AI Agents for Care Coordination

Because nobody is measured on closing them, and because the closing evidence usually lives in someone else's system.

The scale of the problem is well documented. A 2018 analysis in the Journal of General Internal Medicine examined 103,737 primary care referral scheduling attempts in one large health system over twelve months. Only 34.8 percent resulted in a documented completed appointment, and 38.9 percent had no appointment date recorded at all. The second figure is the more revealing one: in nearly four referrals in ten, the system could not tell you whether the patient was seen, because the tracking never existed.

The break points are predictable:

1. Order to schedule. The referral is placed and the patient is told to call. Many do not, and nobody knows which.
2. Schedule to attend. The appointment is six weeks out, circumstances change, and the no show is recorded at the specialist end where the referring clinician never sees it.
3. Attend to report back. The consultation happens and the note goes to a fax number that reaches a scanning queue. The referring clinician learns the outcome from the patient.
4. Authorisation. The referral stalls waiting for a plan decision, which is a prior authorisation problem wearing a care coordination costume.

An agent helps at every point, but only if it can see both ends. Inside one organisation on one EHR this is tractable workflow automation. Across organisations it becomes an interoperability problem.

The transitional care management workflow is unusually well specified, which makes it one of the better automation targets in this whole category. Medicare's transitional care management codes require interactive contact with the patient or caregiver within two business days of discharge and a face to face visit within a defined window, and both of those are checkable events rather than judgements.

The evidence that it matters is reasonable. A 2018 JAMA Internal Medicine study of Medicare beneficiaries found that discharges where transitional care management was billed had adjusted costs of 3,033 US dollars in days 31 to 60 after discharge against 3,358 dollars where it was not, and adjusted mortality of 1.0 percent against 1.6 percent. This is observational work and the groups differ in ways adjustment cannot fully remove, so read it as a strong association rather than a causal estimate.

The gap is uptake. The same study found the service was billed for 3.1 percent of eligible discharges in 2013, 5.5 percent in 2014 and 7.0 percent in 2015. A paid, guideline backed, evidence supported service was being delivered to roughly one discharge in fourteen, and the reason given by practices is almost always the same: the two day contact window is operationally brutal, and nobody has the staff to make the calls.

That is a genuinely good fit for an agent. Placing a bounded outbound call within a defined window, asking a scripted set of questions, escalating anything abnormal to a nurse and booking the follow up visit is exactly the shape of task current voice and messaging agents handle. The clinical judgement stays with the nurse who receives the escalation. Anything that reads like triage rather than logistics belongs in the clinical inbox triage workflow instead, with the safety controls that page describes.

It changes conversion on the patients who were already reachable, and that is a smaller claim than most vendors make but still a real one.

The mechanism is not persuasion. It is volume and timing. A registry of 4,000 overdue patients does not get called, because calling it would take one member of staff several months, by which time the list has changed. An agent works the whole list in a week and hands the ones who say yes straight to a booking. Conversion per contact may be no better than a human's. The number of contacts is an order of magnitude higher.

Three things separate outreach that works from outreach that annoys people.

• Book, do not inform. A message that ends in an appointment slot converts. A message that ends in a phone number to call does not, and it moves the work back to your front desk.
• Suppress properly. Deceased patients, patients already scheduled, patients who opted out, patients in hospice. Every campaign that has caused reputational damage failed at suppression rather than at content.
• Cap the frequency. Once an organisation has three agents running three lists, a patient can receive nine messages in a fortnight. A single contact governor across all outbound automation is not optional, and it is usually built after the first complaint rather than before.

Measure conversion to completed service, never to message delivered or even to appointment booked. Booked and not attended is a common and flattering plateau, and it is where automated outreach programmes tend to report success while the clinical measure does not move.

This is the question that separates a care coordination programme from a marketing campaign, and it deserves to be asked before launch rather than in an annual review.

Automated outreach is cheapest and most effective through the channels the already engaged patient uses: the portal, email, an app notification. Those are precisely the channels least available to the patients whose care gaps are widest. An analysis of national screening trend data published in 2025 illustrates the underlying spread: colorectal cancer screening among insured adults ran at around 78 percent in 2023 against roughly 33 percent among the uninsured, with American Indian and Alaska Native adults well below the national average. An outreach programme that lifts portal responders lifts the top of that distribution and can widen the gap it was funded to close.

Four design choices materially change the answer.

• Lead with voice and SMS, not the portal. Portal first is the cheapest design and the least equitable one.
• Language coverage is a launch requirement, not a phase two. An English only agent produces a measurable gap in contact rates by language within a month, and it stays in your data permanently.
• Keep a human path that is easy to find. Some patients will not engage with an automated caller at all, and the fallback cannot be that they drop off the list.
• Report outcomes stratified. Contact, conversion and completion by language, insurance status and age band. Aggregate reporting hides the divergence.

Where a generative agent composes clinical information for a patient, disclosure obligations may also apply. California has legislated in this area, and we track the detail on the California AI healthcare laws page. For federally qualified health centres, where reachability is the central operational fact rather than an edge case, we set out a different starting sequence on the community health centres page.

Because no single system owns the workflow, and therefore no single vendor can own the product. Every other agent use case has a home: the scribe lives in the encounter, the coding agent lives in the claim, the phone agent lives on the line. Care coordination lives in the space between four systems that were never designed to hand off to each other.

Read down the third column and the shape of the problem is clear. Modelling is not the constraint. Two thirds of the delivery effort in every care coordination project we have scoped is data acquisition, identity matching across organisations, and deciding who is accountable for a loop that crosses a corporate boundary.

So scope should follow data availability rather than clinical priority. Start with the loops that close inside your own walls, where the EHR sees both ends, and add cross organisation loops only where you already have a working exchange path and an agreement about who chases. The integration questions are the ones set out on the Epic integration page, and the interoperability floor that decides what is available at all is covered on the ONC HTI-1 page.

Four things, and the software is the least of them.

A subscription to events. Discharge, admission, referral order, result finalised. Without an event feed the agent polls a list and is always a day late, which for the two business day transitional care window is fatal. An ADT feed from your regional health information exchange is usually the cheapest route and is frequently already available and unused.

An identity that survives the boundary. Matching your patient to the record the specialist created is solved inside one organisation and persistent across several. Budget for a manual reconciliation queue, because there will be one.

A named clinical owner for escalations. Every agent that contacts patients will occasionally hear something needing a clinician within the hour. Who receives that, on what rota, and what happens out of hours are questions to answer before launch, not after the first incident.

A business associate agreement that reaches the contact channel. Voice, SMS and messaging vendors are subprocessors handling protected health information. Retention of call recordings and transcripts needs an explicit answer too, and that ground is covered on the HIPAA and AI compliance page.

On vendors, the market splits between workflow automation platforms sitting across intake, referrals and outreach, and voice specialists that handle the contact but not the list. Neither owns the whole loop. If your gap is contact capacity rather than tracking, start from our AI phone agents for medical offices shortlist.

Pick the closing event, not the activity. Almost every disappointing care coordination programme we have reviewed was measured on contacts made, and almost every convincing one was measured on loops closed.

• Referral loop closure rate. The share of referrals with a documented completed consultation and a returned note, within a defined window. Baseline this before anything is installed, because it is usually far lower than the organisation believes and the shock is instructive.
• Transitional contact within two business days, as a share of eligible discharges, and the follow up visit rate that follows it.
• Care gap closure, measured at completed service rather than booked appointment.
• Escalation volume and time to clinician response, which is the safety metric and the one that tells you whether the human side of the design is holding.
• All of the above, stratified by language, insurance status and age.

Treat two numbers with suspicion. Reduction in readmissions is the outcome everyone wants to claim and the hardest to attribute at a single site over a single year. And messages delivered is not a result at any point.

Where this lands commercially depends on your contracts. Under fee for service, closed loops mostly generate visits and billable transitional care management. Under risk or shared savings, the value is avoided utilisation and quality measure performance, which is easier to write into a business case and harder to prove. Which of the two you are in decides whether this is a revenue project or a cost project, and that is the first thing we establish in an AI readiness audit. Earlier than that, the AI readiness assessment is a shorter way to find out whether the event feeds and ownership you need are in place.

Practices whose bottleneck is the inbound phone line rather than the outbound list. If patients cannot get through to book, outbound outreach generates demand you cannot absorb, and the visible result is a worse patient experience. Fix the line first.

Organisations without an event feed and without a realistic route to one. A care coordination agent with no discharge or referral events is a list worker, which is useful but is a much smaller project than the one being sold, and should be priced accordingly.

Anyone whose registries disagree. If the population health tool, the payer gap file and the EHR each produce a different overdue list, an agent will call patients who are not overdue and miss patients who are. Reconciling those lists is unglamorous data work that comes first, and it is often three months of it.

The pattern in all three is that the constraint sits upstream of the agent, which is why we scope care coordination as an integration and ownership project with an agent attached rather than as a product purchase. Where the prerequisites do hold, this is one of the few agent workflows with a plausible clinical outcome rather than only an administrative one, and it is worth sequencing properly in a deployment roadmap.