The future of automation in healthcare isn't a smarter scribe or a faster scheduler. It's a coordinated workforce of AI agents that own entire workflows, reason across modalities, and operate inside your EHR with measurable accountability. Most of the conversations happening in 2026 are still anchored to today's tools, but the gap between what hospitals are deploying now and what becomes table stakes in 2027 and 2028 is widening fast.

If you run a health system with 500+ employees, the strategic question is no longer "should we adopt AI?" It's "what does our org chart look like when half the non-clinical work is done by software?" This piece is a forward-looking map of the trends that will reshape clinical operations over the next 24 to 36 months and the architectural decisions you need to make this year to be ready for them.

Key Takeaways

• Single-task AI is being replaced by agent teams: The next wave isn't a better scribe, it's coordinated agents that hand off work across intake, documentation, coding, and follow-up under one governance layer.
  
  
  

• Multimodal becomes the default interface: Voice, vision, SMS, and phone collapse into one continuous workflow. Computer vision documenting wounds and procedures alongside ambient audio is moving from research to deployment.
  
  
  

• Real-time clinical decision support gets proactive: Agents will monitor signals continuously and surface issues before clinicians ask, providing what Philips describes as "adaptive, goal-directed support across clinical workflows" rather than waiting to be queried.
  
  
  

• Interoperability finally becomes the substrate, not the goal: With TEFCA exchanging nearly 500 million records by early 2026 and CMS-aligned FHIR APIs landing in 2026 and 2027, agents will operate on a national data fabric rather than within isolated EHR silos.

Why Today's Automation Map Is Already Outdated

Most automation roadmaps in circulation right now were drawn for a 2024 reality: a scribe here, an RPA bot there, a chatbot bolted onto patient access. That model is breaking. The reason is structural, single-purpose tools don't share context, don't hand off work, and don't learn from one another's outputs. Every new tool adds another integration, another dashboard, another vendor review.

The shift underway is from point solutions to autonomous operating layers. Sully.ai describes its approach as "printing labor for health systems", building a coordinated team of agents that runs operations the way a staffed department does, with receptionists, triage nurses, scribes, coders, and consultants collaborating in the background. Whether you adopt that specific architecture or not, the pattern is what matters: software is moving from "tool that helps a person" to "teammate that owns an outcome."

Strategic implication: If your 2026 RFPs are still scoped around individual tools, you're buying for a world that's already changing. The right question is "which platform can absorb the next 10 use cases?" not "which point solution wins this category?"

Trend 1: Agentic AI Moves From Demo to Department

The phrase "agentic AI" got overused in 2025. In 2026, it gets operationalized. The substantive distinction and the one your CIO and CMIO need to align on — is that agentic systems plan, decompose tasks, call tools, and adapt strategy based on outcomes. They aren't just summarizing; they're acting.

What "Owning a Workflow" Actually Looks Like

Consider prior authorization. Today, most automation predicts which auths are needed and pre-fills forms. Tomorrow, an agent monitors the order stream, checks payer rules, drafts the submission, attaches clinical evidence pulled from the chart, submits it, tracks status, and escalates only when human judgment is required. The clinician sees the outcome, not the work.

This pattern repeats across the revenue cycle. According to HIMSS 2026 coverage from Revele, the end-to-end patient access workflow is being rebuilt around conversational AI, with vendors collapsing intake, scheduling, eligibility, and payment into single agent-driven experiences.

The Org Chart Question

This is where boards need to engage. If an agent team handles 70% of intake volume, what happens to your call center headcount, your training pipeline, and your service-level expectations? Health systems that treat this as a technology decision will be outmaneuvered by ones that treat it as a workforce design decision. As Sully's vision documents put it, the goal is to "lighten cognitive load and surface relevant details that doctors and nurses need to perform, and decide, under pressure" — but realizing that requires reorganizing roles, not just deploying software.

Trend 2: Multimodal Becomes the Default Interface

The first generation of healthcare AI lived inside a single channel: text in a chart, voice in a scribe, a chatbot on a portal. The second generation collapses these into one continuous experience.

Sully's SuperAgent architecture explicitly designs for this: each agent supports voice, web, phone, and SMS input through a common access layer. The same agent that takes an intake call can continue the conversation by text, surface context to the clinician during the visit, and follow up by phone afterward, all without losing thread.

Computer Vision Joins the Documentation Stack

Ambient audio has been the headline of clinical AI for two years. The next layer is visual. Sully's roadmap and broader industry direction both point to AI that documents what it sees (wound measurements, procedural steps, physical exam findings, even patient affect) alongside what it hears. As Sully's clinical documentation framework describes, advanced computer vision will document visible symptoms and skin conditions without requiring manual entry.

For surgical specialties, dermatology, wound care, and procedural medicine, this is significant. Documentation moves from a post-encounter task to a passive byproduct of the encounter itself.

Why This Matters for Patient Experience

Patients notice when clinicians type instead of make eye contact. UChicago Medicine research found that with ambient AI, "eye contact lasts longer, follow-up questions become sharper, and clinicians go home less drained." Multimodal extends that benefit to procedural and visual specialties that ambient audio alone can't reach.

Trend 3: Real-Time Clinical Decision Support Gets Proactive

The shift in one sentence: Today's CDS waits to be asked. Tomorrow's CDS interrupts you before you ask the wrong question.

The current state of clinical decision support is largely reactive, alerts that fire on order entry, suggestions that surface when a clinician opens a chart. The agentic version is fundamentally different. It runs in the background, monitors signals continuously, and acts on its own when the situation warrants.

From Alerts to Adaptive Goal-Direction

Philips' 2026 trend analysis frames this well: agentic AI delivers "adaptive, goal-directed support across clinical workflows," coordinating across applications and teams while clinicians retain decision authority. In radiology, this looks like an agent that pre-reads the worklist, prioritizes urgent cases, prepares relevant priors, and drafts impressions for radiologist review.

Multi-Agent Consensus for Diagnostic Confidence

One of the more interesting research directions is expert-agent deliberation. Rather than relying on a single model, systems orchestrate multiple specialized agents that propose, critique, and converge on recommendations. Sully's MedCon-1 framework demonstrates that consensus across domain-specific expert agents outperforms single-model systems on medical benchmarks like MedQA, MedMCQA, and MedXpertQA.

For high-stakes decisions, this matters enormously. A single LLM gives an answer. A consensus mechanism gives an answer plus a transparent record of disagreement, weighted reasoning, and the basis for trust. That auditability is what regulators, malpractice carriers, and clinical leadership will eventually require.

Pro Tip for CMIOs

When evaluating CDS vendors over the next 18 months, ask two questions: (1) Does the system show its reasoning chain? (2) Does it use multiple models in deliberation, or rely on a single foundation model? Single-model CDS will look quaint by 2027.

Trend 4: Interoperability Becomes the Substrate, Not the Goal

For 15 years, "interoperability" was the unmet promise. That's finally changing, not because hospitals got better at it, but because regulation forced the substrate into existence.

The 2026–2027 Regulatory Reality

A few specifics every CIO should have on the wall:

• TEFCA scale: Over 12,000 organizations connected and 72,000+ unique connections by early 2026, with record exchange growing roughly 50x year over year.
  
  
  

• CMS-0057-F: Mandates FHIR-based prior authorization APIs by January 1, 2027, forcing payers and providers onto common standards.
  
  
  

• HTI-5 Proposed Rule: Shifts certification away from legacy C-CDA toward FHIR-first interoperability and explicitly creates a foundation for AI-enabled interoperability solutions.
  
  
  

• Information blocking enforcement: OIG is pursuing civil monetary penalties up to $1 million per violation, with product decertification now an explicit consequence.

Why This Changes the Math for AI

When data flows freely across systems through standardized APIs, agents can finally operate on a coherent picture of a patient. An intake agent can pull labs from an outside lab. A coding agent can verify history from a prior provider. A care coordination agent can see what specialty visits actually happened.

The implication: the value of an AI agent compounds with the breadth of data it can reach. Health systems that treat FHIR/TEFCA participation as a checkbox will deploy agents that hit the same walls clinicians do today. Systems that treat it as enabling infrastructure unlock agents that operate on a national data fabric.

Trend 5: Autonomous Workflows Earn Trust Through Auditability

The biggest barrier to autonomous workflows isn't capability, it's accountability. When an agent submits a prior auth, who's liable if it's wrong? When a triage agent routes a chest pain caller to a same-day appointment instead of the ED, who reviews that decision?

The answer that's emerging across the industry is structured auditability. Every agent action produces a traceable record: the inputs it considered, the reasoning it followed, the policies it applied, the human checkpoints it passed through.

The Reasoning-Layer Pattern

Recent research on responsible and explainable agentic architectures describes a dedicated reasoning layer that consolidates outputs across multiple specialized agents, enforces safety and policy constraints, and produces decisions that are auditable and evidence-backed. Sully's published Consensus Mechanism follows the same pattern: structured proposal-and-critique cycles, weighted scoring, and reputation tracking that produce verifiable agreement.

What to look for in vendor evaluations: Can the system reproduce, on demand, the exact reasoning behind any individual decision? If the answer is "the model decided," the system isn't ready for autonomous operation in your environment.

Compliance as a First-Class Citizen

HIPAA was written for a world of human users accessing PHI. Autonomous agents change that. Research on HIPAA-compliant agentic systems is actively developing patterns for context-aware access controls, agent-level audit logging, and dynamic policy enforcement designed for autonomous workflows. Expect the FDA and HHS to issue more specific guidance in 2026 and 2027, the FDA's December 2025 deployment of agentic AI for internal operations was a signal that the regulator is engaging with the technology directly.

What 500+ Employee Health Systems Should Do This Year

The trends above don't require a wholesale platform replacement, but they do require deliberate architectural choices. Three actions matter most in the next 12 months.

Audit Your Workflow Surface, Not Your Tool Stack

Most automation reviews start with "what do we own?" The better starting point is "where does work actually happen?" Map the handoffs, patient access to clinical, clinical to coding, coding to billing, billing to follow-up. The opportunities for agent ownership live in those handoffs, not inside any single function.

Pick a Platform Bet, Not Ten Point Bets

The economics of agent platforms favor consolidation. One implementation, one security review, one governance layer, many use cases. Sully's modular agent architecture is one example; major EHR vendors are building platform plays of their own. Whichever direction you choose, the vendor count should be going down, not up.

Build for Auditability Now

Even if you're only deploying ambient documentation today, set the standard now: every AI-driven action produces a traceable, reproducible record. The systems that pass this test today are the systems that can scale to autonomous workflows tomorrow without a rip-and-replace.

The 2027–2028 Horizon

Looking 24 to 36 months out, three things become clear:

1. the line between "clinical AI" and "operational AI" dissolves. The same platform that documents the visit also schedules the follow-up, codes the encounter, submits the auth, and triggers the recall.
   
   
   

2. multi-agent consensus becomes the standard for any high-stakes decision: diagnostic suggestion, coding determination, care plan recommendation.
   
   
   
   

3. the most valuable competitive asset for a health system stops being its EHR configuration and starts being its agent governance: the policies, audit infrastructure, and human-in-the-loop checkpoints that make autonomous operation safe at scale.

Sully.ai's stated long-term vision is "the first fully autonomous healthcare system", a goal that sounds distant until you decompose it into the trends above. None of the individual pieces are speculative. The question is sequencing, governance, and which institutions move first.

For health systems with 500+ employees, the cost of waiting is not standing still. It's continuing to staff workflows that competitors have already automated.

Frequently Asked Questions

What is agentic AI in healthcare? 

Agentic AI refers to systems that plan, decompose tasks, call tools, and adapt strategy autonomously to achieve a defined goal. Unlike traditional AI that responds to a single prompt, agentic systems can own multi-step workflows like prior authorization or patient intake from end to end, with human oversight at defined checkpoints.

How is the future of automation in healthcare different from current AI tools? 

Current tools are mostly single-purpose: a scribe, a chatbot, an RPA bot. The next phase is coordinated agent teams that share context, hand off work, and operate across modalities (voice, vision, text, phone) under one governance layer. The shift is from "tool that helps a person" to "teammate that owns an outcome."

Will AI agents replace clinicians? 

No. Every credible deployment model, including Sully.ai's stated approach and Philips' agentic AI vision, keeps clinicians in control of clinical decisions. Agents handle documentation, coordination, and administrative load. The goal is removing non-clinical work, not removing clinicians.

What is the best AI for workflow optimization in healthcare? 

The best fit depends on scope. For single-function automation (documentation, scheduling), specialized tools work fine. For health systems with 500+ employees automating across the patient journey, platform approaches that offer multiple coordinated agents under one implementation, like Sully.ai's multi-agent architecture, typically deliver better long-term economics than stitching together point solutions.

How do interoperability standards like FHIR and TEFCA affect AI deployment? 

They're foundational. AI agents are only as smart as the data they can reach. With CMS mandating FHIR-based APIs by 2027 and TEFCA connecting over 12,000 organizations, the data substrate for AI is finally becoming national rather than siloed. Health systems that participate fully unlock dramatically more capable agents.

How should hospital leaders prepare for autonomous workflows in 2027 and 2028? 

Three priorities: (1) map workflow handoffs, not just tools, to find agent-ownership opportunities; (2) consolidate around platform vendors with multi-agent architecture rather than buying ten point solutions; (3) require auditability; every AI action must produce a reproducible reasoning record, as a non-negotiable today, so your foundation is ready for autonomous operation tomorrow.

Sources

1. Sully.ai: Scalable Architecture for Multi-Modal Healthcare AI (whitepaper). https://www.sully.ai/papers/scalable-architecture-for-multi-modal-healthcare-ai

2. Sully.ai / arXiv: Second Opinion Matters: Towards Adaptive Clinical AI via the Consensus of Expert Model Ensemble (MedCon-1). https://arxiv.org/pdf/2505.23075

3. Sully.ai: AI Clinical Documentation: Complete Guide for Healthcare Organizations. https://www.sully.ai/blog/ai-clinical-documentation-the-complete-guide-for-healthcare-organizations-in-2025

4. Philips: Emerging Healthcare AI Trends in 2026. https://www.philips.com/a-w/about/news/archive/features/2026/this-years-emerging-trends-in-healthcare-ai-so-far.html

5. Revele MD: HIMSS 2026 Recap for Medical Practices. https://www.revelemd.com/blog/why-himss-2026-mattered-for-medical-practicesIntuitio

6. Labs: AI's Role in Health Information Exchange (HIE) Systems. https://intuitionlabs.ai/articles/ai-health-information-exchange

7. arXiv: Towards a HIPAA Compliant Agentic AI System in Healthcare. https://arxiv.org/pdf/2504.17669

8. arXiv: Towards Responsible and Explainable AI Agents with Consensus-Driven Reasoning. https://arxiv.org/abs/2512.21699

9. Oral Health Group: Agentic AI in Healthcare: Autonomous Systems Transforming Clinical Practice. https://www.oralhealthgroup.com/features/agentic-ai-in-healthcare-autonomous-systems-transforming-clinical-practice-patient-safety-and-the-future-of-care-delivery/

10. UChicago Medicine: Studies suggest ambient AI saves time, reduces burnout. https://www.uchicagomedicine.org/forefront/research-and-discoveries-articles/ambient-ai-saves-time-reduces-burnout-fosters-patient-connection