Transitional Care Management: Alerts vs AI Agent vs Staff

By Jude Lee · · Workflow

Office manager and nurse reviewing a hospital discharge follow-up list on a monitor at a small medical practice front desk

The 30-day window most practices lose by default

A patient on your panel gets discharged from a hospital on Thursday afternoon. Someone at your practice needs to make interactive contact quickly, get a visit on the books inside a tight window, reconcile medications, and document enough to bill the transitional care management (TCM) codes. If that slips, you still owe the patient good care — you just may not get paid for the coordination, and readmission risk goes up.

The timing rules are specific. Per the CMS Transitional Care Management Services MLN booklet, TCM requires interactive contact with the patient or caregiver within two business days of discharge, a face-to-face visit within 7 days (high medical complexity, 99496) or 14 days (moderate complexity, 99495), medication reconciliation no later than the date of that visit, and a 30-day service period.

Importantly, a missed call is not automatically a lost claim. The same CMS booklet describes handling for unsuccessful contact: if you make two or more separate attempts in a timely manner, document them in the medical record, and meet all other TCM requirements, the service may still be reported. That single provision changes how you design the workflow — the system needs to log attempts, not just log successes. CMS has adjusted TCM policy over the years, including what may be billed concurrently, so confirm the current-year rules in the CMS booklet and with your Medicare Administrative Contractor rather than relying on a blog post, including this one.

The other half of the picture: CMS’s Interoperability and Patient Access final rule added Conditions of Participation requiring hospitals with certain EHR capabilities to send admission, discharge, and transfer (ADT) notifications to the patient’s established primary care practitioner or practice. For many practices, the signal already arrives. The failure is downstream: it lands in a portal, an interface inbox, a fax queue, or a direct message nobody opens until Monday.

2 business days
TCM interactive contact window — verify current CMS guidance
CMS MLN Transitional Care Management Services booklet
7 or 14 days
Face-to-face visit window by complexity (99496 / 99495) — verify current CMS guidance
CMS MLN Transitional Care Management Services booklet
30 days
TCM service period starting at discharge — verify current CMS guidance
CMS MLN Transitional Care Management Services booklet

Three ways practices actually run post-discharge follow-up

Alerts + EHR worklist
Cheapest path, no new vendor, no PHI leaves your stack. Output is a pile of events with no prioritization, no chart context, and no draft work product. Capture depends entirely on whoever opens the queue — and on them opening it on the right day. Fails quietly: an unworked queue looks identical to an empty one.
AI agent on top of the same feed
Reads the same event, resolves identity, pulls relevant chart context, flags complexity signals for the clinician to confirm, drafts the contact script and scheduling offer, and tracks the 2-day and 7/14-day clocks. Costs money, needs a BAA, needs weekly oversight, and introduces a new class of errors (below).

What the agent actually does, step by step

  1. Ingest the discharge event

    Watch the ADT feed, HIE notification, direct message, or fax inbox — and deduplicate across them. Faxed discharge summaries are still common; if that’s your reality, the parsing problem is the same one described in medical fax automation.

  2. Match and qualify

    Confirm the patient is active on your panel, identify the attributed provider, check whether a TCM service period is already open, and read the patient’s recorded preferred language and communication needs. Low-confidence matches go to a human, not to the outreach queue.

  3. Assemble context

    Pull the discharge summary, problem list, current med list, and last visit note into one short brief, with every clinical fact linked back to the source document.

  4. Draft the contact

    Produce a call script or secure message covering new meds, symptoms, transportation, and home support, plus the documentation fields the call must capture. If the patient’s preferred language isn’t one your caller speaks, route the task to a bilingual staff member or attach the interpreter-line workflow — a polished English script handed to the front desk is a capture gap and a safety gap.

  5. Work and log the attempts

    Timestamp attempt one. If unsuccessful, schedule attempt two on a different day and time of day, log it, and escalate to a named human after two documented attempts so the practice can decide whether the TCM criteria are still met. Attempt logs belong in the chart, not in the agent’s memory.

  6. Offer and book

    Propose real open slots inside the 7- or 14-day window with the right provider and visit length, re-checking availability at the moment of the call. Booking should be gated by your write-access policy — see the read-only vs approval vs full-auto breakdown.

  7. Track the clocks and prep the packet

    Maintain the 30-day service period, flag anything at risk of missing a window, and assemble what a biller needs to decide whether the encounter supports 99495 or 99496 — leaving the complexity determination to the clinician.

Where the human stays in the loop

How this breaks in practice, based on the failure modes worth testing for before you trust any of it:

The agent’s job is to make sure the right human has the right three facts before the clock runs out — not to replace the call.

Building it: MCP, a skill, and what stays inside your walls

If you go custom, the plumbing is less exotic than it sounds. MCP (the Model Context Protocol) is an open standard for giving an AI assistant governed access to specific tools and data — one scoped connection per system, with permissions you define. A practical build is a small custom MCP server exposing exactly four or five operations: find_patient, get_recent_encounters, list_open_slots, log_contact_attempt, and nothing else. Minimum necessary, in the HHS Office for Civil Rights sense, expressed in code. Our guide to connecting Claude to an EHR via MCP walks the pattern.

A skill is the second piece: a packaged, versioned instruction set that makes the agent run your TCM process the same way every time — which fields to capture, what counts as low-confidence, when to escalate after a failed attempt, what the packet looks like. The value of a skill is consistency and auditability, not intelligence.

On PHI: any vendor touching discharge data needs a Business Associate Agreement, and consumer AI tools without one are out. Confirm your configuration with your privacy officer or counsel before a single record moves.

Build, buy, or just fix the worklist

Be honest about volume. If you see a handful of discharges a month, an agent is overkill — put a recurring calendar block on one person and a shared checklist, and spend your automation budget elsewhere. If your EHR or care-management module already runs a TCM workflow, evaluate that first; layering on what you own is usually cheaper than replacing it. And if you already pay a chronic care management vendor, ask whether TCM is in scope before buying anything new.

Custom builds earn their keep when the bottleneck is your specific mess: four intake channels, a PM system with a weak API, a panel attribution rule nobody else models correctly.

A revenue model you fill in yourself

Don’t take a vendor’s dollar figure. Build your own:

Captured revenue upside = A × (target capture − B) × C. Labor shift = A × (D − D′) ÷ 60 × E, where D′ is minutes after automation. Then subtract the platform or build cost and the oversight time someone spends reviewing agent output every week — that line is real, and it never goes to zero. If the result doesn’t clear your cost with room to spare under conservative assumptions, the honest answer is “not yet.”

Choosing a tool for this job

There is no single best AI for a medical practice, and anyone who names one is selling it. The right answer is job-by-job: scribing, scheduling, denial triage, and discharge follow-up have different data, different risk, and different buyers — our job-by-job buyer’s map is organized that way deliberately. For TCM specifically, ask every vendor something narrow: what does your system do in the 48 hours after a discharge event, and what does it hand my staff?

You’ll also see people cite a “30% rule” for AI — roughly, “expect a third of cases to need human handling.” We’re not aware of an authoritative standard behind it; treat it as rhetoric, not a benchmark. Before signing anything, get a written list of which TCM elements the system documents automatically versus which still require staff entry, confirm the BAA covers the discharge feed itself, and run one month of real discharges end to end while counting exceptions. That number — yours, measured — is the only one worth managing to.

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