Scribe Vendors Add AI Agents: Buy, Layer, or Build

By Jude Lee · · Comparison

Physician and office manager reviewing post-visit documentation on a laptop at a clinic workstation

The ambient scribe quietly became an agent platform

Several moves reported in the healthcare IT trade press in late 2025 point the same direction. Fierce Healthcare reported that Heidi launched a platform extending its ambient documentation product into AI agents that take on clinical administrative tasks. Healthcare IT Today covered Hello Patient’s acquisition of Converse Health to stretch AI agents across the full patient conversation. HIT Consultant wrote up a Capstone Partners report describing a rise in AI-enabled healthcare IT M&A. I’m recalling these as reported in late 2025 rather than citing dated links — check each publisher’s archive before you quote any of them in a board memo or a budget request.

The direction of travel is the signal, not any single deal, and reading it as “point tools are racing to become platforms” is my interpretation of vendor announcements, not a measured trend. For an independent practice, it creates a specific decision. The scribe you already bought — the one your physicians actually like — is about to offer to do the referral letters, the order queue, the prior-auth packet, and the patient instructions. Is that the cheapest win available to you, or scope creep into workflows that belong somewhere else?

What an admin agent actually is, honestly

A scribe is a transcription-and-structuring tool: audio in, structured note out, clinician edits and signs. An agent is different in kind. It takes a goal, decides on a sequence of steps, calls tools (read a chart, search prior notes, draft a letter, create a task, send a fax, check a portal), and either executes or stops for approval. The useful mental model is a very fast, very literal new hire who never gets tired and never asks a clarifying question unless you’ve explicitly taught it to.

Agents are strong at drafting, summarizing, assembling packets, and pattern-matching against rules you’ve written down. They are weak at judgment under ambiguity, at payer portals that change layout, and at anything where being wrong is expensive and hard to detect later. That last category deserves a concrete example, because it’s where practices get hurt: an agent drafts a referral letter and faxes it to the PCP number in the chart, which went stale eighteen months ago, so the letter lands nowhere and nobody notices until the patient asks why their doctor never heard back. Or an agent queues an imaging order against the wrong laterality — left knee instead of right — because the note discussed both. Both outputs look correct at a glance. Both surface weeks later, in a complaint or a denial. Any step with clinical consequence should be confirmed by a qualified clinician before it executes.

If a workflow is fully deterministic — “if CPT in list, attach form X” — a rules engine is cheaper and more auditable than an agent, a tradeoff worth reading in detail in AI agents vs RPA vs rules.

A scribe that’s wrong produces a note a clinician catches in ten seconds. An agent that’s wrong produces an action somebody has to undo.

Three paths for the twenty minutes after a visit

Take a concrete workflow: a specialist visit ends. Someone has to send a referral letter to the PCP, queue imaging orders, start a prior auth for the imaging, confirm the charge is captured, and get after-visit instructions to the patient in their language. Today that’s split across a medical assistant, the biller, and the front desk.

Scribe-vendor agents
Closest to the clinical context — the agent already has the encounter it documented, so referral letters and after-visit instructions are near-free. Fast to turn on; your clinicians are already logged in. Weaknesses: write access back into the EHR is usually limited to what the integration permits, revenue-cycle depth is shallow compared to a PM vendor, and you’re deepening dependence on a company whose roadmap you don’t control. Ask pointed questions about what happens to the admin agents if the vendor is acquired.
EHR / PM vendor agents
Native write access, native audit trail, no extra integration to maintain, usually one BAA you already signed. Best for anything that must land inside the chart, the schedule, or the claim. Weaknesses: release timelines are not yours, configurability is often shallow, and “AI feature” sometimes means a template with a model behind it. Evaluate these claims carefully using the questions in our EHR vendor AI agent evaluation guide.

The third path is a custom agent. Here you connect an AI assistant (Claude, or another model your BAA covers) to your own systems and expose a small, deliberate set of tools: get_encounter_summary, find_pcp_fax, draft_referral_letter, create_task. MCP — the Model Context Protocol, an open standard for giving an AI governed access to specific tools and data — is one way to wire that up, and it’s convenient when you want one governed tool layer several assistants can share. It is not the only way: a vendor’s REST API is simpler when you only need one system, HL7/FHIR interfaces are the right call when you’re moving standard clinical data between clinical systems, and an iPaaS is often cheaper when the job is scheduled record-shuffling rather than agent reasoning. The MCP pattern is walked through in connecting Claude to your EHR via MCP.

Custom deserves the same weaknesses list as the vendor paths. You own maintenance forever: when your EHR changes an API or a payer portal changes a field, nobody ships a patch for you. You carry key-person risk — the person who built it understands it, and that understanding is rarely written down. And the integration contract is yours, which means version upgrades, credential rotation, and log retention become a line item in somebody’s week. Custom earns its keep when a workflow crosses three or more systems no single vendor owns, when your rules are genuinely idiosyncratic (a payer mix, a referral network, a state requirement), or when you’ve bought two tools that each do most of the job. It’s the wrong answer when an existing vendor ships the same thing in two quarters.

Guardrails that aren’t optional

Everything above touches PHI. HHS Office for Civil Rights guidance on HIPAA is the primary source to work from — and a vendor “supporting HIPAA” is not the same as a signed Business Associate Agreement covering the specific AI feature. Ask whether the admin agents fall under the BAA you already signed for the scribe, or whether they’re a separate product with separate terms. Apply the minimum-necessary principle at the tool layer, not the policy layer: an agent drafting a referral letter does not need read access to the full chart history. Keep a human approving anything that writes to a chart, submits a claim, or messages a patient — see the permission models in read-only vs approval vs full auto. Have your privacy officer or counsel review any new data flow before it goes live.

There is no authoritative “30% rule” for agent deployment

People search for “the 30% rule for AI” expecting an industry standard. No regulator, standards body, or primary source we can point to defines one, and you should treat anyone citing it as a benchmark with suspicion. The usable version is an opinion, offered as such: assume the first deployment of any agent will handle a minority of your volume cleanly and escalate the rest, and design the escalation path first. If a workflow only pays off when the agent handles nearly everything autonomously, it’s probably the wrong first workflow.

Which roles change, and which don’t

My read is that AI compresses the typing-and-assembling parts of healthcare jobs and barely touches the deciding-and-persuading parts. Transcription, retyping between systems, and packet assembly shrink. Clinical judgment, calling a payer and arguing, de-escalating an upset patient, catching the thing that looks wrong but isn’t flagged — that work stays, and tends to absorb the freed hours. The practical risk for an office manager isn’t mass displacement; it’s a staff member whose job quietly becomes pure exception handling with no change in training or title.

Doing the math before you sign

Don’t accept a vendor’s ROI slide. Build your own with numbers you can defend — including the line practices most often forget, which is the staff time spent reviewing agent output:

hrs/wk × loaded hourly rate × 52
Annual labor value of one recovered task — fill in your own figures
Worked example, not a benchmark
(visits added/wk × avg collected per visit)
Revenue side: hours reallocated to backfilling cancellations or recall
Worked example, not a benchmark
items/wk × review minutes ÷ 60 × loaded rate × 52
QA cost: e.g. referral letters drafted per week × minutes to scan each
Worked example, not a benchmark
per-user fee + integration + maintenance + QA
True cost — subtract from the two lines above before you call it ROI
Worked example, not a benchmark

Run the QA line honestly. An agent that drafts a high volume of letters still needs someone reading them, and if review takes a few minutes each, that time comes straight off the top of whatever you thought you were saving.

How to decide in two weeks

  1. Pick one workflow, not a category

    Name the exact task, who does it now, and how many times a week. “Post-visit referral letters” — not “documentation.”

  2. Check whether your incumbents already ship it

    Email your EHR/PM rep and your scribe rep the same question. One of them may make the build decision for you.

  3. Price all three paths including review time

    Use the formulas above. Custom builds carry maintenance and key-person risk; vendor features carry roadmap risk.

  4. Run a shadow pilot

    Agent drafts, human does the task normally, compare. Two weeks of side-by-side tells you more than any demo.

  5. Decide the escalation rule before go-live

    What does the agent do when it’s unsure? If the answer is “guess,” don’t ship it.

For practices mapping several of these decisions at once, the job-by-job breakdown in our buyer’s map for AI in a practice is a reasonable starting grid — and “which AI tool is best” almost always resolves to “which job are you solving,” not to a single product name.

Not sure where to start?

Get a free automation audit: we map your scheduling, intake, insurance, billing, and patient communication and show you what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides