Diagnostic AI vs Scribes vs Admin Agents: Where to Start

By Jude Lee · · Comparison

Office manager and clinician reviewing schedule and screens at the front desk of an independent practice

Three different products, one confusing label

When a rep says “we do AI for practices,” they could mean any of three fundamentally different things. Confusing them is how practices end up with a pilot that stalls in compliance review, or a tool that clinicians quietly stop opening.

Tier one is clinical decision support (CDS) — software that helps a clinician reason about a specific patient: differential diagnosis, workup suggestions, drug interaction flags. Glass Health is the tool people usually name here; it markets itself to clinicians for differential and plan drafting. This is the tier where regulators pay attention.

Tier two is ambient documentation — the scribe category. Microphone in the room, structured note out the other end. In my read of the market it’s the most-adopted clinical AI in independent practices right now, and at least one vendor is pushing past the note itself: FDB (First Databank) announced an AI-powered prescribing agent that extends ambient listening into medication actions, per its own announcement on PR Newswire. Whether that becomes the category norm is still an open question — treat it as one vendor’s direction, not a settled trend.

Tier three is administrative and agentic AI — multi-step software that takes actions in your practice management system, scheduling, phone, fax, and billing stack. Hippocratic AI is a well-known name aimed at patient-facing outreach, and it markets itself around non-diagnostic patient-facing tasks (verify current positioning directly with the vendor). General assistants like Claude, connected to your systems, live here too.

Where the FDA line actually falls

The FDA published final guidance titled Clinical Decision Support Software in September 2022, which describes how the agency interprets the CDS exclusion added to the Federal Food, Drug, and Cosmetic Act by the 21st Century Cures Act. Broadly, the criteria turn on whether the software merely supports a clinician who can independently review the basis for its recommendation, versus whether it drives a decision the clinician can’t reasonably second-guess. Read it directly: Clinical Decision Support Software guidance. The FDA also maintains a public list of AI-enabled medical devices it has authorized. Check both before you evaluate anything in tier one — don’t take a vendor’s summary of its own regulatory status at face value, and confirm clinical use questions with your medical director or a qualified clinical professional.

Which brings us to the search query. People type “best free AI for medical diagnosis” and the honest operations answer is: there isn’t one you should route patients through. Free consumer chatbots aren’t cleared as diagnostic devices, they carry no business associate agreement by default, and a wrong answer lands on your clinician’s license, not the vendor’s. Clinicians use general assistants as a reference and a thinking partner — that’s real and widespread — but that’s a personal-productivity habit, not a practice workflow you can standardize, audit, or defend.

What kind of AI ChatGPT and Claude actually are

The technical answer to “what type of AI is ChatGPT classified as” is: a general-purpose large language model — a generative, transformer-based system trained on broad text, not a domain-specific medical device. Same category as Claude, Gemini, and the rest. Regulatorily, general-purpose assistants are not cleared medical devices, and vendors’ usage policies typically warn against relying on them for diagnosis.

Operationally, the relevant question for an office manager isn’t classification, it’s coverage: does the vendor sign a business associate agreement, and does your paid tier fall under it? HHS Office for Civil Rights guidance is clear that a vendor creating, receiving, maintaining, or transmitting PHI on your behalf is a business associate and needs an agreement. We walk through the specifics in our piece on whether ChatGPT is HIPAA compliant for practices.

Every tier of medical AI eventually needs a human signature. The only real question is how much chart prep, prior-auth follow-up, and fax chasing happens unpaid before that signature.

Where each tier breaks — before you decide what to buy

CDS breaks on trust and workflow fit: a suggestion that arrives at the wrong moment gets dismissed, and dismissal becomes habit. Ambient scribes break on accents, crosstalk, complex multi-problem visits, and specialty vocabulary — and they break silently, which is worse than an obvious error. Admin agents break on ambiguity: a patient message that’s part scheduling request and part clinical complaint, a payer portal that changed overnight, a referral fax with three pages missing.

Let those failure modes drive the purchase. Ask every vendor how its product fails, not just how it succeeds, and make sure a named human has the authority — and the time — to say “no, redo that.”

Buying a category versus building one

Buy the tool (clinical + documentation tiers)
Ambient scribes and CDS require clinical training data, model validation, specialty tuning, and in some cases regulatory clearance. No independent practice should attempt to build these. Buy from a vendor that will sign a BAA, show you an audit trail, and tell you plainly what its regulatory status is. Evaluate on note quality in your specialty, EHR write-back, and how easily a clinician can edit and reject.
Build or assemble (administrative tier)
Admin work is where your practice is idiosyncratic — your payer mix, your fee schedule, your recall cadence, your fax queue. Off-the-shelf tools cover the common cases; the exceptions are where custom agents earn their keep: an assistant connected to your practice management data through MCP, with narrowly scoped, minimum-necessary access and a human approving anything that touches a patient or a claim. See our build guide for connecting Claude to an EHR via MCP.

MCP — the Model Context Protocol — is an open standard for giving an AI assistant governed access to a specific set of tools and data. In practice, a custom MCP server for a practice exposes a handful of deliberately narrow operations (look_up_appointment, get_eligibility_status, draft_referral_packet) rather than raw database access. Paired with skills — packaged, reusable instructions that make the agent do a job the same way every time — that’s how you get consistency instead of a clever assistant improvising differently each Tuesday.

Not everything deserves an agent. If the task is deterministic and high-volume, a rules engine or your PM system’s built-in automation is cheaper, faster, and easier to audit. Reach for an agent only when the work genuinely requires judgment across messy inputs.

A defensible sequence for a small practice

  1. Fix one administrative bottleneck first

    Pick the job that generates the most phone tag or the most rework — usually eligibility verification, referral intake, or recall. Automate it with the least clever tool that works. Measure the before state manually for two weeks so you have your own numbers rather than a vendor’s.

  2. Add documentation AI only if clinicians ask for it

    Scribes succeed on clinician enthusiasm and fail on mandate. Pilot with two willing providers, in one specialty, for a full month.

  3. Wire your own systems together

    Once one or two AI tools are earning their keep in isolation, connect an assistant to your scheduling and PM data through MCP under a BAA — read-mostly at first, write actions gated behind human approval.

  4. Treat clinical decision support as a governance project, not a purchase

    Who reviews the output? What’s documented? What’s the regulatory status? If you can’t answer those in writing, you’re not ready.

Modeling the payback without inventing numbers

Don’t accept a vendor’s ROI slide. Build your own from inputs you can actually observe:

hours × rate
Recovered staff time — your measured minutes per task, your loaded hourly rate
visits × margin
Captured revenue — fewer abandoned calls and unfilled cancellation slots
claims × rework cost
Errors avoided — denials prevented at the front end

Here’s the shape of the calculation, using illustrative placeholders only — substitute your own measured figures. Assume verification takes 11 minutes per new patient and you see 40 new patients a month: 11 × 40 = 440 minutes, or about 7.3 hours. Multiply by your loaded hourly rate R and the monthly recovered-time figure is 7.3 × R. If automation only removes half the task, halve it.

The leg that gets forgotten: recovered hours only become money if they’re reallocated. Seven hours a month returned to a front-desk coordinator is real value only if it goes to balance follow-up, recall outreach, or capturing the same-day slot — not to absorbing an already-overloaded day. Decide where that time gets pointed before you sign anything.

If you’re deciding what to buy first across the whole practice rather than within one tier, our job-by-job buyer’s map sorts the options by the specific work they do rather than by vendor category. As of 2026 this landscape is moving monthly — re-verify vendor claims, BAA terms, and regulatory status at the point of purchase, not from an article.

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