Key Takeaways
- Johns Hopkins Medicine found that the best AI agent resolved only 28% of complex healthcare tasks on the first attempt.
- No agent in the study cleared 20% on strict, repeated testing, and performance collapsed to 3.8% in single end-to-end sessions.
- Most health systems remain in pilot mode, with few having expanded AI use across departments or functions.
- Prior authorization alone consumes an average of 13 hours of physician and staff time per week.
- Governance, reliability measurement, and workflow-first thinking must come before financial ROI when deploying healthcare AI agents.
Healthcare organizations are under increasing pressure.
Administrative costs consume an estimated 15 to 30 percent of all US healthcare spending. Staffing shortages are intensifying due to increasing demands for healthcare services. Cyberattacks threaten to disrupt critical operations, requiring stronger security measures. And amid all these challenges, industry regulators continue to push for higher-quality care.
Artificial intelligence has the potential to help solve many healthcare challenges, placing it solidly on CIO’s radars.
However, many of them are uncertain if AI can handle the complex workflows that exist in healthcare settings. Johns Hopkins Medicine decided to find out before deploying anything.
Can AI Agents Do Healthcare Work?
Healthcare administrative work depends on dense policies, numerous software systems, and repeatable decision-making. Can AI accurately do such complex work?
That was the first question Dr. T. Y. Alvin Liu, inaugural director of the Gills AI Innovation Center, asked before choosing any workflow to automate. Liu is a member of the Johns Hopkins Medicine AI Oversight Team.
Johns Hopkins partnered with actAVA to evaluate AI agents using CHI-Bench. Published in May 2026, the study tested AI agents across provider prior authorization, payer utilization management, and care management. Each task ran in a high-fidelity simulator of 20 healthcare applications exposed through 87 MCP tools, guided by a 1,290-page managed-care operations handbook. The results came from testing 30 distinct agent harness and model configurations.
Across all configurations, the best agent resolved only 28% of tasks on the first attempt. No agent cleared 20% on strict, repeated testing. When agents attempted complete end-to-end sessions, performance collapsed to 3.8%. The Healthcare IT News report on the study, Liu said researchers were surprised that most problems involved reasoning through policy-rich situations, rather than hallucinations or software integration failures.
"The model is necessary but not sufficient," Dr. Liu said. "The harness built for the workflow is what turns a capable model into a deployable agent."
The Gap Between Pilots and Production
KLAS Research's 2025 healthcare AI perception report found that most health systems remain in pilot mode. Few have expanded AI use across departments or functions. The primary barriers are a lack of governance frameworks, the need for ROI validation, and challenges in integrating AI into existing workflows.
When an agent stalls mid-workflow or produces an incorrect output in a regulated process, it can impact compliance, data security, revenue, and, most importantly, the quality of care.
It also erodes trust in the technology, and healthcare already has an AI adoption problem. Benchmarking establishes whether a system works before it touches real patients or processes, which helps increase trust and adoption.
Measure Reliability Before You Measure ROI
Johns Hopkins has not yet reached the stage where it can report financial ROI from its agentic AI program. However, it can measure reliability in terms of first-pass completion rates, handoff consistency, reasoning failures, tool-use failures, and unsafe completions. All of these are established as baselines before any production deployment begins.
This means testing against that organization's specific data conditions and compliance requirements. The failure modes that matter most are rarely the ones that appear in pre-deployment testing. They emerge when the agent encounters policy-dense edge cases at production volume. Without a proper implementation harness, the AI that performed well in the isolation of a pilot collapses under realistic operational load.
Taazaa built mediPulse AI Assist on this principle. The system evaluates the quality of its own intermediate results before responding. It reformulates queries when results are insufficient and maintains context across follow-up questions. Reliability was the design foundation, not a post-deployment validation step.
Design for Effective Healthcare AI Governance
Johns Hopkins established AI governance policies, vendor evaluation processes, and oversight structures earlier than most health systems. In healthcare, AI adoption has outpaced governance maturity across nearly every organization deploying it.
The Johns Hopkins AI governance program covers policy management, workforce education, audit-ready training records, and an agent registry aligned with the NIST AI Risk Management Framework, HIPAA, CMS health equity criteria, and ONC transparency requirements.
"Accountability requires knowing which agent did what, under which policy, and with what result," Dr. Liu said.
That level of specificity is not achievable without immutable auditing and failure reconstruction. Organizations deploying AI agents without it cannot answer regulator questions, resolve audit findings, or trace the cause of a workflow failure back to its source.
Understanding the benefits and risks of AI in healthcare is the foundation of any governance decision. The risks are not hypothetical. They are operational, regulatory, and clinical. Organizations that design for governance from the outset outperform those that address governance reactively.
Workflow Design Comes Before Technology Selection
At Johns Hopkins' Wilmer Eye Institute, staff members identify automation opportunities through structured workflow assessments. These examine frequency, complexity, manual effort, waiting time, data re-entry, interruptions, and burnout signals. Rather than beginning with technology selection, Johns Hopkins begins with understanding the work itself.
AI agents fail in production most often because the workflow was never fully understood before deployment began. Teams that start with technology build agents that perform well in controlled conditions. Production conditions are never controlled. The actAVA study confirms this: even the best-performing agents resolved fewer than three in ten tasks under realistic operational conditions.
Johns Hopkins is also building an "AI-native brain" that captures employees’ institutional expertise before they leave the organization. This valuable operational knowledge currently lives in email threads, PDF documents, and the memories of long-serving staff. Queryable, preserved knowledge becomes an infrastructure asset rather than a retention risk.
The landscape that agentic AI in healthcare must be benchmarked against includes use cases, deployment requirements, governance structures, and the specific criteria that determine whether an agent is ready for production.
Three Healthcare AI Recommendations
For healthcare organizations looking to deploy agentic AI, three moves consistently separate successful implementations from those that stall in production.
1. Define success across clinical, operational, and financial dimensions before the pilot begins. Agree on KPIs with operational leaders first. Decision-makers who are not involved from the outset will not be positioned to support the transition from pilot to enterprise deployment. Financial ROI measured before reliability baselines are established provides no way to detect whether an agent is creating downstream problems.
2. Benchmark the agent harness, not just the model. ActAVA tested 30 agent harness and model configurations and found that performance varied dramatically based on orchestration, not just the underlying model. Evaluate the full workflow system in conditions that resemble production. Demo environments make every agent look more capable than it is.
3. Establish governance infrastructure before the first agent goes live. An agent registry, audit trails, and accountability cannot be retrofitted after deployment. The organizations that skip this step consistently discover the gap at the worst possible moment: when a regulator asks, when an audit surfaces an anomaly, or when an agent produces an output that nobody can trace back to its source.
The organizations that benchmark carefully and govern deliberately before scaling will build systems that patients and regulators trust. Those who don't will discover the gap when reversing course is cost-prohibitive.
Contact Taazaa to build the benchmarking infrastructure, governance frameworks, and workflow design discipline that responsible agentic AI deployment in healthcare requires.
Frequently Asked Questions
Why benchmark healthcare AI agents before deployment?
ActAVA found that the best agents resolved only 28% of complex healthcare tasks on the first attempt under realistic conditions. Without benchmarking against production-equivalent scenarios, organizations have no reliable way to determine whether an agent is ready before it touches real patient data or administrative workflows.
What is an agent harness, and why does it matter more than the model?
The harness is the workflow-specific orchestration layer that connects your existing systems to your AI tools. ActAVA tested 30 harness and model configurations and found that performance varies dramatically based on orchestration. Model selection alone is not sufficient.
What does a healthcare AI governance program require?
At minimum, healthcare AI governance requires policy management, workforce education, audit-ready training records, and immutable auditing that shows which agent acted under which policy with what result. Governance policies should align with NIST AI RMF, HIPAA, CMS health equity criteria, and ONC transparency requirements.
Where should health systems start with agentic AI?
Start with workflow assessment before technology selection. Examine frequency, complexity, manual effort, waiting time, data re-entry, interruptions, and burnout signals. High-volume, manual-heavy workflows are the most defensible starting points. Model selection should be the last consideration.
