By Laksh Krishnamurthy, CTO
Healthcare AI has reached an inflection point—though not the one most people are celebrating.
Investment in the space hit $18 billion in 2025, accounting for nearly half of all healthcare investment. By every financial measure, this is a boom.
But capital flowing in and value flowing out are different things entirely. And right now, we have a troubling execution gap that threatens to turn all that promise into write-offs.
Over the past several years, our industry has proven that AI can extract information, summarize clinical records, classify documents, and automate discrete tasks. Those capabilities are no longer novel. What remains unresolved is far more consequential: whether healthcare AI systems can withstand real-world complexity, regulatory scrutiny, and sustained production scale.
Most healthcare AI initiatives don’t fail because the models are weak. They fail because the systems surrounding those models were never designed to operate in regulated, high-volume environments.
As healthcare organizations look toward 2026, the question is no longer whether AI works. The question is whether we’re building it to endure.
The Pilot Graveyard
Across payers, PBMs, and healthcare enterprises, I keep seeing the same pattern.
Pilots show promise. Demos perform well. Production deployments stall, fragment, or quietly retreat.
Recent, widely-shared research from MIT suggests that 95% of enterprise generative AI pilots fail to deliver measurable business value. Here’s what makes that particularly frustrating: among the minority that do reach production, 74% of healthcare executives report positive ROI on at least one use case.
The technology works. The execution doesn’t.
This breakdown is often attributed to data quality challenges or organizational readiness. Those factors matter, but they obscure a deeper architectural mismatch. Healthcare workflows aren’t linear, static, or standardized. They’re shaped by policy variation, clinical nuance, evolving regulation, and human judgment. Systems designed for experimentation buckle when exposed to these forces at scale.
The stakes extend beyond sunk investment. According to the Chief Healthcare Executive, administrative burden currently consumes 28 hours per week of clinician time—nearly half their workweek. Physicians spend 49% of their day in the EHR and only 27% actually caring for patients. Every failed AI initiative perpetuates this burden rather than relieving it, and worse, it erodes trust in future automation attempts.
AI that succeeds in healthcare must be designed as infrastructure, not as a collection of experiments.
Domain Knowledge Isn’t Optional
In regulated healthcare workflows, general intelligence is insufficient.
Utilization management, pharmacy authorization, appeals, correspondence, care management—these operate within dense layers of clinical policy, contractual rules, and regulatory oversight. The constraints vary by payer, by line of business, often by geography. They evolve continuously.
When domain expertise is treated as an afterthought, AI systems become brittle. They require constant rework, generate inconsistent outputs, and introduce operational risk.
Research from MIT NANDA backs this up: AI projects implemented through specialized vendor partnerships succeed at 67%; roughly twice the rate of internal builds at 33%. In healthcare, the reasons are structural. Most health systems operate across multiple EHRs, legacy databases, and payer portals—environments where building an internal AI solution that integrates seamlessly is prohibitively complex.
Vendors with healthcare-specific expertise arrive with connectors, compliance frameworks, and workflows already in place. They also bring battle-tested approaches to HIPAA, HITECH, and state-level data protection—a compliance minefield that internal teams often underestimate until they’re deep into a build. And perhaps most critically, they compress time-to-value. Internal builds can take years and frequently stall before reaching production. Proven vendors can deploy pilots in weeks and scale once ROI is demonstrated. For organizations facing urgent cost pressures, that speed matters as much as the technology itself.
Durable systems require domain knowledge baked into the architecture itself. In healthcare, domain understanding isn’t a feature. It’s the boundary condition for success.
Explainability as Engineering, Not Afterthought
In payer operations and administrative decision workflows, explainability is no less critical than in clinical settings. Prior authorization, utilization management, and appeals decisions are routinely reviewed by regulators, auditors, and external stakeholders. Every determination must be traceable to policy, evidence, and process logic—not just statistically probable.
With CMS-0057-F as just one example of interoperability and prior authorization reforms, healthcare organizations are being forced to operationalize transparency, consistency, and accountability at scale. As CMS, NCQA, and state regulators increase scrutiny on prior authorization timeliness, consistency, and fairness, healthcare organizations are being held accountable not only for outcomes, but for how decisions are made. Systems that cannot surface decision lineage, policy references, and model behavior under review introduce operational and regulatory risk.
In this environment, explainability is not a feature layered onto models. It is an architectural requirement that spans agents, workflows, data sources, and evaluation frameworks.
Platforms that treat governance as an overlay inevitably struggle. By the time oversight is bolted on, behaviors are already opaque, accountability is fragmented, and trust has eroded. Systems designed with evaluation frameworks, traceability, and human oversight from the start can scale safely.
The Gap Between Demonstration and Operation
There’s a meaningful distinction between AI systems that demonstrate capability and those that operate reliably. Most organizations underestimate how wide that gap actually is.
Model drift—encompassing data drift, concept drift, and label drift—can silently degrade performance as patient populations evolve, clinical protocols change, and disease patterns shift. What worked in the lab doesn’t automatically work in the real-world operations.
This is why 84% of healthcare CFOs are investing as much or more in AI in 2025. Organizations need statistical distribution tests, prediction distribution tracking, and scheduled drift reviews with retraining governance.
Production-grade healthcare AI requires more than model performance. It demands observability across workflows, versioned and reproducible agent behavior, controlled deployment pipelines, and measurable operational outcomes. Without these foundations, organizations remain trapped in cycles of revalidation and reimplementation.
As systems theorist John Gall observed: “A complex system that works is invariably found to have evolved from a simple system that worked.”
Healthcare AI platforms that attempt to leap directly to sophistication without stable foundations rarely endure. Reliability precedes scale. Durability precedes complexity.
The SaaS Imperative: Why Platform Thinking Matters
Healthcare enterprises don’t need dozens of disconnected AI tools. They need a repeatable operating model.
Every point solution carries its own integration tax—connecting to EHRs, analyzing legacy systems, achieving compliance certifications. Without a common platform foundation, organizations pay these taxes repeatedly, creating technical debt that compounds faster than capability scales.
This is exactly why health enterprises are moving toward SaaS-based AI platforms. The economics are straightforward:
- CapEx-to-OpEx Efficiency: Eliminate heavy upfront infrastructure procurement in favor of a predictable operational model; remove the overhead of unbudgeted maintenance and patch cycles.
- Accelerated Deployment Cycles: Transition from multi-year waterfall builds to week-level production releases, utilizing a modular architecture that supports iterative updates without core refactoring.
- Elastic Resource Allocation: Automatically scale to meet volatile demand spikes, eliminating the need for manual, reactive resource provisioning.
- Extensible Interoperability: Ensure seamless data exchange via standard APIs to integrate with legacy stacks, third-party vendor ecosystems, and autonomous agents.
That last point deserves emphasis. The smartest health plans aren’t looking for a single vendor to do everything. They’re building ecosystems where AI platforms coexist with their claims systems, their care management tools, and their existing vendors and solutions. The value compounds when data and intelligence flow across these boundaries.
Think about what this means practically. A SaaS platform with over 100+ healthcare-native agents doesn’t just give you pre-built workflows. It gives you an entry point into an ecosystem, one where your teams can create new AI Agents and workflows, test them in sandboxes, publish them to your internal marketplace, and iterate based on real operational feedback. That’s a fundamentally different model than buying point solutions and hoping they integrate.
Platforms provide shared infrastructure that enables reuse, governance, and consistency across workflows. When intelligence is built on a common foundation, improvements compound. When it’s built engagement by engagement, complexity multiplies.
Repeatability is what allows organizations to move beyond isolated wins toward enterprise-level transformation. Without it, even successful deployments remain fragile and expensive to sustain.
2026: From Extraction to Reasoning
The next phase of healthcare AI won’t be defined by marginal improvements in extraction or summarization. It will be defined by systems that can reason across clinical context, policy, and operational constraints.
In-production results validate this architectural shift. A single Autonomize prior authorization workflow, in-product with a Fortune 150 health plan, achieved a 49% faster turnaround time, shaving off an average of 18 minutes per request within a large volume environment. All while carrying a 76% auto-intake case rate with 98%+ accuracy in clinical data abstraction. McKinsey estimates that 50 to 75 percent of manual prior authorization steps are automatable—significant value given that Medicare Advantage plans issued nearly 50 million prior authorizations in 2023, up 40% since 2020.
Multi-agent architectures, governed orchestration, and application-level intelligence will become the primary units of value. These systems will support human decision-making rather than attempting to replace it.
Here’s the key insight for software developers and architects. Building agentic systems isn’t enough. You need to think about how those systems solve real challenges for knowledge workers. The nurse reviewing prior authorizations. The analyst processing appeals. The care manager coordinating transitions. They don’t need smarter chatbots. They need AI that understands their workflows, their policies, their edge cases.
Agents and applications, not standalone models, will shape how work gets done.
Architecting for Reality: Legacy Systems and the AI Mindset
Healthcare AI platforms must meet enterprise SaaS expectations while respecting regulatory realities.
Beyond SaaS mechanics, healthcare enterprises are being forced to confront a deeper architectural reality: data gravity. Privacy, regulatory exposure, and operational risk mean that most enterprise data will continue to live within in-house systems of record. At the same time, AI-driven workflows and agentic systems require curated, aggregated data to sit close to systems of intelligence to operate with speed and reliability.
This is pushing organizations toward hybrid models by necessity, not preference.
Systems that assume wholesale data centralization will struggle. Systems designed to operate across controlled boundaries—where sensitive data remains protected, but intelligence can be executed where it creates business value, will define the next generation of healthcare platforms.
What does this mean practically? Your AI layer needs to sit alongside your legacy claims system, your enterprise CRM, your existing care management platform. It needs FHIR R4 and HL7 v2 connectors that actually work. It needs to integrate with EHRs through APIs, MCPs, and agent-to-agent interoperability. It needs to respect your security boundaries while still delivering real-time intelligence.
The operational burden these systems need to address is real. AMA surveys show that 94% of physicians report that prior authorization delays patient care, and according to AJMC, 78% abandon treatment entirely. Electronic prior authorization could save hundreds of millions of dollars annually, but adoption remains limited because implementations lack what healthcare enterprises actually require: secure multi-tenant architectures, predictable performance and cost behavior, clear operational ownership, and compliance-ready audit trails.
There are no shortcuts here. Platforms that aren’t built for regulated scale from the outset struggle to adapt later.
Beyond LLMs: Where Agentic AI Creates Real Value
Let me be direct about where I think the opportunity lies.
LLMs are impressive. They’ve unlocked capabilities we couldn’t have imagined five years ago. But for healthcare operations, the LLM is just the foundation. The real value comes from what you build on top of it.
Look for opportunities where agentic AI can drive and speed up healthcare work processes in a more autonomous manner:
- Intake and triage: Agents that classify, route, and prioritize cases before a human ever sees them
- Evidence retrieval: Pulling relevant clinical documentation, policy language, and guidelines automatically
- Case preparation: Summarizing charts, identifying gaps, flagging inconsistencies
- Decision support: Presenting reviewers with everything they need to make faster, more accurate determinations
- Documentation generation: Creating compliant, policy-aligned narratives that clinicians can review and approve
The pattern here is important. You’re not replacing human judgment. You’re eliminating the 70% of the work that’s just gathering, organizing, and formatting information. You’re giving knowledge workers their time back so they can focus on the decisions that actually require expertise.
This is how you keep members happy with faster turnaround times. This is how you keep providers happy with less administrative friction. This is how you keep your operations team happy with manageable workloads.
And critically, this is how you stay compliant. Every output is linked to evidence. Every decision is traceable to policy. Every workflow is governed through AgentOps with full audit trails.
The Road Ahead
The window for getting architecture right is narrowing.
Healthcare is now adopting AI more than twice as fast as the broader economy. But velocity without foundation creates fragility at scale.
Healthcare doesn’t need more AI promises. It needs systems that endure scrutiny, adapt to regulation, and earn trust over time.
The organizations that succeed in the coming decade won’t be those that adopt the most tools. They’ll be those that invest in durable foundations for governed, scalable intelligence.
