Forget the Hype. These Are the Three Things AI Can Actually Do in Healthcare Right Now.
Ramani Narayan · May 2026 · 6 min read
We’ve spent several posts in this series diagnosing a problem. The broken data layer. The EHR that became a billing system. The clinical ontologies most healthcare AI skips entirely. If you've been following along, you might reasonably wonder: so, what can AI do?
That's the right question. And the honest answer is both more modest and more powerful than most of the hype suggests.
AI can't fix a physician’s judgment. It can’t replace the relationship between a clinician and a patient. And despite what you may have read, it isn't about to render most of the healthcare workforce obsolete.
But once the data layer is right — once clinical information is normalized, grounded, and semantically understood — there are exactly three things AI can do in healthcare. And all three are possible today.
Retrieve the right information. Take the right action. Write the right words. That’s it. And that's enough to change how medicine is practiced.
01. Find It — Retrieval
The most underappreciated capability of clinical AI isn't generation. It’s retrieval.
Consider what a physician does before they walk into a patient room. They're mentally assembling a picture: what's the patient's history, what medications are they on, have they had this complaint before, what did the last specialist say? In a complex patient with years of records spread across multiple systems, that assembly can take fifteen minutes of hunting — before a single word is spoken to the patient.
AI retrieval changes that equation. Not just by summarizing the chart (though it can do that). By answering the specific clinical question a physician has right now, drawn from the full longitudinal record, including notes, labs, imaging reports, and prior visit summaries, across every system that holds the patient's data.
"Does this patient have any prior episodes of hyponatremia?" That question, which might require a nurse to spend twenty minutes pulling labs across three systems, becomes a two-second answer.
The prerequisite — and this is the part the glossy AI demos never show — is that the retrieval layer must be grounded in the actual source record. Every answer must cite its source. Every claim must be traceable. A retrieval system that hallucinates in a clinical context isn't a productivity tool. It's a liability.
The bottleneck in most physician workflows isn't decision-making. It's information assembly. Retrieval AI solves the right problem.
This is why the data layer underneath retrieval matters so much. If the patient's records are fragmented across systems, if lab values are coded inconsistently, if medication names aren't resolved to a common terminology, the retrieval layer fails — not because the AI model is bad, but because it has nothing reliable to work with.
Get the data layer right, and retrieval AI does something genuinely remarkable: it gives a physician standing in a hallway between appointments the same depth of patient context that used to require a pre-visit prep call.
02. Do It — Agentic Action
The second capability is the one that makes administrators nervous and engineers excited in equal measure: AI that doesn't just answer questions but takes actions.
In healthcare, ”agentic” AI means AI that can interact with clinical systems on behalf of a physician or care team — placing orders, scheduling follow-ups, routing referrals, pre-authorizing procedures, queuing tasks to the right person at the right time.
The canonical example is the physician inbox. Physicians at large practices routinely receive hundreds of messages per week: refill requests, lab result notifications, patient questions, administrative tasks, results that need review. Each one requires opening a message, reading context, deciding, and taking an action. The average primary care physician spends two hours per day on this alone.
Agentic AI works through that queue. It reads the message, retrieves relevant context from the patient record, proposes an action — refill approved, and results communicated to patient, referral queued — and presents it for a single physician approval click rather than a ten-step manual workflow.
The physician is still making every decision. But instead of also being the person who assembles context, routes messages, and executes tasks, they're the person who reviews and approves. That’s a meaningful difference.
The more ambitious version of agentic AI works across the full care continuum. Prior authorization is a particular target:: a process that currently requires a human to assemble clinical evidence, navigate a payer portal, and often resubmit multiple times takes up an estimated three hours per physician per week. An agentic system that understands the clinical record and the payer's criteria can automate most of that — not by bypassing the process, but by doing the assembly and submission work that currently falls on the clinical staff.
The caution here is real. Agentic AI operating inside clinical systems is high-stakes. An action taken erroneously — the wrong order placed, the wrong patient's record queried — has consequences that a bad search result doesn’t. This is why agentic systems in healthcare need something almost no AI vendor talks about: a complete, auditable record of every action taken, every source consulted, every decision made. The physician who approved the action needs to be able to reconstruct exactly what the AI saw and what it did.
That audit layer isn’t a nice-to-have. In a regulated environment, it’s the thing that separates deployable AI from a demo.
03. Write It — Generation
The third capability is the one most people think of when they hear “AI in healthcare”: generation. Writing things.
The most mature application is ambient clinical documentation — AI that listens to a physician-patient encounter and generates a draft clinical note in the appropriate format. After-visit summaries. Referral letters. Patient instructions. Discharge documentation. These are the tasks that consume an estimated two hours of every clinical shift in documentation time.
Generative AI doesn’t eliminate that work. It drafts it. The physician still reviews, edits, and signs. But the difference between starting from a blank screen and starting from a well-structured draft is enormous — not just in time, but in cognitive load at the end of a twelve-hour shift.
The less obvious application is patient communication. The average physician spends forty minutes per day writing responses to patient messages in the EHR inbox — messages that require reading context, formulating a clinically appropriate response, and writing it at a level the patient will understand. Generation AI drafts those responses. The physician reviews and sends (or edits and sends). The time savings compound across a practice.
Generation AI’s value isn’t that it writes better than a physician. It’s that it writes faster, so physicians can spend their cognitive energy on the decisions that require their training.
The prerequisite for good clinical generation is exactly what it is for retrieval: grounded context. A generation system that drafts a patient message without access to the patient’s actual record is producing creative writing, not clinical communication. The note must reflect what happened. The lab result must be accurate. The medication instructions must match the prescription.
Which brings us back, again, to the data layer.
The Three Together
Retrieval, action, and generation are individually useful. Together, they're transformative — but only when they're integrated and working from the same clinical substrate.
Consider the pre-visit workflow. Retrieval surfaces the relevant history, outstanding care gaps, and recent results before the physician walks in. During the encounter, generation captures the note. After the encounter, agentic action processes the orders, schedules the follow-up, queues the referral, and drafts the patient communication. The physician has been present for every decision. The AI has handled everything else.
This is what ”clinical intelligence at the speed of care” means. Not AI making decisions independently. AI eliminating the administrative surface area that currently sits between clinical decisions and clinical action.
Today, a physician’s workflow looks like: decide, then spend twenty minutes on the logistics of acting on that decision. The three capabilities flip that ratio.
The reason most attempts at this fail isn’t the AI. It's the data layer. A retrieval system without grounded records retrieves noise. An agentic system without a complete patient record acts on incomplete information. A generation system without clinical context produces text that sounds right but isn't. The model is only as good as what it can see — and in most healthcare IT environments today, what the model can see is a fraction of what the physician needs.
That’s the gap that must be closed before any of this works at scale. And closing it is harder, slower, and less glamorous than building a better model. Which is exactly why most AI vendors prefer not to talk about it.
What’s Actually Hard
Building models that can retrieve, act, and generate in healthcare isn’t the hard part anymore. The models exist. The inference infrastructure exists. The APIs exist.
What's hard is the layer between those models and the raw clinical data they need to work with: the normalization, the ontology mapping, the FHIR transformation, the grounding infrastructure, the audit trail that makes every output accountable to a source record.
That layer is what determines whether AI in healthcare is a demo or a deployment. Whether it's something a physician uses once and doesn't trust, or something that becomes as essential to their workflow as their stethoscope.
We built RISA to close that gap. Not to build a better LLM. To build the infrastructure layer that makes the LLMs that already exist work in clinical environments. The retrieval, the agentic action, the generation — those are the application. The data layer is the foundation. And foundations, it turns out, are what everything else depends on.
This post is part of The Clarity Protocol, ThetaRho‘s ongoing series on AI, clinical workflow, and healthcare data.
The next piece asks the question that follows naturally from these three capabilities: when AI retrieves, acts, and writes on behalf of a clinician — who is accountable when something goes wrong? The governance layer that most AI vendors are quietly hoping nobody asks about.
ThetaRho (thetarho.ai) builds clinical AI infrastructure for healthcare organizations. RISA is our clinical intelligence platform — HIPAA-compliant, AICPA SOC certified, and live on the athenahealth Marketplace.