Executive overview
In 2025, the conversation around artificial intelligence in healthcare is evolving beyond simple question-and-answer systems to something far more dynamic: agentic workflows. AI agents—systems that can perform multi-step tasks autonomously—and agentic browsers that can read, reason, and act on web pages are moving from research demos to early deployments in healthcare dispatch, discharge follow-up, and EHR automation. These technologies promise a new level of speed and consistency, but they also introduce new and complex safety and security considerations (Forbes, Healthcare Dive, Fierce Healthcare).
The trend is exemplified by the launch of tools like Perplexity Comet, an AI-first browser designed for native automation and research across the web, with similar "agentic Browse" concepts emerging across the industry (Perplexity AI, Opera News). This article provides a guide for UK clinicians and healthcare leaders on what this new wave of agentic AI is, where it's already being used, and the critical governance required for its safe and effective implementation.
Definitions & scope
To understand this shift, it's important to be precise with our terms.
- AI agent: An AI system that can perceive its environment, make a plan, and act across different tools and APIs to achieve a specific goal. This could be anything from triaging patient outreach and matching them to a clinical trial, to scheduling appointments. Crucially, a well-designed agent knows when to escalate to a human (McKinsey & Company).
- Agentic browser: This is a web browser with a built-in AI agent. Instead of just rendering content for a human to read, it can interpret the content on a page, follow links, fill in forms, and automate tasks. Perplexity Comet and Opera’s Operator are leading examples of this new category (Perplexity AI, Opera News).
What exists today (market signals & exemplars)
While still an emerging field, agentic systems are already being deployed in real-world healthcare settings.
- Agentic browsers: Perplexity Comet is the most prominent example, offering an AI-first Browse experience that blends research with automation. Other entrants like Arc Search’s “Browse for Me” feature and Opera’s Operator concept signal a wider industry shift toward this "do it for me" model of web interaction (Perplexity AI, Arc, Opera News).
- Healthcare agents in the wild:
- Discharge follow-up: Hippocratic AI has partnered with US health system UHS to deploy voice agents that autonomously call patients after a hospitalisation to check on their recovery and escalate any concerns to a human clinician (Hippocratic AI, Healthcare Dive).
- EHR agentic roadmaps: EHR giant Epic has announced a roadmap that includes embedding agentic AI directly into its workflows for tasks like pre-visit preparation, clinician inbox triage, and generating patient-facing guidance (Fierce Healthcare).
- Emergency dispatch: Corti’s AI, which listens to emergency calls in real time, has demonstrated the ability to recognise the subtle cues of an out-of-hospital cardiac arrest, providing on-task assistance to the human call-taker (EENA, ScienceDirect).
Healthcare use cases where agents/agentic browsers add value now
- Knowledge & evidence co-pilot: An agentic browser can be tasked to synthesise information from NICE guidance, local Trust pages, and recent journal articles, returning a summary with all sources logged for audit.
- Patient communications & follow-up: Autonomous voice agents can handle routine post-discharge calls, send medication reminders, and use escalation logic to flag patients who need a human follow-up.
- Operational automations: Agents embedded in an EHR can draft referral packets, coordinate multi-departmental appointments, or prepare pre-visit summaries from recent results.
- On-scene/virtual triage augmentation: In emergency dispatch, an AI agent can extract signals from audio in real time to prompt the human call-taker with potential red flags, speeding up recognition of time-critical conditions.
How agentic browsers (e.g., Comet) fit clinical workflows
Perplexity Comet and similar tools offer two key advantages for clinicians:
- Research at the speed of care: A clinician can ask a complex question, and the agentic browser will read multiple web pages, synthesise the findings, and return a single, sourced answer. This avoids the "tab-sprawl" and wasted time of manual searching during a busy clinic.
- Task chains: The agent can be instructed to carry out multi-step tasks, such as finding and filling out a web-based prior authorisation template, collecting relevant PDFs from a trial registry, and assembling them into a case packet—all under human review.
Risks, new failure modes, and security
This new level of autonomy introduces new risks.
- Security of agentic Browse: Recent research flagged a significant vulnerability in an agentic browser that could have enabled a takeover via a path traversal attack. While it was patched, it serves as a stark reminder that these tools need to be built with hardened permissions and rigorous security reviews (Tom's Guide).
- Hallucination/action risk: An AI that can act (e.g., submit a form) is inherently more risky than one that only provides information. Agents that can act must provide traceable sources, support "dry-run" modes to preview their actions, and require explicit human confirmation for any high-impact step. This aligns with WHO and NHS guidance on oversight (World Health Organization, NHS England).
- Regulatory frontier: Fully autonomous clinical agents may outpace current medical device regulations. Expect increased scrutiny from regulators over their explainability and auditability. The MHRA’s AI Airlock provides a vital pathway for testing novel AI-as-a-medical-device (AIaMD) behaviours in a supervised sandbox (Medical Xpress, GOV.UK).
Governance & assurance (UK-centric, globally relevant)
- WHO LMM guidance: The World Health Organization's guidance on large multimodal models provides the key principles for transparency, risk management, and data protection that must be applied to agentic systems.
- NHS England guidance: The practical guidance already released for AI ambient scribes—covering human verification, DPIAs, and the need for a clinical safety case—can and should be adapted as a baseline for broader agent deployments.
- MHRA AI Airlock: For any agent that performs a clinical function that could classify it as a medical device, the AI Airlock is the appropriate route to trial its behaviour with UK regulators.
Architecture patterns (what to build/buy)
- RAG with provenance: Agents must be grounded in vetted clinical sources using a Retrieval-Augmented Generation architecture. They must render citations and log the provenance of each autonomous action.
- Guardrails & roles: The system must have clear "read-only" versus "can act" modes, with strong consent and approval checkpoints.
- EHR integration: Plan for agentic hooks near key decision points in the EHR, such as order-entry and inbox triage, which is the clear direction of travel from major vendors like Epic.
Near-term roadmap (30 / 60 / 90 days for a safe pilot)
- Days 1–30: Design the lane. Choose one low-risk, high-volume task (e.g., evidence lookup and packet assembly) and use an agentic browser like Perplexity Comet in "dry-run" mode. Measure the time-to-answer and citation correctness.
- Days 31–60: Add “act with approval.” Enable form-filling or download actions, but place them behind a mandatory human confirmation step. Run a DPIA and rehearse escalation paths as per NHS guidance.
- Days 61–90: Scale cautiously. Expand the pilot to a task like discharge outreach or pre-visit summaries. Track clear KPIs (turnaround time, override rate, safety incidents) and consider an MHRA AI Airlock-style engagement if the use case is clearly clinical.
KPIs that matter (beyond vanity metrics)
- Quality & safety: Citation accuracy, error intercepts by human reviewers, and incident reports.
- Efficiency: Time-to-answer for research tasks; time-to-packet for administrative tasks; staff satisfaction.
- Trust & robustness: Performance across different user groups; resilience to prompt-injection tests on hostile web pages.
FAQ
- What is an “agentic browser,” practically?
- It's a browser with a built-in AI agent that can understand the content on a page, navigate between pages, and complete tasks for you, such as filling a form. Perplexity Comet and Opera’s Operator are key examples.
- Are agentic systems already used in healthcare?
- Yes, early examples are live. These include post-discharge voice agents (UHS/Hippocratic AI), AI in emergency dispatch (Corti), and the agentic AI roadmaps from EHR vendors like Epic.
- What governance should we apply in the UK?
- Start with the WHO LMM guidance, adapt the principles from the NHS guidance on AI scribes (DPIA, safety case), and for any clinical device functions, engage with the MHRA AI Airlock pilot program.
- Are there security concerns?
- Yes, agentic Browse introduces new potential attack surfaces. It is essential that these tools undergo rigorous security testing and are deployed with strict permissioning.