize information across the organization, and analytics applications/models on top to turn data into insights and actions. Without this full “stack,” AI remains fragmented: different applications generate inconsistent recommendations, data quality varies, policies are harder to enforce, and costs and complexity rise.

In healthcare, this fragmentation is common. An AI capability embedded in a provider’s EHR can use clinical data, but it won’t incorporate payer rules, authorization status, supply constraints, the availability of post-acute care, or operational context unless those sources are integrated. The result is decision-making with blind spots—for example, care plans that ignore coverage limits, discharge planning that doesn’t reflect staffing constraints, or revenue cycle AI that denies claims without sufficient clinical context.

Relevant context often extends beyond traditional clinical systems to include operational data from human resources and supply chain systems—for example, whether medications are in stock, whether durable medical equipment has been delivered to a patient’s home, or whether staffing constraints may delay discharge or follow-up care.

The principle is simple: embedding AI into applications is necessary but not sufficient. To deliver system-wide impact, embedded AI must be supported by a unified data platform that continuously refreshes context at inference, enabling accurate, current, and coordinated intelligence across the enterprise. Absent this integration, the status quo, whereby data isn’t available at the right place and time, will remain fundamentally unchanged.

Preparing enterprise data for AI

Successful AI strategies depend on a broader technological foundation and investment in modern data architectures capable of integrating, structuring, and governing data across the enterprise. This is challenging because healthcare data is inherently fragmented and heterogeneous, spanning structured fields, unstructured clinical notes, images, and genomic information, often stored in systems that use different standards and vocabularies.

A modern data platform helps by taking fragmented healthcare data and turning it into a single, trusted foundation that AI can use reliably. It connects AI models to enterprise information so the model can pull the most relevant facts when a question is asked and base its responses on current, organization-specific context.

A semantic data layer makes the data consistent and comparable. It translates messy, inconsistent inputs into standard terms and formats, for example, treating “myocardial infarction” and “heart attack” as the same condition, and ensuring lab results are interpreted the same way across systems and sites.

On top of that, knowledge graphs add the missing context by showing how things relate over time. They connect patients, diagnoses, treatments, providers, encounters, and resources so AI can “see” the real-world relationships behind the data.

For example, instead of seeing a diagnosis code, an abnormal lab value, a medication order, and a hospital admission as separate events, AI can understand they all relate to the same patient, the same episode of care, and the same underlying clinical context. That makes it far more likely the AI will draw the right connections, avoid false assumptions, and generate insights grounded in how patients, conditions, treatments, and outcomes actually relate to each other.

Together, the semantic layer and knowledge graph provide the clarity and context AI needs to interpret healthcare data accurately, both when building models and when using them in day-to-day operations.

Interoperability: the missing link between data and impact

Even the strongest data platform and AI strategy will underperform without interoperability—the ability for systems to exchange data reliably, securely, and in usable form across the care continuum. In most healthcare environments, patient and operational information is spread across multiple EHRs, labs, imaging systems, payer platforms, post-acute providers, and public health systems. When these systems cannot share data effectively, organizations are left with partial, delayed, or conflicting views of the patient and the operation, which limits care coordination and reduces the accuracy and usefulness of AI-supported insights.

Interoperability is also a governance and operating-model issue. Organizations need clear policies for data access, patient matching/identity resolution, provenance (where data came from), consent and privacy controls, and auditing, especially when AI is involved.

Equally important is addressing non-technical barriers, including contractual constraints, inconsistent workflows, and “data blocking” behaviors that can limit exchange even when standards exist. Leaders should treat interoperability as a strategic capability, not an integration afterthought, because it directly determines whether AI can operate on a complete, trusted, and current picture of patient and system state.

Older technology is not built for AI

Preparing data for AI is hard because the legacy core IT systems of healthcare organizations, such as MUMPS databases, were designed for transactional processing, not for the real-time data sharing and interpretation critical for effective use of AI. As a result, applications, even those with embedded AI, often can’t pull in and harmonize data from other systems quickly enough, because the needed integration, data standardization, and real-time pipelines aren’t in place.

Without modernization, organizations are often forced to maintain hybrid architectures in which legacy systems function as systems of record while separate platforms serve as systems of intelligence. This dual architecture introduces complexity, cost, and latency, and it makes it harder to fully embed AI into workflows.

By contrast, modern, AI enabled data platforms integrate structured and unstructured data, support semantic indexing, and enable retrieval augmented generation (RAG) by retrieving relevant information in real time and using it to ground model responses in enterprise data within a unified environment. This reduces expenses as well as architectural complexity, and it enables deeper integration of AI into enterprise systems.

Governance, validation, and trust

Successfully deploying AI in healthcare requires more than technical capability; it also demands robust governance, validation, and ongoing oversight. Trustworthy AI depends on human oversight, with qualified professionals retaining responsibility for final decisions and validating AI-generated recommendations before acting on them. Organizations must clearly define accountability and put safeguards in place when outputs are uncertain or conflicting.

Transparency is also essential. Clinicians and administrators must be able to understand the purpose of models, as well as their data inputs, limitations, and appropriate uses. This aligns with emerging regulatory expectations, including FDA guidance on software as a medical device and ONC requirements for decision-support transparency within certified health IT.

Healthcare organizations must evaluate AI systems across their full lifecycle, including initial validation, deployment, and continuous monitoring. That includes assessing performance across different patient populations, monitoring for model drift over time, and helping ensure that outputs remain clinically appropriate as workflows evolve.

The path forward

If healthcare AI fails to achieve the systemwide transformation many predict, it will not be because the technology is incapable, but rather because organizations underestimate what it takes to operationalize this powerful capability effectively. Success depends on moving beyond fragmented, bolt-on solutions toward deeply embedded, enterprise-wide intelligence supported by unified data, modern architecture, and rigorous governance. AI must be integrated into real clinical and operational workflows, grounded in high-quality, context-rich data, and aligned with the realities of healthcare delivery.

For healthcare leaders, the path forward is clear but demanding. It requires investing not only in AI models, but also in the data platforms, interoperability, and governance frameworks that allow those models to function safely and effectively at scale. It also requires maintaining human oversight, ensuring transparency, and building trust across clinicians and administrators. Organizations that treat AI as a strategic capability, rather than as a standalone tool, will be best positioned to unlock its full potential to improve health outcomes, provider efficiency, and the patient experience.

Seema Verma is executive vice president and general manager of Oracle Health and Life Sciences.

• A Strategic Blueprint for Digital Health Innovation

  Digital health solutions that retrofit AI onto legacy architectures have proved insufficient. Learn about Oracle’s AI-first approach in this IDC analysis.

  Download the white paper

• Oracle Health and Life Sciences Virtual Events

  Hear from subject matter experts and Oracle customers, see live demonstrations, and participate in Q&As.

  Register for an event