Alex Fairweather argues that AI agent adoption has moved beyond theory and that the window to lead remains open, though it will not for long.
“What is now proved was once only imagined,” said William Blake. For much of the past two years, organisations have focused their AI efforts on generative AI; tools that respond to prompts, summarise records, draft letters and answer questions. Generative AI is, in essence, a very capable assistant that waits to be asked. Agentic AI represents the next evolution: systems that can set their own sub-goals, execute multi-step workflows and take actions autonomously, human-supervised but without needing a prompt for every action.
The distinction matters enormously. As McKinsey noted early this year, agentic AI “can function more like a coworker than a tool”. Where a generative AI tool might summarise a patient’s notes when asked, an agentic system can independently review outstanding results, flag anomalies, schedule the relevant follow-up and update the clinical record all within a single workflow cycle. This is not science fiction; it is already being piloted in revenue cycle management across leading US health systems, where McKinsey analysis projects a 30% to 60% reduction in cost-to-collect through agentic automation.
The opportunity in healthcare
Research published in Frontiers in Digital Health last year describes agentic AI as a system capable of “adaptive, efficient and ethical healthcare”, with particular promise across hospital operations, electronic health record management, clinical decision support and administrative workflow. A systematic review published by the Institute of Electrical and Electronics Engineers last year, covering studies from 2024 and 2025, found that agentic AI is being applied in diagnostics, autonomous documentation and simulation-based clinical training with the critical caveat that success depends on thoughtful workflow design and organisational readiness, not technology alone.
IBM’s Institute for Business Value healthcare report last year reinforces this view. Already, 39% of healthcare executives are using AI for inpatient monitoring and early warning systems. This is expected to be the area where full agentic AI implementation is achieved within the next three years.
In administrative functions, agentic AI is being widely adopted across the public and private sectors. Functions such as note scribing, letters, referral and discharge documents are produced and sent in two clicks, and automated call answering and integrated booking capabilities are already mature technology.
Governance: the non-negotiable condition
What separates agentic AI from its predecessors is that the consequences of errors compound differently. A generative AI producing a flawed summary can be caught and corrected. An agentic system that makes a flawed decision early in a multi-step workflow can propagate that error downstream before a human reviews it. This is why governance is a foundational prerequisite.
Research from Springer last year highlights that symbolic AI systems, which underpin safety-critical domains like healthcare, currently lack adequate governance models. McKinsey’s 2025 Technology Trends Outlook echoes this, stressing the “urgent need for governance frameworks” before organisations scale agentic deployments. Leaders should identify a small number of lower-risk, back-office use cases; billing, coding, booking, note generation and scheduling as the starting point. Build trust with stakeholders through measurable outcomes in controlled environments before advancing to any patient-facing workflow.
A practical starting point
The journey from generative to agentic AI need not be disruptive. The four-stage Identify, Align, Evaluate, Deliver framework I have outlined in previous articles applies directly here: identify the highest-volume, most error-prone process in your organisation; align an agentic solution with your strategic priorities and clinical governance structures; evaluate the capability, data and oversight requirements; and deliver with continuous monitoring in place.
Healthcare organisations that invest now in designing AI-native workflows, rather than layering automation onto broken processes, will be best positioned to capture lasting value.
As Capgemini’s 2025 research confirms, AI agent adoption has moved beyond theory, with 14% of organisations already implementing them at partial or full scale and 23% running pilots.
The window to lead, rather than follow, remains open. But it will not remain open indefinitely.
