'AI scribe' is the floppy disk of healthcare AI — what an agentic EHR looks like instead

Every EHR founder is now racing to build the 2021 version of healthcare AI. Most of the major allied-health platforms shipped AI scribes through 2024 and 2025 the way SaaS companies shipped "mobile-first" pages in 2014 — late, loud, and already overtaken by something bigger. I get the urgency; the demos look magical and the marketing writes itself. But the category is aging in real time. The future of an agentic EHR is not a faster note. It is a chain of small agents that coordinate the whole patient journey, end to end, while the clinician does the work only the clinician can do.

Here's what most people miss. A scribe is one function. An agent network is the operating system around that function. You can have the best scribe on earth and still leave the clinician chasing referrals, prior auths, no-shows, and the inbox at 9 PM. The scribe doesn't reach those things. The architecture around it doesn't let it.

The floppy disk wasn't bad — it was just narrow

The 3.5-inch floppy was beautiful when it shipped in 1985. It killed the 5.25-inch overnight. By 2000 it was gone, not because it failed but because the abstraction beneath it changed. We stopped thinking in disks and started thinking in files that travel — email attachments, USB sticks, the early web. The medium didn't matter once the surrounding system grew up around it.

That is the arc the AI scribe is on. Two years from being magical. Five from being Memorex.

I've watched the same thing happen up close in three demos this month. Beautiful transcription. Clean SOAP output. And then I asked the rep, "What does it do between visits?" Pause. "We're working on that." That pause is the whole story.

The scribe, stripped to function

A clinical scribe — the AI kind that ships inside Jane, SimplePractice, Carepatron, Heidi, and twenty others — is a specific, narrow piece of software. It may pull recent notes, templates, or chart snippets for context, but the core loop is still documentation: visit input in, draft note out, clinician edits before signing.

That's the centre of gravity. A scribe is transcription-plus-summarization wrapped around the clinical note.

What it doesn't do, by design:

• It doesn't maintain durable workflow state across the whole patient journey.
• It doesn't own the referral the clinician just verbally promised.
• It doesn't turn intake into structured chart fields unless the wider EHR gives it that job.
• It doesn't coordinate live checks, billing context, referrals, and follow-up as one workflow.
• It doesn't chase the prior auth, draft the referral letter, schedule the follow-up, and update the chart as a connected chain.
• It stops at documentation unless another module or a human picks up the thread.

Look at what the major vendors actually market on their AI pages. Jane's AI Scribe page is centred on documentation: ambient capture, transcription, structured note generation. SimplePractice's AI Note Taker reaches a step further — it produces pre-session summaries from the prior progress note and, according to its support docs, can surface clinical alerts — but the surface is still session prep and documentation, not the workflow around it. Carepatron markets a broader "first booking to final invoice" platform, with the AI scribe as one bundled component generating notes and templates. The shared pattern across these scribe pages is that the AI itself is sold as a documentation aid. I don't see these public AI pages claiming autonomous orchestration across intake, follow-up, billing, and re-engagement. That's where the gap lives.

The scribe is the floppy disk: useful, narrow, single-surface — and now table stakes.

What replaces it: a chain of agents, not a smarter scribe

The next layer is not "scribe with more features." It is a network of small agents, each owning one task in the patient journey, all sharing state with the chart and with each other. Kore.ai's 2026 healthcare agent report catalogues twelve real-world use cases for AI agents across healthcare — patient engagement, intake, scheduling, claims, and care coordination among them. The five-agent skeleton I describe below is my own synthesis of that landscape, mapped specifically onto the structure of an allied-health EHR. It's what an agentic EHR should be hosting natively, not bolting on.

Five agents I think every clinic will be running by 2027:

1. The intake agent. Runs a conversational interview with the patient before the first visit, extracts structured data (medications, allergies, comorbidities, presenting complaint, social history), populates the chart, and flags clinical risks the clinician should know before they walk in. It is not a smarter form. It is the form eliminated, replaced by a short asynchronous conversation that produces clean structured fields.

2. The pre-visit agent. Reads the chart, the new intake answers, any uploaded labs or imaging reports, and produces a one-page brief: what's changed since last visit, what to follow up on, what new evidence has been logged. The clinician opens the chart at 8:55 AM and the brief is already there.

3. The visit agent. Yes, it transcribes. But it also cross-references against history while the visit is happening. When a patient says "I've never been on metformin," and the chart shows two years of metformin from 2023, the visit agent quietly flags the contradiction in the practitioner's view. The clinician decides what to do about it. That's the part the scribe-only world cannot reach.

4. The post-visit agent. Drafts the chart note, drafts the referral letter, queues the follow-up appointment, runs an insurance eligibility check on the new procedure, and surfaces the lab orders for the clinician to sign. Five small jobs that today consume the clinician's evening, handed off to a coordinated set of small workers.

5. The re-engagement agent. Notices when a patient missed a follow-up, sends a multi-channel nudge (SMS, email, portal), rebooks them, and updates the chart. This one is quiet but compounding. Every solo and small-clinic operator I've talked to in the past year describes no-shows and follow-up drift as one of their biggest unrecoverable revenue leaks — the exact percentage varies by specialty, but the pattern is consistent. The re-engagement agent is the one that pays for the rest.

The difference between an AI scribe and this is not "smarter." It is a different unit of automation. A documentation function versus a workflow chain. A note surface versus the whole journey.

Why the existing vendors can't make the jump

Not because they don't want to. The pattern I keep observing — across vendor demos, public roadmaps, integration docs, and conversations with engineers I've met at industry events over the past year — is that the AI scribe behaves like a documentation service attached to a non-AI core. It ingests audio or note context, calls a model, and writes a note back into the chart. From the outside, I have not found public evidence that these scribe features share durable workflow state with scheduling, billing, intake, and follow-up in the way a native agent network would need to. Four ponds, not a river. I'd love to be proven wrong on any of this; I'm describing what's externally observable, not insider knowledge.

Agentic AI requires the river. An intake agent that doesn't write structured fields into the chart is useless. A re-engagement agent that can't read scheduling and rebook through it is decorative. A post-visit agent that can't see the billing eligibility check is a chatbot drafting referral letters in a vacuum.

This is an architecture question, not a feature roadmap question. You cannot ship five coordinated agents on top of a system that was designed in 2010 to be a relational database with a calendar and an invoicing module. You can ship one — the scribe — because the scribe only needs to write a note back to one table. The chain needs cross-module state, tool access, and a way for one agent to call another. That requires re-plumbing the whole platform. I have not seen public evidence that any established allied-health vendor has committed to a rebuild of that scope, and the commercial reasons are obvious: re-architecting a profitable platform is not how you make next quarter's number. But not doing it is how you become the floppy disk.

What the architecture looks like when it's built right

This is the part where I should be honest. Oli Health is built on the bet I just described, and the structural choices show up at every layer of the platform.

We use the Model Context Protocol as the substrate for agent tool access — MCP gives us schema-defined tools each agent can call, covering chart, intake, scheduling, billing, and uploads. The auditing — every tool call logged with the requesting agent, the patient context, the inputs, and the result — is Oli's own implementation on top of that protocol. We use the Vercel AI SDK 6 agent abstraction so each agent is a small, replaceable worker with a clear job, not a 2,000-line megaprompt that a single model has to keep straight. And the chart, intake, AI patient overview, and billing all sit on a single shared state, so when the intake agent writes a new medication, the visit agent already sees it before the clinician walks into the room. That's the part the four-ponds vendors structurally cannot replicate without a multi-year rebuild.

What this gets the clinician, in plain terms: an intake the patient finishes in their kitchen the night before, a chart that's already populated when the visit starts, a draft note ready before the patient leaves the room, follow-ups that schedule themselves, and a no-show recovery loop running quietly in the background — all on a single $19.95/month plan, no AI add-on SKU, no per-agent surcharge. The architecture doesn't just enable the chain; it lets us price it as one product instead of seven.

I'm not going to pretend every agent on that list is fully live today. The intake agent is. The post-visit drafting agent is. The pre-visit brief is shipping. The re-engagement agent is in private beta with three clinics in Austin and Mississauga, and the visit-time contradiction agent is still in the lab. The architecture is agentic-ready; the agents are landing one by one. Anyone telling you they have all five running in production right now is selling you the demo.

The point of describing this isn't to pitch the product. It's to show that the architecture choice is the choice. You either built the platform on a single shared state with agent-callable tools, or you didn't. If you didn't, you can ship the scribe. You cannot ship the chain. That's the structural fork in the road for the next two years of this category, and it's why a clinic shopping for an EHR in 2026 should be asking about architecture, not about which add-on shipped most recently.

For more on what agent-coordinated documentation looks like in day-to-day use, the post on AI clinical notes and documentation automation walks through the practitioner-side workflow. And the comparison of traditional vs. AI-first practice management maps the broader architectural divide that this article is one slice of.

Where this lands by the end of 2027

A prediction. By December 2027, "AI scribe" as a product category will sound the way "internet-connected" sounded in 2008. Once-bold, then a default, then unmentionable. The vendors who treated AI as a module — a SKU bolted on for $15 to $35 a month — will compress in the market. Some will consolidate downward into white-label scribe providers for niche specialties. Some will be acquired for their book of business and quietly migrated to whoever built the agent layer first. A few will rebuild and survive.

The winners will be the platforms that treated AI as architecture from the start. They won't market themselves as "AI scribes." They will market themselves as the EHR that handles the workflow end to end while the clinician sees patients. The scribe will still be there. It will just be one of five agents, the way the spell-checker is one of fifty things your word processor does.

I have noticed something in the conversations with founders this past quarter: the smart ones already know this. They are not panicking about adding a scribe; they are quietly hiring agent-systems engineers and re-plumbing the chart. The loud ones are still demoing transcription. Watch which group is left standing in 18 months.

If you want a fuller view of how this looks in a working clinic — pricing, workflow, the parts that are live and the parts that are not — the deeper write-up on Oli Health as an AI-first EHR is the next read.

Frequently asked questions