This panel discussion at Google Cloud Next features leaders from HCA Healthcare, Independence Blue Cross, Davita, and Elevance Health discussing their journeys from pilot projects to production-scale deployment of AI agents across healthcare operations. The organizations address common challenges including pilot purgatory, fragmented use cases, and change management while implementing AI solutions across clinical workflows, revenue cycle management, patient engagement, and administrative tasks. Key success factors identified include domain-based architectures, tight workflow integration, comprehensive governance frameworks, employee upskilling, and moving beyond pure financial ROI to measure patient outcomes and clinician productivity improvements.
This case study captures insights from four major healthcare organizations—HCA Healthcare (operating 190+ hospitals), Independence Blue Cross (serving 6 million members), Davita (global kidney care operating in 14 countries), and Elevance Health (formerly Anthem)—as they transition AI agent deployments from pilot phases to production scale. The discussion reveals common patterns in healthcare LLMOps challenges and solutions, with all organizations leveraging Google Cloud Platform and Gemini capabilities as their core AI infrastructure.
The session addresses a critical inflection point in healthcare AI adoption where organizations must move beyond proof-of-concept demonstrations to enterprise-wide deployment while navigating regulatory requirements, clinical safety standards, and complex multi-stakeholder workflows. The organizations collectively identify five key trends driving agentic transformation: personal agents for every employee, agents embedded in every workflow, patient-facing agent interfaces, agentic cybersecurity defense, and comprehensive AI literacy programs.
Independence Blue Cross articulates three paradigm shifts forcing healthcare organizations beyond pilot purgatory. First, regulatory compliance requirements have tightened dramatically, with penalties for non-compliance creating existential business risk that eliminates the option of minimal compliance strategies. Second, consumer expectations for real-time insights and actionable recommendations have evolved beyond what batch processing architectures can deliver. Third, cost pressures combined with mandates like prior authorization acceleration and access-to-care requirements create a “do more with less” imperative that makes AI adoption an operational necessity rather than an innovation experiment.
These pressures create what the organizations describe as “a sense of urgency” that fundamentally changes the calculus around AI investment. The window for experimental pilots has closed, with 2025 characterized as a year of experimentation and 2026-2027 positioned as years of execution and production deployment at scale.
HCA Healthcare describes a fundamental shift from fragmented use-case thinking to domain-based architecture as their key enabler for scaling. Rather than pursuing isolated AI applications, they organize around core operational domains including clinical AI, revenue cycle and supply chain, and nursing-specific workflows. Each domain maintains cross-functional teams that include process engineers working directly with frontline staff, along with dedicated change management teams focused on AI literacy and user adoption.
This domain structure allows for value tracking consistency across the organization. Domain leaders establish quarterly OKRs that map to specific value propositions, with individual projects defining key results tied to operational metrics. A dedicated value tracking team monitors progress quarterly and coordinates with finance to measure impact. Critically, HCA emphasizes that at their scale (60,000 nurse handoffs daily across 190 hospitals), success depends less on innovation novelty and more on execution excellence and ease of use.
The domain approach directly addresses the workflow integration challenge. Healthcare workers face significant cognitive burden and physical exhaustion, making them resistant to any tools that add friction to existing workflows. By organizing around domains rather than point solutions, HCA ensures that AI capabilities live within the natural flow of work rather than requiring users to context-switch to separate systems.
Elevance Health takes a complementary approach focused on building reusable capability frameworks that embed governance, explainability, observability, and fairness controls from the ground up. They identify common intelligence patterns across healthcare operations including document intelligence (for processing the massive volumes of unstructured healthcare documents), workflow intelligence (for long-running multi-step processes), data intelligence (for structured and unstructured data integration), voice intelligence (for conversational interfaces), and SDLC intelligence (for software development productivity).
Each framework comes with built-in governance mechanisms, significantly reducing the time required for compliance reviews. Prior to implementing this approach, Elevance required governance meetings with 70+ people for each AI initiative. The reusable framework strategy allows them to clearly delineate areas where faster deployment is appropriate (administrative tasks, non-clinical workflows) versus areas requiring more rigorous review (clinical decision support, patient-facing applications).
This approach directly addresses the challenge of avoiding governance silos. Without unified frameworks, organizations risk creating thousands of ungoverned agents scattered across departments, each with different security postures, audit trails, and compliance controls. The framework approach provides what the panel describes as a “system behind the system”—a single pane of glass for creating, monitoring, securing, and operating agents at scale.
HCA Healthcare highlights ambient documentation as a significant production success case. The system silently captures doctor-patient interactions and automatically generates clinical notes for physician review and submission. This represents a mature deployment of multi-modal AI combining voice recognition, natural language understanding, and structured data generation tightly integrated into clinical workflows.
The organization views this as a foundation for expanded multi-modal capabilities. Their vision for 2027 includes integrating computer vision alongside voice to transform hospital operations more broadly, reducing cognitive burden on healthcare staff while improving care delivery. The key success factor has been making the technology invisible within the workflow—physicians don’t need to learn new systems or change their interaction patterns with patients.
The nurse handoff use case demonstrates the scale impact of workflow-embedded AI at HCA Healthcare. With 60,000 handoffs occurring daily across their hospital network, even a conservative five-minute time savings per handoff translates to 300,000 minutes of nursing time returned to direct patient care rather than administrative documentation. HCA emphasizes that this metric represents just the beginning of value measurement rather than the complete ROI picture.
The deployment illustrates the importance of domain expertise in LLMOps success. Understanding the specific workflow patterns, information requirements, and handoff protocols in nursing requires deep process knowledge that informs model design, prompt engineering, and user interface decisions. The cross-functional domain teams ensure this process knowledge remains central to technical implementation decisions.
Multiple organizations highlight revenue cycle management, prior authorization, and eligibility verification as areas where agentic AI deployment can move faster due to more deterministic outcomes and clearer success metrics. These long-running, multi-step workflows with complex dependencies represent ideal use cases for agent orchestration.
Independence Blue Cross specifically calls out customer service as their most successful domain for demonstrating direct ROI, where automation of routine inquiries through voice agents produces clear cost avoidance compared to human agent costs. However, they emphasize looking beyond this surface-level ROI to downstream value creation. Their example involves a customer service interaction that begins with a billing inquiry but surfaces care gap information (an overdue mammogram), leading to preventative care completion. This demonstrates how connecting discrete capabilities (billing system access, clinical data access, care gap algorithms, member communication) creates value beyond the initial use case scope.
While patient-facing AI deployment moves more cautiously due to regulatory considerations and clinical risk, organizations envision agents serving as always-on concierges for navigation tasks including appointment booking, insurance coverage verification, co-pay inquiries, and care recommendations. The strategic vision distinguishes between administrative “chore” tasks that AI can handle 24/7 versus high-touch clinical interactions where human judgment remains essential.
Davita articulates this as agents helping patients “advocate for themselves” and navigate complex healthcare systems during vulnerable moments. The key design principle involves creating space for meaningful human touchpoints by offloading routine navigation tasks to AI interfaces. Organizations acknowledge that regulatory caution is appropriate for patient-facing applications but emphasize that 44% of healthcare stakeholders are actively exploring this domain.
The panel strongly emphasizes that AI strategy requires robust data strategy as a prerequisite. Healthcare organizations contend with significant legacy data distributed across multiple systems, requiring intentional prioritization about which data domains to tackle first. Davita describes organizing their data strategy around intent—focusing first on patient data versus clinician data versus operational data based on strategic priorities.
The distinction between structured and unstructured data emerges as particularly important in healthcare. Massive volumes of information reside in documents, emails, faxes, and images. Elevance describes building intelligence platforms that govern both structured and unstructured data, making them accessible and usable across the enterprise while maintaining security controls. This unified data governance approach enables the reusable framework strategy described earlier.
The organizations uniformly leverage Google Cloud Platform as their data and AI stack, with specific mentions of GCP infrastructure, Gemini models, and Google’s document intelligence capabilities. The choice of a unified cloud platform supports the goal of avoiding fragmented agent deployments with inconsistent governance.
A significant theme across all organizations involves evolving ROI measurement beyond pure financial returns to encompass patient outcomes, clinician satisfaction, and employee productivity. HCA Healthcare maintains elaborate OKR-based goal structures focused on reducing administrative burden, improving workflows, and enhancing patient care quality, with the philosophy that financial value will follow from these operational improvements.
Davita explicitly distinguishes between “return on employee” and “return on investment,” willing to prioritize employee productivity gains first before measuring traditional ROI. They focus on reducing friction in administrative work so employees can concentrate on their core competencies. This approach recognizes that in a labor-intensive industry facing workforce shortages, metrics like clinician retention and job satisfaction carry significant economic value even when difficult to quantify directly.
Independence Blue Cross encourages all team members regardless of role to “build their own business case” rather than treating ROI as solely a finance or business owner responsibility. They advocate for developers and technical staff to think holistically about value creation, connecting upstream and downstream impacts of their work. This democratization of value thinking supports identifying opportunities that might not surface through formal business case processes.
The panel strongly emphasizes that agentic AI expands the security risk surface beyond protecting data to protecting workflows, decisions, and ultimately patient outcomes. Organizations reference the concept of “agentic defense” where AI-based security systems can reason about threats, adapt to attack patterns, and detect anomalies at scale. The acquisition of Wiz by Google is mentioned in the context of AI security focus, with 49% of healthcare organizations already deploying AI-based cybersecurity operations.
The governance challenge intensifies as organizations deploy agents across departments. Without unified governance frameworks, the risk of creating ungoverned agents with inconsistent security postures and audit trails becomes significant. Elevance’s framework approach directly addresses this by embedding governance controls into reusable capabilities rather than treating governance as a separate layer applied to each implementation.
The panel emphasizes that in healthcare, trust is the fundamental currency and healthcare moves at the speed of trust. This makes auditability, explainability, and consistent governance non-negotiable for production deployments. Organizations cannot accept the complexity that would arise from thousands of agents deployed with different governance standards across the enterprise.
Multiple speakers identify change management and workforce upskilling as the critical factors determining whether organizations remain stuck in pilot purgatory or achieve production scale. The panel highlights a significant gap where 82% of healthcare executives believe AI literacy is important but only 30% feel their organizations have adequate curriculum in place.
The vision involves AI literacy programs spanning every role in the organization, not just engineers and data scientists. Staff need training to operate AI systems, manage their outputs, audit their decisions, and understand their limitations. HCA Healthcare’s domain-based approach explicitly includes change management teams that work with frontline staff on AI literacy and gather user feedback to inform system refinement.
The panel emphasizes that AI adoption is fundamentally a change management challenge rather than a purely technical problem. Technology decisions must account for the human beings who will use these systems. The most technically sophisticated solution will fail at scale if it doesn’t fit naturally into user workflows and if users lack the literacy to work effectively with AI capabilities.
Organizations emphasize the need for conscious budget investment in workforce upskilling, positioned as equally important as technology infrastructure investment. This represents a shift from viewing training as overhead to recognizing it as a core enabler of technology ROI.
Looking forward to 2027, organizations envision multi-agent enterprise architectures where agents handle handoffs between departments and functions. Davita articulates a vision of patient experience driven by orchestration of caregiver agents working in coordination. This represents a significant evolution from single-purpose agents to ecosystems of specialized agents that collaborate.
Elevance describes their goal as AI becoming embedded in every workflow, process, and interaction rather than remaining a separate tool users must access. The vision involves AI capabilities so seamlessly integrated into daily work that they become invisible infrastructure, focused on member outcomes, provider needs, broker requirements, and associate productivity.
Independence Blue Cross envisions every employee having a self-governing personal AI assistant that proactively prepares work rather than reactively responding to queries. An underwriter would begin their day with proposals already drafted based on overnight data analysis. A marketing professional would find campaign options prepared for review. The vision compresses active work time while freeing hours for strategic thinking and wellbeing.
This multi-agent future requires the unified governance and framework approaches described earlier. Without common infrastructure for agent creation, monitoring, and security, organizations risk recreating the fragmented application landscape that has historically plagued healthcare IT, where thousands of applications exist but few meaningfully integrate with patient data in interoperable ways.
Several key lessons emerge across the organizations’ experiences moving from pilots to production. First, tight workflow integration is non-negotiable—solutions that require users to step outside their natural work patterns will fail to achieve adoption regardless of technical sophistication. Second, domain expertise must inform technical decisions, requiring cross-functional teams that include process engineers and frontline staff alongside data scientists and developers. Third, reusable frameworks with embedded governance enable faster deployment at scale compared to bespoke solutions reviewed individually.
Fourth, organizations must balance rigor and speed by clearly delineating use cases where faster deployment is appropriate versus those requiring more careful review. Administrative and back-office applications can generally move faster while clinical decision support requires more comprehensive validation. Fifth, measuring value broadly beyond financial ROI to include patient outcomes, clinician satisfaction, and employee productivity creates sustainable justification for continued investment.
Sixth, unified data and AI platforms reduce complexity compared to point solutions from multiple vendors. All four organizations standardized on Google Cloud Platform and Gemini, viewing this consolidation as essential for maintaining governance consistency and avoiding integration challenges. Seventh, executive sponsorship and organization-wide AI strategy set at the top level enables coordinated deployment rather than disconnected experiments.
While the panel focuses more on organizational strategy than deep technical details, several architectural patterns emerge. Organizations emphasize multi-modal capabilities combining voice, vision, and text processing. They highlight RAG patterns for grounding agent responses in organizational knowledge and patient-specific data. Document intelligence represents a critical capability given the volume of unstructured information in healthcare operations.
The emphasis on workflow intelligence suggests sophisticated orchestration capabilities where agents manage multi-step processes with dependencies and handoffs. Voice intelligence powers conversational interfaces for both patient-facing and internal applications. The mention of SDLC intelligence indicates applying LLMOps principles to the software development process itself, using AI to improve developer productivity.
The reusable framework approach implies significant investment in platform capabilities rather than application-by-application development. This platform thinking includes prompt engineering standards, evaluation frameworks, monitoring and observability tooling, and security controls that can be consistently applied across use cases.
The organizations position 2026-2027 as critical years for moving from experimentation to execution at scale. They emphasize that the window for pilots has closed and operational imperatives now demand production deployment. Success will depend on solving organizational challenges around change management, governance, and cross-functional collaboration as much as on technical innovation.
The vision of ubiquitous AI agents embedded in every workflow, supporting every employee, and augmenting every patient interaction represents a fundamental reimagining of healthcare operations. Achieving this vision requires the unified platforms, governance frameworks, and change management approaches the panel describes rather than continuing the historical pattern of layering point solutions onto existing complexity.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
Medable developed Agent Studio, a comprehensive platform for deploying AI agents in clinical trial operations to address the lengthy drug approval process that currently takes over 10 years. The platform enables both internal teams and customers to build configurable multi-agent systems that tackle problems like document classification in electronic trial master files and clinical research monitoring across multiple data systems. By taking a platform-first approach with support for model-agnostic agents, RAG knowledge integration, MCP connectors, workflow functionality, and robust evaluation frameworks, Medable has deployed multiple agentic applications that help clinical research associates process over 80,000 documents per year and monitor data across 13+ disparate systems, with the ambitious goal of reducing clinical trial timelines from 10 years to one year.
Anthropic developed a production multi-agent system for their Claude Research feature that uses multiple specialized AI agents working in parallel to conduct complex research tasks across web and enterprise sources. The system employs an orchestrator-worker architecture where a lead agent coordinates and delegates to specialized subagents that operate simultaneously, achieving 90.2% performance improvement over single-agent systems on internal evaluations. The implementation required sophisticated prompt engineering, robust evaluation frameworks, and careful production engineering to handle the stateful, non-deterministic nature of multi-agent interactions at scale.
