AI Agents in Healthcare: Diagnosis, Triage & Coordination

The first time I sat in a clinical informatics meeting and heard a chief medical officer describe an agent as a member of the care team, I had to remind myself how recently that sentence would have been unthinkable. By 2026 it is not just thinkable, it is being operationalised in academic medical centres, integrated delivery networks, and community hospitals across the United States, the United Kingdom, India, and most of continental Europe. The combination of crushing administrative burden, persistent clinician burnout, and the explosive maturity of large language models has produced one of the fastest technology adoptions I have ever seen in healthcare.

What makes the healthcare conversation different from finance or marketing is the stakes. A poorly designed marketing agent loses an opportunity. A poorly designed clinical agent can harm a patient. That tension between the promise and the peril shapes every paragraph of this guide.

In what follows I want to share what I have learned from advising hospital systems, life sciences companies, and digital health vendors on how to deploy AI agents in healthcare responsibly. The aim is not to hype the technology, and it is not to dismiss it. The aim is to give clinical, technical, and operational leaders a clear map of where the genuine wins are, what the regulatory landscape looks like in 2026, and how to avoid the failure modes that I have seen up close.

Healthcare has the largest administrative overhead of any major industry. Estimates from the National Academy of Medicine suggest that physicians spend roughly two hours on documentation and administrative work for every hour of direct patient care. That ratio is the single most important reason to be optimistic about agents in clinical settings. If a well-supervised system can recover even thirty per cent of that time, the productivity, quality, and burnout improvements are transformational.

The same characteristics that make healthcare promising also make it difficult. Clinical decisions are high stakes, the regulatory environment is unforgiving, the data is fragmented across decades of legacy systems, and the workforce is rightly cautious about delegating judgement to a machine. The wins are real, but they require patience and discipline.

“I have learned to ask three questions before approving any clinical agent deployment. What is the worst thing this agent can do? Who notices when it does it? And how quickly can we turn it off?”

Clinical decision support has existed for decades in the form of rule-based alerts inside the electronic health record. The alerts are famously over-triggered, under-evidenced, and largely ignored. Agentic clinical decision support is qualitatively different. It can read the chart, summarise the relevant history, surface the latest guidelines, and reason about the specific patient in front of the clinician.

The most credible deployments I have seen do not try to make diagnoses. They do four things very well.

• Pull the relevant prior labs, imaging, and notes from across the longitudinal record.
• Compare the current presentation against the most recent guideline.
• Identify gaps in the work-up that the clinician may want to consider.
• Draft the differential and the orders for the clinician to review.
   

These numbers come from peer-reviewed pilots and from internal data shared with me under non-disclosure. I would not treat any of them as guaranteed. I would treat them as evidence that the category is real.

Triage is one of the highest-leverage places to deploy an agent because the decision is structured, the consequences of error are well understood, and the system is already drowning in volume. I have seen good results in two settings.

In the emergency department, a triage agent reads the chief complaint, the vital signs, and the initial nursing assessment and proposes an acuity level with the supporting reasoning. The nurse retains the decision. The agent’s value is consistency and the surfacing of subtle red flags that a tired clinician at the end of a twelve-hour shift may miss.

In primary care and urgent care, a patient-facing triage agent gathers the history, suggests the appropriate care level, and books the visit. The win is not just operational efficiency. It is keeping non-urgent cases out of the emergency department and routing genuine emergencies to the right setting faster.

The risk that keeps me cautious is under-triage. A system that misses one case of meningitis in ten thousand is a system that has done immense net good and will still be remembered for the case it missed. That asymmetry has to be designed into the controls.

Care coordination is the unglamorous backbone of modern medicine. It is also where I expect agents to deliver some of the largest patient-experience improvements over the next three years. A care coordination agent can do the following:

• Track every open referral and follow up automatically.
• Reconcile medication lists across the primary care provider, the specialist, and the pharmacy.
• Identify patients who are overdue for screening and queue outreach.
• Coordinate transport, social services, and home care for complex patients.
• Draft the after-visit summary in the patient’s preferred language and reading level.

These tasks are currently performed by armies of medical assistants and case managers who are themselves stretched thin. The agent does not replace them. It amplifies them by handling the structured work so the humans can focus on the conversations that require empathy and judgement.

If there is one category where the value is already proven beyond reasonable doubt, it is ambient documentation. Tools like Abridge, Suki, and Nuance DAX listen to the clinical encounter, produce a structured note, and place it in the electronic health record for the clinician to approve. The published evidence and the deployment data are consistent. Clinicians using these tools recover roughly an hour a day, report substantially lower documentation burden, and stay in the profession longer.

The scribe market is interesting because it shows what a mature agent category looks like. The vendors have invested heavily in EHR integration, billing code suggestions, and specialty-specific templates. The pricing has settled into a per-physician-per-month model that the chief financial officer can model easily. The legal and compliance questions have been answered well enough that procurement is no longer the bottleneck.

I expect the next wave of scribe innovation to move from documenting the visit to driving the visit. Imagine a scribe that not only writes the note but also surfaces the right order set, books the follow-up, and educates the patient, all while the clinician is still in the room.

Every clinical agent has to be deployed under a privacy framework that the institution’s general counsel can defend. In the United States that means HIPAA business associate agreements with every model and infrastructure vendor, encryption at rest and in transit, role-based access controls, and audit logs that satisfy the Office for Civil Rights. In the European Union it means GDPR Article 9 special-category processing with a defensible lawful basis. In the United Kingdom the same framework applies through UK GDPR and the Data Protection Act. In India the Digital Personal Data Protection Act of 2023 has added a domestic compliance layer.

The practical implication is that the choice of model provider matters. The frontier labs now offer healthcare-tier deployments with zero data retention and business associate agreement coverage. Anything less is not deployable in a regulated clinical setting.

“I have walked away from more than one promising vendor because they could not provide a business associate agreement signed by the model provider, not just the application vendor.”       

The line between clinical decision support and a medical device is the line between an unregulated tool and an FDA-regulated product. The 21st Century Cures Act exempts clinical decision support that meets four criteria, the most important of which is that the clinician can independently review the basis for the recommendation. Agents that meet those criteria can be deployed without FDA clearance. Agents that do not have to go through the de novo or 510(k) pathway.

In 2026 the FDA has begun to articulate a more nuanced framework for what it calls predetermined change control plans for adaptive algorithms. The relevance for agentic systems is that the agency now expects vendors to declare in advance how the system will be updated and monitored in production. That is a constructive development. It also raises the bar for the second tier of vendors that have been shipping unregulated products with regulated-product capabilities.

If you are deploying an agent that touches diagnosis, treatment selection, or risk stratification, get regulatory advice early. The cost of getting it wrong is not just a fine. It is a recall and a reputational hit that the institution will spend years recovering from.

The electronic health record is the operational nervous system of every hospital. Any agent that does not integrate with it is, in practical terms, a science project. Epic, Oracle Health Cerner, MEDITECH, and Athenahealth dominate the United States market, and each has its own integration philosophy.

Epic has invested heavily in its own agentic capabilities through the MyChart platform and its developer programmes. Third-party agents typically integrate through FHIR APIs, SMART on FHIR apps, and the Epic App Orchard. Cerner offers a similar pattern through Oracle Health’s developer platform. The integration patterns are now mature enough that a competent engineering team can wire up a new agent in weeks rather than quarters.

The pattern I recommend is to keep the agent’s intelligence outside the EHR and to use the EHR as the system of record. The agent reads through FHIR, writes through structured endpoints, and never tries to be the source of truth.

The legal question I get asked most often is who is responsible when an agent makes a mistake. The honest answer in 2026 is that the law has not caught up with the technology. The working assumption, supported by the early case law and by guidance from the American Medical Association and the General Medical Council, is that the clinician who uses the agent remains responsible for the standard of care.

That answer is unsatisfying to clinicians who feel that they are being asked to underwrite a technology they did not build. It is also the only answer that is consistent with the existing professional liability framework. The practical implication is that institutions need to invest in training, supervision, and documentation so that clinicians can defend their decisions if they are ever challenged.

The vendor liability question is evolving more quickly. Several major vendors now offer indemnification for output-level errors caused by their systems, subject to caps and exclusions. I would not deploy a clinical agent from a vendor that refuses to share any liability.

There is a real tension between what the marketing materials claim and what the peer-reviewed literature supports. The honest assessment as of mid-2026 is that frontier models perform impressively on board-style examinations, perform unevenly on real-world clinical reasoning tasks, and remain prone to confident-sounding errors that a clinician would catch but that a junior trainee might miss.

The implication is not to avoid the technology. It is to design the deployment around the model’s actual capabilities. A scribe that drafts a note for clinician review is well within the current capability envelope. A fully autonomous diagnostic agent is not.

Patient-facing agents are where I am most excited about the long-term impact. The average patient struggles to understand their own diagnosis, their medication list, and their care plan. A well-designed agent can explain a condition at the patient’s reading level, in their preferred language, with cultural sensitivity, around the clock.

The deployments I have seen working well share three properties. They are scoped to a clear set of conditions or workflows. They escalate to a human at well-defined triggers. They never make claims about diagnosis or treatment that exceed the evidence.

The category is sensitive because patients are vulnerable, because the information is personal, and because the consequences of error fall hardest on the people with the least ability to absorb them. Get this one right and you transform care. Get it wrong and you erode trust in the entire institution.

The most common reason clinical agent deployments fail is not the technology. It is the failure to redesign the workflow around the agent. I have seen institutions spend millions on a scribe deployment only to discover that the note review process is now the bottleneck because nobody changed the clinic schedule.

The discipline I recommend is to map the current workflow in detail, identify the steps the agent will change, and redesign the human role around the new capability. The change management work is at least as important as the technology work, and it cannot be delegated to a vendor.

A short, opinionated map of the categories I am tracking most closely.

• Ambient scribes. Abridge, Suki, Nuance DAX, DeepScribe. The category leaders are pulling away.
• Clinical decision support. Glass Health, Atropos Health, OpenEvidence. Early but promising.
• Care coordination. Notable, Memora Health, Hyro. Strong product-market fit in specific niches.
• Patient triage. K Health, Buoy Health, Ada Health. Maturing into integrated care models.
• Revenue cycle. Olive, Notable, Waystar agents. Real money in a previously sleepy category.
• Life sciences. Causaly, BenchSci, OpenAI for Science partnerships. The boundary with research is blurring.

If I were advising a health system on its first major agent deployment, here is the rough sequence I would follow.

• Establish a cross-functional governance committee with clinical, technical, legal, and patient representation.
• Pick one workflow with strong clinical sponsorship and clear measurable outcomes.
• Run a structured pilot with documented baseline metrics and pre-registered success criteria.
• Build the monitoring, logging, and incident response capability before the agent goes live.
• Roll out in waves with mandatory training and clear escalation paths.
• Publish the results internally and externally so the field can learn from what worked and what did not.

The institutions that are doing this well are the same institutions that have always been good at translating evidence into practice. The technology rewards the same disciplines that have defined modern medicine.