Blab AI
Healthcare AI Adoption Shifts Focus From Models to Foundations, Experts Say
Health‑care leaders, once dazzled by generative‑AI prototypes, are now turning their attention from selecting the right model to building the foundations that will support it at scale.
The sector has long wrestled with aging infrastructure, cybersecurity threats, workforce shortages and the need to prove a return on technology spend. While vendors continue to showcase increasingly powerful AI models, many providers find that the real bottlenecks lie in data, governance and interoperability rather than in the models themselves.
According to Lukasz Lazewski, CEO of health‑IT consultancy LLInformatics, most AI pilots fail because the organization is not ready. “Most AI pilots in healthcare do not fail because of the wrong model or a flawed implementation – they fail because the organization simply is not ready,” he said.
Data fragmentation is a core problem. Clinical, operational and financial information remains scattered across electronic health records, departmental systems and third‑party applications. “The first is data. Health systems store data across multiple systems with no proper aggregation, cleanup or normalization,” Lazewski explained. Legacy environments that have accumulated customizations over decades can support transactional workflows but create obstacles for modern AI architectures.
Beyond technology, AI requires alignment across clinical, operational and administrative stakeholders. Pilots often succeed within a single department but stall when enterprise‑wide adoption demands cooperation. Governance is shifting from a technical discussion to a strategic one. “The questions that should be answered before a pilot launches, including who approves the model, who monitors it, who is accountable when it produces a wrong output, and how that is tracked and corrected, are typically only raised once problems emerge,” Lazewski said.
Lifecycle management is another overlooked area. Software has a life cycle – it is built, it matures, it ages, and eventually it sunsets. “Software has a life cycle. It is built, it matures, it ages, and eventually it sunsets, and at every stage of that life cycle, you need people in place to look after it and ensure business continuity,” Lazewski said. Monitoring, traceability and accountability must be built into AI programs from the beginning rather than retrofitted later.
The pace of AI innovation also imposes architectural implications. Models that were cutting‑edge two years ago are already eclipsed. “The single most important quality of a future‑proof AI architecture is how plug‑and‑play it is,” Lazewski said. He advocates modular architectures that separate data, integrations, AI services and governance functions. “In concrete terms, this means keeping a clear separation between your data layer, your integration layer that covers both internal and external tooling, your AI model layer, and your governance and monitoring layer,” he added.
Technical debt is a hidden cost that can derail AI initiatives. Years of deferred modernization, undocumented integrations and legacy code create integration delays, data remediation costs and security gaps. “Technical debt is probably the single biggest hidden cost in healthcare AI,” Lazewski said. “If a regulatory body asks you to demonstrate how your AI reached a particular decision, and your underlying architecture is built on top of technical debt, you may find yourself unable to answer. That is not a compliance risk. That is a compliance failure.”
Lazewski concludes that successful AI programs rest on three pillars: trusted data, flexible architecture and strong governance. “Every successful healthcare AI program I have seen rests on three things: strong data that is high‑quality, accessible and well‑governed; strong architecture that is modular, secure and interoperable; and strong governance,” he said. For CIOs under pressure to demonstrate value, the lesson is clear: the model is easy to acquire; the organization must first prepare the foundations that will allow it to scale safely.
The sector has long wrestled with aging infrastructure, cybersecurity threats, workforce shortages and the need to prove a return on technology spend. While vendors continue to showcase increasingly powerful AI models, many providers find that the real bottlenecks lie in data, governance and interoperability rather than in the models themselves.
According to Lukasz Lazewski, CEO of health‑IT consultancy LLInformatics, most AI pilots fail because the organization is not ready. “Most AI pilots in healthcare do not fail because of the wrong model or a flawed implementation – they fail because the organization simply is not ready,” he said.
Data fragmentation is a core problem. Clinical, operational and financial information remains scattered across electronic health records, departmental systems and third‑party applications. “The first is data. Health systems store data across multiple systems with no proper aggregation, cleanup or normalization,” Lazewski explained. Legacy environments that have accumulated customizations over decades can support transactional workflows but create obstacles for modern AI architectures.
Beyond technology, AI requires alignment across clinical, operational and administrative stakeholders. Pilots often succeed within a single department but stall when enterprise‑wide adoption demands cooperation. Governance is shifting from a technical discussion to a strategic one. “The questions that should be answered before a pilot launches, including who approves the model, who monitors it, who is accountable when it produces a wrong output, and how that is tracked and corrected, are typically only raised once problems emerge,” Lazewski said.
Lifecycle management is another overlooked area. Software has a life cycle – it is built, it matures, it ages, and eventually it sunsets. “Software has a life cycle. It is built, it matures, it ages, and eventually it sunsets, and at every stage of that life cycle, you need people in place to look after it and ensure business continuity,” Lazewski said. Monitoring, traceability and accountability must be built into AI programs from the beginning rather than retrofitted later.
The pace of AI innovation also imposes architectural implications. Models that were cutting‑edge two years ago are already eclipsed. “The single most important quality of a future‑proof AI architecture is how plug‑and‑play it is,” Lazewski said. He advocates modular architectures that separate data, integrations, AI services and governance functions. “In concrete terms, this means keeping a clear separation between your data layer, your integration layer that covers both internal and external tooling, your AI model layer, and your governance and monitoring layer,” he added.
Technical debt is a hidden cost that can derail AI initiatives. Years of deferred modernization, undocumented integrations and legacy code create integration delays, data remediation costs and security gaps. “Technical debt is probably the single biggest hidden cost in healthcare AI,” Lazewski said. “If a regulatory body asks you to demonstrate how your AI reached a particular decision, and your underlying architecture is built on top of technical debt, you may find yourself unable to answer. That is not a compliance risk. That is a compliance failure.”
Lazewski concludes that successful AI programs rest on three pillars: trusted data, flexible architecture and strong governance. “Every successful healthcare AI program I have seen rests on three things: strong data that is high‑quality, accessible and well‑governed; strong architecture that is modular, secure and interoperable; and strong governance,” he said. For CIOs under pressure to demonstrate value, the lesson is clear: the model is easy to acquire; the organization must first prepare the foundations that will allow it to scale safely.
