Healthcare AI in 2026: Collaboration, cleaner documentation, and smaller models
Healthcare's next move with AI is about practical collaboration. David Lareau, president and CEO of Medicomp Systems, expects three shifts to define 2026: broader use of the Model Context Protocol (MCP), a push for cleaner documentation with real validation, and a pivot to smaller, domain-specific models that are cheaper and safer to run at scale.
MCP connects AI tools without heavy EHR surgery
The Model Context Protocol gives large language models and agent-based apps a standard way to connect with trusted data and services. That means teams can build targeted tools-like clinical prompts, summarizers, or audit helpers-without deep, brittle EHR integrations.
Early momentum is showing up in voice-driven workflows. Ambient listening and voice commands can plug into enterprise platforms through defined APIs that follow MCP, reducing custom code and speeding up deployment across sites. For more on the standard, see the Model Context Protocol.
Documentation quality moves from "nice to have" to "must have"
With Medicare Advantage growth and tighter audits, payers and providers will need stronger proof that documented diagnoses reflect conditions that are present and being managed. If the evidence isn't there, the financial and compliance risks stack up fast.
Expect tools that review encounters in real time, confirm clinical evidence for diagnoses, and flag gaps before data ever leaves the building. Payers will use the same capabilities to validate accuracy before submitting to Medicare.
HCC specificity raises the bar for coding and care plans
The latest CMS Hierarchical Condition Category model calls for greater specificity and clearer clinical support. As the transition completes in 2026, organizations will depend on tech that surfaces gaps and confirms clinical accuracy across populations.
Done right, this improves reimbursement integrity and patient outcomes-every documented condition should match the clinical reality and tie to an active care plan. Learn more from CMS on risk adjustment: CMS Risk Adjustment.
Smaller, domain-specific models win on cost and control
Enterprise-scale LLMs pile up compute and token costs quickly when thousands of clinicians use them daily. Security and reliability also get harder as usage grows.
Smaller models focused on clinical domains are more practical. Many can run on standard CPUs inside a health system's protected environment-keeping PHI in-house, controlling spend, and still delivering high accuracy for specific use cases.
What to do now
- Pick two workflows for MCP-enabled agents (e.g., CDI queries, prior auth, discharge summaries). Define inputs/outputs and how evidence is logged.
- Pilot ambient/voice tools via API, not one-off integrations. Measure time saved, note quality, denial rates, and user satisfaction.
- Stand up a documentation validation program: real-time prompts in the encounter, post-visit reviews, and periodic sampling with feedback to clinicians.
- Tighten HCC specificity: update templates, problem list hygiene, and order sets; publish gap lists monthly with clinician sign-off.
- Run a cost model: compare GPU-heavy LLM usage to CPU-friendly small models. Set token budgets and rate limits for any LLM in production.
- Procurement guardrails: require MCP support, evidence trails for every diagnosis, and clear audit logs accessible to coding and compliance.
- Governance: define data sources of truth, model monitoring, fallbacks when AI is uncertain, and a process to retire underperforming use cases.
Bottom line
The path to real clinical value is getting clearer: connect specialized AI tools through MCP, validate documentation with evidence, and shift to smaller models where scale and security matter. As Lareau puts it, no single vendor can cover every need-so build an ecosystem that plays well together and proves its work at the point of care.
Helpful resources
- Model Context Protocol (MCP)
- CMS Risk Adjustment and HCC
- AI upskilling by job role (Complete AI Training)
Your membership also unlocks:
