AI Automation in CCM

What is AI automation in CCM?

AI automation in CCM means using models and rules to draft, sort and check the work of a chronic care management program, such as call summaries, care plan updates, outreach lists and time logs, while licensed clinical staff still talk to patients, make clinical decisions and sign the record. The aim is fewer clerical minutes per patient, not fewer clinicians.

That distinction matters for money as well as safety. CCM is paid for time that people spend on care, so automation pays off by freeing staff from documentation and searching, which lets the same team serve more enrolled patients well.

Which CCM tasks can AI take on?

AI can take on the clerical and preparatory tasks that surround each patient contact. In delivery work, the best early targets are the ones where a nurse already reviews the output before it reaches the record.

• Pre-call briefs: A one-screen summary of recent visits, results, medication changes and open care plan goals before a nurse calls.
• Call notes: A draft note from the call recording or the nurse's shorthand, structured into problems, actions and follow-ups for the nurse to edit and sign.
• Care plan updates: Suggested changes to goals and interventions after a call, shown as tracked edits against the current plan.
• Outreach queues: Ranking enrolled patients by who most needs a call this week, using results, missed appointments and messages.
• Message triage: Sorting portal messages into clinical, refill, scheduling and billing queues, with urgent wording flagged for a person straight away.
• Time log checks: Spotting entries that look duplicated, overlap another service or lack a matching note before month-end.
• Patient education: Drafting plain-language material at the right reading level for a nurse to approve and send.

Each of these produces a draft or a list. None of them should write to the chart, send a clinical message or close a task without a named staff member accepting it.

Which tasks should stay with clinical staff?

Clinical judgement, patient conversations and anything that changes treatment should stay with clinical staff. The CMS model of CCM assumes people under the direction of a billing practitioner do this work, and AI should support them rather than stand in for them.

• Symptom assessment: Deciding whether a new symptom needs a same-day visit, an emergency referral or a routine follow-up.
• Medication changes: Any change to dose, timing or drug, even when the model's suggestion looks obvious.
• Consent conversations: Explaining CCM, cost sharing and the right to stop, and recording the patient's answer.
• Care plan sign-off: Accepting the comprehensive care plan and each revision as the plan of record.
• Escalations: Calls to the billing practitioner, specialists or caregivers when something is going wrong.

A useful test is to ask who would answer to the patient if the output were wrong. If the answer is a nurse or a physician, that person must see and accept the output first.

What data does an AI assistant for CCM need?

An AI assistant for CCM needs the same record a care manager reads before a call: problems, medications, recent results, visits, the current care plan and the history of past contacts. Gaps in that data show up directly as gaps or errors in the drafts.

• Clinical record: Problem list, medication list, allergies, recent labs and vitals, pulled through FHIR resources such as Condition, MedicationRequest and Observation, or from HL7 v2 feeds.
• Encounters: Recent visits, admissions and emergency department stays, with discharge summaries where they exist.
• Care plan: The current comprehensive care plan, its goals and interventions, and the date of each change.
• Contact history: Prior call notes, portal messages and outreach attempts, so drafts do not repeat questions already answered.
• Program data: Enrollment status, consent date, minutes logged this month and the assigned care manager.
• Social context: Caregiver details, language preference and known barriers such as transport, which change how outreach should happen.

Limit each model call to the minimum data the task needs. A pre-call brief needs the last few months of activity, not the full chart, and a message triage model rarely needs lab history at all. Smaller inputs cost less, leak less and are easier to audit.

How do CCM billing rules shape automation?

CCM billing rules shape automation because the codes pay for human time. CMS's MLN909188 booklet describes CPT 99490, 99439, 99487 and 99489 as clinical staff codes under general supervision, and counts time spent by clinical staff or the billing practitioner toward each threshold.

• Model time is not staff time: Minutes a model spends drafting a note do not add to the monthly total; the nurse's time reviewing and editing it does.
• Honest logs: Time entries should record what a person did, so an AI-drafted note still needs the staff member's real minutes behind it.
• Consent and plan first: The MLN booklet requires documented consent and an electronic comprehensive care plan, which automation can prepare but not replace.
• One biller a month: Only one practitioner can bill CCM per patient per calendar month, so automated outreach must check enrollment before contacting anyone.

The business case therefore rests on capacity and quality. If AI cuts the clerical part of each contact, nurses spend more of their logged minutes on actual care, and the program can grow without hiring at the same rate.

Which federal rules apply to AI in a CCM program?

Two federal rules come up in almost every CCM automation project: FDA's position on clinical decision support software and the HHS nondiscrimination rule on patient care decision support tools. HIPAA applies as well, because every model sees protected health information.

• FDA device status: FDA's Clinical Decision Support Software guidance explains that software is excluded from the device definition only when it meets all four criteria in section 520(o)(1)(E) of the FD&C Act, including that a health care professional can independently review the basis for its recommendations.
• Signals and images: Software that analyzes a medical image, an in vitro diagnostic signal or a pattern from a signal acquisition system fails the first criterion, so a model reading continuous monitoring data needs a separate device review.
• Nondiscrimination: Under 45 CFR 92.210, covered entities must make reasonable efforts to identify patient care decision support tools that use race, color, national origin, sex, age or disability as inputs, and to mitigate the risk of discrimination from their use.
• Privacy and security: Vendors that host models handling care plans and call notes act as business associates and need a business associate agreement and HIPAA Security Rule safeguards.

In practice, outreach ranking tools deserve the closest look under 45 CFR 92.210. Age is almost always an input, and a model that ranks older patients lower because they answer the phone less often can quietly steer care away from the people CCM was designed for.

How should AI be rolled into a running CCM program?

Roll AI into a running program one task at a time, starting with drafts that nurses already review, and measure before and after on the same team. Programs that switch on several features at once cannot tell which one helped or caused trouble.

• 1. Baseline: Measure minutes per contact, documentation time, notes returned for correction and patients reached per nurse for one full billing month.
• 2. Pick one task: Start with pre-call briefs or call note drafts, where the nurse reads every output.
• 3. Shadow mode: Run the model for two to four weeks with outputs visible only to reviewers, and score accuracy against what nurses wrote.
• 4. Limited release: Turn it on for a few nurses, with an easy way to flag a bad output and a weekly review of flags.
• 5. Audit sample: Have a clinical lead review a set of AI-assisted notes each month for omissions, invented details and wrong medications.
• 6. Expand or stop: Move to the next task only when the first one is stable, and keep the option to switch any feature off.

Keep the prompts, model versions and review outcomes in a log tied to each output. When a payer or a patient questions a note, the program should be able to show what the model drafted and what the nurse changed.

Where does AI in CCM go wrong?

AI in CCM goes wrong when drafts are trusted without review or when automation creates contact the patient did not expect. These are the failure points to design against.

• Invented details: A call note includes a symptom or a dose the patient never mentioned.
• Copy-forward drift: Each month's note is drafted from the last one, and outdated problems keep reappearing.
• Unreviewed messages: An automated reminder answers a clinical question it should have routed to a nurse.
• Biased queues: Outreach ranking favours patients who are easy to reach over those at highest risk.
• Inflated time: Staff log the time a draft would have taken to write by hand rather than the time actually spent.
• Hidden failures: The model silently fails on some calls, and nobody notices because the queue still looks full.

How Agnotic can help

Agnotic Technologies builds AI chronic care manager tools for the client inside its existing CCM workflow: pre-call briefs, call note and care plan drafts with tracked edits, outreach ranking with bias checks, and time log review before month-end.

Agnotic designs every feature with a named reviewer, an audit log of model inputs and outputs, and a switch to turn it off, and connects it to the EHR through FHIR APIs or HL7 v2 so drafts land where nurses already work.

Sources

• CMS MLN909188: Chronic Care Management Services
• FDA: Clinical Decision Support Software, Guidance for Industry and FDA Staff
• 45 CFR 92.210: Nondiscrimination in the use of patient care decision support tools