Unscheduled Treatment Follow-Up: AI Agent vs PMS vs Staff

By Jude Lee · · Comparison

Two practice staff members reviewing an unscheduled treatment report on a computer at a clinic front desk

The revenue that’s already in your chart

This is not recall. Recall handles time-based cycles — six-month hygiene, annual physicals, med checks. Unscheduled treatment is different: a clinician looked at a specific patient, recommended something specific, and it never got on the schedule.

In a dental practice it shows up as the unscheduled treatment plan report — crowns, quadrants of SRP, an implant case that stalled after the consult. In primary care and specialty it’s ordered imaging, a referral that never converted, a procedure quote the patient never called back about. In behavioral health it’s the client authorized for twelve sessions who stopped at five.

The work is unglamorous: pull the list, figure out who’s actually worth a call today, remember what was discussed, reach out without sounding like a collections agency, and book. Most practices do this in bursts when someone has a slow afternoon — which is to say, rarely and unevenly.

Three ways to work the list

PMS recall & engagement module

PMS recall & engagement module — try this first. Dentrix, Open Dental, Eaglesoft, athenahealth practice management and most modern EHRs ship something here, and many have added AI-flavored features.

Strengths: Already inside an existing BAA. No new data path, no build, usually already paid for. If your list is homogeneous (“everyone with unscheduled hygiene-adjacent treatment in the last 90 days”), a templated campaign off a filter is honestly the smarter move.

Weaknesses: Filters, not judgment. It can’t weigh a stalled implant case against an expiring benefit. Messages read templated because they are.

Staff calling from a report

Staff calling from a report.

Strengths: A coordinator who knows the patient can read the room — cost objection vs. fear vs. scheduling conflict — and adjust in real time. Payment plan, time slot, closed in one call.

Weaknesses: It’s the first thing dropped when the phones ring. Prioritization is gut feel. Who was called and what was said lives in someone’s head or a sticky note. The list grows faster than anyone works it.

What the agent actually does — and where MCP fits

An AI agent here is not a chatbot. It’s an assistant that carries out a multi-step job: query, read, rank, draft, hand off. To do that it needs governed access to your systems, which is what the Model Context Protocol (MCP) provides — an open standard for exposing specific tools and data to an AI assistant like Claude, with you deciding exactly which operations exist. You’d typically build a small custom MCP server over your PM database or API exposing a handful of read tools (get_unscheduled_treatment, get_patient_financials, get_open_slots) and, at most, one narrowly scoped write tool (log_followup_note). We walk through the mechanics in connecting Claude to your EHR via MCP, and the permission tiers in read-only vs. approval vs. full auto.

The second building block is a skill — a packaged set of instructions that makes the agent do this job the same way every time: how to rank, what tone to use, what it may never say, when to stop and flag a human.

  1. Pull and normalize the list

    The agent queries unscheduled treatment plans plus last visit date, remaining insurance benefit, outstanding balance, and whether the patient is already scheduled for anything.

  2. Rank by realistic bookability

    A skill encodes your rules: accepted-but-unbooked ahead of presented-but-undecided; benefits expiring this plan year ahead of fresh diagnoses; anyone in active collections or with a documented do-not-contact excluded entirely.

  3. Draft the outreach, one patient at a time

    Short, specific, clinically neutral: a reminder that recommended care is outstanding and an offer to find a time. Plus a three-line call brief for the coordinator — what was recommended, when, what the patient’s stated objection was.

  4. Human reviews the queue

    Your coordinator scans the drafts — approve, edit, or kill. This is the step you do not skip. It’s also your quality signal: if approval takes longer than writing from scratch, the skill needs work.

  5. Book and log

    Approved messages send from your existing patient communication platform. Outcomes get logged back so the same patient isn’t hit three weeks later with the same message.
The agent’s job is to make sure the right twenty patients are on someone’s screen before lunch. Closing is still a human skill.

The compliance line most people miss

Follow-up about recommended treatment sits close to a real regulatory boundary. Under the HIPAA Privacy Rule, communications about an individual’s own treatment are generally permitted, while communications that encourage the purchase of a product or service can meet the definition of marketing at 45 CFR 164.501 and require patient authorization. Start at the HHS Office for Civil Rights HIPAA Privacy Rule hub and read the regulatory text rather than trusting a summary. “Dr. Lee recommended a crown on #14; want to find a time?” reads very differently from “Limited-time whitening offer.” Keep your agent firmly on the first side, and have counsel review your templates.

Two more guardrails. First, minimum necessary: HHS OCR’s minimum necessary standard should shape your MCP server’s tool design — expose the fields the job needs, not the whole chart. Second, texting and automated calling are governed by FCC rules implementing the Telephone Consumer Protection Act, which include specific treatment for healthcare-related calls and messages at 47 CFR § 64.1200; confirm the current text, your consent capture, and your opt-out handling with counsel before scaling outbound volume. Anything touching PHI needs a signed BAA with every vendor in the path, including the AI provider — see our HIPAA-compliant AI workflow notes.

Modeling the money without making numbers up

We have no benchmark data for your practice, and anyone quoting a universal conversion rate for unscheduled treatment is guessing. Build the model yourself:

rows × avg case value × conversion %
Recoverable revenue in the list (fill in your own three numbers)
hours/week × loaded hourly rate
Staff time currently spent working the list
build + license + review time
True cost of the agent option, including human approval minutes

Pull your actual unscheduled treatment report. Count the rows, average the case values, and pick a deliberately conservative conversion assumption — then ask whether the gap between that number and your current run rate justifies anything more than a templated campaign. Note the second-order effect: hours a coordinator gets back from list-wrangling go into scheduling and case presentation, which is where revenue actually gets captured.

The questions office managers keep asking

Is there one best AI for a medical practice? No, and the question is the problem. “Best” is per job — scribing, phone coverage, denial triage, and follow-up lists have different winners. Our job-by-job buyer’s map lays this out; this workflow in particular often doesn’t need AI at all.

Is there a medical version of ChatGPT? Sort of, but they’re clinical tools, not office tools. OpenEvidence and Glass Health target clinician-facing evidence lookup and documentation support, Ada Health sits in consumer symptom assessment, and Hippocratic AI builds patient-facing voice agents for non-diagnostic tasks. Product positioning here changes fast — categorizations last verified January 2026, so check each vendor’s own documentation and applicable FDA classification before anything touches patient care. For admin work, general assistants like Claude under a BAA plus your own data connections are usually the more relevant path, as we compare in ChatGPT vs. Claude vs. EHR AI.

Can I do this for free? Partially. A free AI assistant can help you design the ranking rules, write message templates, and draft the outcome-code taxonomy — none of which requires uploading PHI. The moment patient data enters the workflow, free consumer tiers are off the table.

Which front-office roles survive this? Our opinion, not a forecast: the ones involving judgment, persuasion, and accountability. Nobody’s agent talks a nervous patient into a root canal or negotiates a payment plan. Pulling lists and typing the same message 200 times is the genuinely automatable part.

How to decide

Go with your PMS module if your list is small or uniform and you haven’t yet exhausted templated campaigns. Go with staff plus a disciplined weekly block if case values are high and personal relationships drive acceptance — a custom build won’t beat a good treatment coordinator on a 40-row list. Go custom-agent when the list runs to hundreds of rows, prioritization genuinely requires reading chart and financial context together, and you already have a clean way to reach your PM data. Start read-only, keep a human approving every outbound message for at least the first several weeks, and watch the approval edits — if your coordinator rewrites most drafts, the skill isn’t done yet.

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