Agentic Practice Management: Switch or Layer AI on Top?

By Jude Lee · · News

Practice manager and biller reviewing a claims worklist on two monitors in an independent medical office

The funding wave, and what it actually signals

Several agentic-healthcare deals landed in 2026. Forus raised a $150M Series C at a reported $3B valuation to put AI agents into medical practices, as reported by HIT Consultant in September 2026 — and worth saying plainly: round sizes and valuations in coverage like this are figures reported by the company and its investors, not audited numbers. GenHealth.ai announced a $16.5M Series A to build AI agents for the medical back office, per Newswire, also reported in September 2026. Other vendors market the category directly — Athelas, which now sits under the Commure brand, is one name you’ll encounter — so check current product naming and claimed capabilities on the vendor’s own documentation before a demo, because branding in this space has moved fast.

What all of this signals is capital conviction, not proven outcomes in a five-provider practice. Funding rounds measure investor belief. They are not evidence that a given agent will post your ERAs correctly or stop your Medicare Advantage denials. Treat the wave as a reason to look seriously at agents — and as a reason to be a demanding buyer, because a lot of money is now chasing your signature.

$150M
Forus Series C, as reported
HIT Consultant, September 2026
$16.5M
GenHealth.ai Series A, as reported
Newswire, September 2026

What “agentic” means, without the marketing

An AI agent differs from a chatbot in one concrete way: it takes multi-step actions in your systems. A chatbot answers “what’s this denial code?” An agent pulls the remit, matches it to the claim, checks the payer policy, drafts the corrected claim, and drops it in a queue for your biller to approve. Same model underneath — the difference is tools and permissions.

Two building blocks matter for independent practices. Skills are packaged, reusable instructions that make an assistant do a job the same way every time — your appeal letter format, your specific payer quirks, your intake checklist. MCP (Model Context Protocol) is an open standard for giving an AI governed, auditable access to a system’s data and actions; it’s what lets an assistant like Claude read your scheduling or billing data through a defined, minimum-necessary interface instead of someone pasting screenshots into a chat window. We walk through that build pattern in connecting Claude to your EHR via MCP.

The interesting question isn’t whether agents work. It’s who owns the permissions, the audit trail, and the switching cost.

Three paths, compared

Path 1: Replace your PM with an agentic platform. The genuine case for this is architectural, not just “my current system is bad.” One vendor means one BAA, one audit trail covering both the record and the agent’s actions, and no integration drift — nothing breaks because a third party changed an API version, and you don’t wake up with orphaned agent logic that pointed at an endpoint that no longer exists. Agents wired natively to the data can also see context an external integration never gets exposed to. The cost side is real: a data migration, a revenue-cycle transition period, staff retraining, and a deeper dependency on one roadmap. Migrations are where practices lose money quietly — in AR that ages out during cutover and in hours nobody budgeted.

Path 2: Layer agents on the PM you already run. If you’re on athenahealth practice management, Epic, eClinicalWorks, Dentrix, or a behavioral-health platform like SimplePractice, you may be able to keep the system of record and add agents via the vendor’s API or a custom MCP server scoped to specific workflows. One caveat to confirm early: at several of these vendors, write-scope API access is gated — partner-program-only, subject to review, or separately priced — so don’t walk into a demo assuming the layer path is available to you. Ask in writing which endpoints allow writes, under what approval process, and at what cost. More depth in EHR AI agents vs point tools vs custom builds.

Path 3: Buy one point tool — or skip AI entirely. Sometimes the honest answer is that your problem is a rule, not a judgment call. Eligibility checks at a fixed cadence, reminder cadences, claim edits with deterministic logic — those are better served by plain automation than an LLM. Cheaper, more predictable, easier to audit.

Switch to an agentic PM platform
Best when your current PM is genuinely failing you, or when you value a single vendor, a single BAA, and one unified audit trail over flexibility. Agents are tightly integrated and you don’t maintain the plumbing or absorb breakage when someone else’s API changes. Cost: migration, retraining, and a deeper dependency on one vendor’s roadmap. Ask what happens to your data and your agent configuration if you leave.
Layer agents on your current PM
Best when the PM works but specific workflows don’t. You keep your system of record, scope agents to one or two jobs, and expand only where results hold. Cost: you own integration, permissions design, and monitoring — or you pay someone to. Confirm the API actually supports the write actions you need, on your contract tier.

Modeling the decision with your own numbers

Don’t accept a vendor’s ROI slide. Build your own, with assumptions you can defend.

Recovered hours. (hours per week on the workflow) × (loaded hourly rate) × 52 × (share you believe an agent handles without rework). Worked illustration, with every input labeled as an assumption you own, not a benchmark from us: assume 12 hours/week of eligibility calls, assume a $32 loaded rate, assume 50% of that volume is genuinely deflected. That’s 12 × 32 × 52 × 0.5 ≈ $10,000/year of recovered time — a number that exists only because you picked those three inputs. Change the deflection assumption to 25% and it halves. Run it both ways before you sign anything.

Captured revenue. (claims or appointments currently lost per month) × (average value) × (share you expect to recover). Backfilled cancellations and reworked denials belong here.

Costs. Platform fees + implementation + migration hours + the ongoing time someone spends reviewing agent output. That last line is permanent, and it’s real work, because agents fail in specific, detectable ways:

If the honest version of that math is close, don’t switch platforms. Layer instead. Fuller framework in practice automation ROI and a HIPAA vendor checklist.

Where agents tend to hold up in real practices

This is our editorial read, not a survey finding: the deployments we see holding up cluster in administrative work with a clear right answer and a reviewable output — ambient documentation, eligibility and benefits verification, claim scrubbing and denial triage, inbound call handling and scheduling, records requests, prior-authorization packet assembly, chart prep. Clinical decision support is a different regulatory category; the FDA has published guidance on when clinical decision support software is regulated as a device, and that’s a conversation for your clinical leadership, not your ops budget.

The pattern that separates working deployments from stalled pilots is scope. One workflow, one measurable output, a human approving anything that touches money, the medical record, or a patient’s care. See read-only vs approval vs full auto for EHR access.

What this does to front-office and billing roles

The most exposed jobs are pure transcription and lookup: rekeying intake forms, calling payer IVRs, posting simple ERAs. The ones that hold up require judgment, relationship, and accountability — the biller who knows which payer rep to escalate to, the manager who reads a schedule and knows which patients will no-show. In our view, the practical shift for most practices isn’t headcount reduction; it’s the same staff spending less time on queue work and more on collections, patient access, and the exceptions agents hand back.

  1. Name one workflow and its current cost

    Pick the queue that’s actually hurting — AR over 90 days, unfilled slots, intake rework. Measure baseline hours and dollars for two weeks before you talk to anyone.

  2. Ask whether it's a rule or a judgment

    Deterministic and stable? Automate with rules. Unstructured inputs and exceptions? That’s agent territory.

  3. Test the layer option first

    Ask your current PM/EHR vendor what agentic features ship this year, and what their API or MCP support actually allows on your tier. A migration you don’t need is the most expensive option on the table.

  4. Run a demo on your own messy data

    Bring real (de-identified, or under an executed BAA) examples — your ugliest denials, your worst fax. Generic demos prove nothing.

  5. Lock down the compliance layer

    Executed BAA, minimum-necessary access scoping, audit logging of every agent action, and a documented human-approval step. Verify HIPAA specifics against HHS Office for Civil Rights guidance, and have your privacy officer or counsel review before PHI moves.

  6. Write the success metric and the kill criterion

    Before launch, name one number (clean-claim rate, days in AR, hours on the queue), the target, and the date you check. Then write the kill criterion: the result at that date that ends the pilot. Most practices skip this and let an underperforming tool coast on sunk cost.

  7. Define the exit

    Data export format, who owns the agent configuration and skills you wrote, and contract length. Ask before signing, not after.

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