From Experimentation to Clinical-grade AI in Healthcare
For years, enterprise AI technique has operated on a easy assumption that when the fashions get ok, adoption will comply with. That assumption is now being examined — and failing. Model functionality has, by most measures, arrived. What hasn’t arrived but are the infrastructure, safety posture, and workflow structure wanted to allow autonomous techniques to function safely throughout the enterprise.
One clear sign of this hole comes from the federal authorities itself. When NIST published a proper Request for Information on AI agent safety in January 2026, it drew 932 public feedback earlier than its March 9, 2026 shut—a rare response quantity that displays how urgently practitioners are grappling with issues present frameworks don’t tackle.
NIST’s personal evaluation of why is blunt: autonomous brokers are being embedded into manufacturing environments with out the id administration infrastructure, entry controls, or audit mechanisms that govern conventional software program, at the same time as they write and execute code and chain software calls throughout dozens of built-in providers.
The safety information underlying that concern is sobering. Novel assault methods in opposition to AI brokers succeeded 81% of the time in early-2025 red-team workout routines referenced in NIST’s inside analysis — a failure charge that has nothing to do with how effectively the underlying mannequin causes.
Healthcare illustrates the workflow aspect of the issue most sharply. A proper remark submitted to HHS’s Office of the National Coordinator for Health IT argues that capabilities associated to information readiness, interoperability, life cycle monitoring, and auditability are foundational to realizing AI’s potential whereas safeguarding sufferers, clinicians, and establishments, significantly for AI embedded in documentation and operational workflows that affect care with out being regulated as a medical system.
The sample throughout each supply is similar: the bottleneck isn’t intelligence. It’s readiness.
Alex Tyrrell, SVP and CTO of Health at Wolters Kluwer, joined Emerj’s Matthew DeMello on the AI in Business Podcast to clarify why agentic AI success now is dependent upon modernizing enterprise infrastructure, safety posture, and workflow structure slightly than enhancing mannequin efficiency.
This article examines three insights that make clear why agentic AI adoption is constrained by enterprise readiness slightly than mannequin functionality.
- Infrastructure readiness because the gating issue: Agentic AI can not function reliably when legacy techniques, brittle APIs, and monolithic architectures stop autonomous execution throughout multi‑step, regulated workflows.
- Domain‑tailored reasoning because the engine of agentic efficiency: Multi‑mannequin orchestration, high-quality‑tuning, and dynamic chain‑of‑thought are required to decompose advanced duties into executable steps that persistently ship outcomes.
- Autonomous safety posture as the brand new enterprise requirement: Machine‑initiated actions develop the assault floor, complicate compliance, and demand id, entitlement, and observability controls constructed for brokers slightly than human operators.
Episode: From Experimentation to Clinical-grade AI in Healthcare – with Alex Tyrrell of Wolters Kluwer
Guest: Alex Tyrrell, SVP and CTO of Health at Wolters Kluwer
Expertise: Artificial Intelligence, Machine Learning, Healthcare Technology, Data & Product Engineering
Brief Recognition: Alex Tyrrell is a expertise govt with experience spanning AI, machine studying, information platforms, and healthcare expertise. He at present serves as Executive Vice President and CTO of Health at Wolters Kluwer, the place he leads product engineering, expertise technique, and AI and Data Centers of Excellence throughout the Health Division’s portfolio of scientific resolution assist, analysis, and healthcare options. Previously, Alex held engineering and product management roles at Thomson Reuters and Refinitiv, the place he led work throughout search, pure language processing, machine studying, information platforms, and enterprise content material options. Earlier in his profession, Alex was a postdoctoral fellow at Massachusetts General Hospital and Harvard Medical School, conducting analysis in high-performance computing, picture evaluation, and computational modeling that contributed to publications in journals together with Nature Medicine, Nature Methods, and the New England Journal of Medicine. He holds a Ph.D. in Computer Engineering from Rensselaer Polytechnic Institute and an M.S. in Computer Science from Rochester Institute of Technology.
Infrastructure Readiness because the Gating Factor
Alex opens the episode by drawing a pointy line between what agentic AI can do and what enterprises are literally ready to assist. His core message is that autonomous techniques don’t fail as a result of the fashions are weak — they fail as a result of the enterprise stack beneath them was constructed for human‑paced, display‑primarily based workflows. Agentic AI removes that friction completely, exposing architectural weaknesses that had been beforehand hidden.
According to Alex, the most important constraint just isn’t an urge for food for innovation however technical debt: monolithic purposes, brittle APIs, coarse entitlements, and operational tooling designed for human operators slightly than autonomous brokers. When brokers start executing duties at machine pace, these weaknesses turn out to be systemic blockers.
He emphasizes that enterprises should modernize the underlying infrastructure earlier than they’ll safely deploy brokers into regulated workflows. That modernization consists of decomposing monoliths, tightening entitlements, instrumenting APIs, and upgrading observability so groups can hint autonomous actions throughout distributed techniques.
Alex states::
“When you had all these screens and these buttons and human operators… the amount of site visitors was virtually considerably metered by how rapidly you’ll be able to press these buttons. Now brokers are doing the work. There’s no extra friction. The charge and pace and quantity they’ll ship to your again finish — you higher be prepared for that.”
- Alex Tyrrell, SVP & CTO, Health, Wolters Kluwer
From this, Alex units out a number of sensible steps enterprises should take earlier than deploying agentic techniques:
- Decompose monoliths into modular, observable parts: Agents can not function inside black‑field techniques; they require clear, auditable endpoints and predictable habits.
- Redesign APIs for high-quality‑grained, least‑privilege entry: Coarse or brittle APIs break beneath autonomous load, and inadequate entitlements create compliance danger in regulated workflows.
- Engineer for machine‑pushed site visitors spikes: Human workflows naturally throttle system load; brokers don’t. Infrastructure have to be ready for sudden, excessive‑quantity execution.
- Modernize observability and monitoring: When brokers act throughout a number of techniques, conventional logging can not clarify the place errors originate or why spikes happen. Distributed tracing turns into obligatory.
- Strengthen id and entry controls for brokers: Autonomous techniques require id fashions and permissioning frameworks tailor-made to machine‑initiated actions.
Together, these steps type Alex’s core argument: agentic AI succeeds when the enterprise basis is rebuilt for autonomy slightly than human operation.
Domain‑Adapted Reasoning because the Engine of Agentic Performance
Rather than starting with structure or infrastructure, Alex pivots this a part of the dialog towards the habits of agentic techniques — particularly, how they assume. His argument is that enterprises persistently underestimate the quantity of area adaptation required for brokers to carry out reliably. The query is rarely “which mannequin,” however “what number of layers of area logic have to be embedded earlier than the mannequin can truly execute a regulated process?”
Alex describes a sample he sees throughout each critical deployment: the off‑the‑shelf mannequin is simply the place to begin. Once it enters an actual workflow, it have to be reshaped — by supervised instruction, high-quality‑tuning, low‑rank adaptation, and dynamic chain‑of‑thought — till it displays the reasoning patterns of skilled professionals. In his view, that is the place the enterprise’s true differentiation lives.
A key second in the dialog comes when Alex explains how dramatically a mannequin modifications as soon as it’s tailored to a site:
“You may begin with a foundational frontier mannequin, however you’re going to area‑adapt it… and it’s not going to appear to be what you took off the shelf. It’s going to be higher — and also you’re doubtless going to do that a number of instances.”
- Alex Tyrrell, SVP & CTO, Health, Wolters Kluwer
Instead of itemizing steps, Alex illustrates the idea by the character of agentic workflows themselves. A human may see a single process — a pre‑authorization test, a declare evaluation, a scientific abstract — however an agent should break that process into reasoning items which can be usually invisible to the human operator. Sometimes the agent decomposes the duty extra finely than a human would; different instances it collapses a number of human steps into one. The adaptation course of teaches the mannequin how to make these selections.
Alex’s sensible steering emerges from this framing:
- Domain experience turns into the scaffolding for agent reasoning. Subject‑matter consultants are not simply reviewers — they outline the reasoning pathways brokers should study.
- Multi‑mannequin techniques turn out to be the norm. Different elements of a workflow require totally different reasoning behaviors, and no single mannequin can reliably cowl the whole chain.
- Dynamic chain‑of‑thought turns into important. Agents should regulate their reasoning to the precise details of every occasion, slightly than counting on static prompts or generic directions.
- Explainability comes from granularity. When duties are decomposed into high-quality‑grained reasoning steps, enterprises acquire the transparency required for regulated environments.
Alex’s overarching level is that agentic efficiency just isn’t a property of the mannequin — it’s a property of the variation course of. Enterprises that deal with area reasoning as a primary‑class engineering self-discipline will see brokers behave with the consistency and reliability their workflows demand.
Autonomous Security Posture because the New Enterprise Requirement
Alex’s ultimate theme shifts from workflow and reasoning to the operational actuality enterprises face as soon as brokers start appearing inside regulated environments. His level is simple: the safety assumptions which have ruled cloud and SaaS techniques not maintain when autonomous techniques provoke actions, talk with different brokers, and generate machine‑paced site visitors throughout the enterprise. The assault floor expands, the id mannequin modifications, and compliance frameworks should evolve to account for machine‑pushed habits.
Rather than treating safety as a supporting concern, Alex frames it because the first constraint enterprises encounter when brokers transfer from experimentation to manufacturing. Identity, entitlements, auditability, and observability — all traditionally designed for human customers — have to be rebuilt for autonomous actors able to initiating workflows, accessing delicate information, and interacting with exterior techniques. In his view, this shift just isn’t incremental; it’s structural.
He highlights a second dimension: menace actors now profit from the identical generative capabilities enterprises use. LLM‑enabled assaults turn out to be more durable to detect, simpler to scale, and extra adaptive. This forces enterprises to rethink how they validate agent id, monitor agent‑to‑agent communication, and keep compliance beneath HIPAA, SOC 2, ISO, and comparable frameworks.
Alex’s most pointed remark in this part comes when he describes how conventional operational alerts break down as soon as brokers change human clicks:
“Where did the spike come from? Where did this error come from? The observability, the monitoring — it’s a problem, so you will have to adapt.”
- Alex Tyrrell, SVP & CTO, Health, Wolters Kluwer
From this framing, a number of sensible insights emerge:
- Agent id turns into a primary‑class safety primitive. Enterprises have to be ready to confirm which agent is appearing, beneath what permissions, and with what provenance — not simply which human person initiated a session.
- Least‑privilege entitlements have to be redesigned for autonomous habits. Agents require narrower, extra specific permission boundaries than human operators, particularly in healthcare and monetary workflows.
- Auditability should lengthen to machine‑initiated reasoning steps. It is not adequate to log API calls; enterprises have to be ready to hint how an agent arrived at a choice and what information it used.
- Observability should evolve to detect non‑human site visitors patterns. Traditional monitoring can not clarify machine‑pushed spikes or cross‑system cascades triggered by autonomous workflows.
- Compliance frameworks should incorporate agent habits. HIPAA, SOC 2, and ISO controls have to be interpreted by the lens of autonomous execution slightly than human‑mediated workflows.
Alex’s overarching message is that agentic AI modifications the safety posture from the bottom up. Enterprises that modernize id, entitlements, observability, and compliance for autonomous techniques can be positioned to deploy brokers safely and sustainably throughout regulated environments.