Credit Balances & Refunds: AI Agent vs PM Reports vs Biller

By Jude Lee · · Comparison

Office manager and biller reviewing a credit balance report on monitors in a small medical practice back office

The report nobody wants to work

Open your practice-management system and run the credit balance report. Most independent practices find a list that is older and longer than they expected — a line like $14.22 from two years ago (illustrative, but you’ll recognize the shape), a duplicate primary payment on a surgical claim, a Medicare overpayment sitting next to a patient prepayment that was never applied.

Credits accumulate for boring, structural reasons. Secondary payers pay before coordination of benefits is resolved. A patient prepays an estimate, then the contractual adjustment lands and the estimate was too high. A payer recoups on one claim and pays on another in the same remit, and the posting splits wrong. A refund check goes out, never gets cashed, and the credit reappears.

None of that is interesting work. All of it is work your practice is responsible for.

Three clocks, not one

Federal overpayments. Under the Affordable Care Act’s overpayment provision and CMS’s rulemaking on Reporting and Returning of Overpayments, identified Medicare overpayments must generally be reported and returned within 60 days of identification. CMS’s 2016 final rule for Medicare Parts A and B established a six-year lookback period, and CMS has revisited the standard for when an overpayment counts as “identified” in later rulemaking — so as of 2026, confirm the operative text directly with CMS and your compliance counsel rather than relying on any summary, including this one.

The quarterly CMS-838. Separately, CMS requires institutional providers to file the Medicare Credit Balance Report (Form CMS-838) quarterly, certifying Medicare credit balances on hand. Per CMS’s form instructions, this obligation attaches to institutional providers billing on the Part A/UB-04 side — hospitals, SNFs, hospital-based departments. A practice billing only professional claims on the CMS-1500 generally does not file it, but if you have any provider-based or facility component, check with your MAC before assuming you’re out of scope.

State refund and escheat deadlines. Many states have prompt-pay or prompt-refund statutes requiring that patient and commercial-payer refunds be issued within a fixed window after the credit is identified — a clock that starts independently of the federal one. And credits you genuinely cannot refund (bad address, uncashed check) usually don’t become practice revenue: state unclaimed property (escheat) statutes set dormancy periods and reporting calendars. Both vary meaningfully by state; your state insurance department and state treasurer or comptroller are the primary sources.

Two builds, one buy

PM report + biller

How it works: run the credit balance report monthly, sort by age and dollar amount, and have a biller research each line — pull the EOB, check COB status, decide payer refund vs. patient refund vs. posting correction.

Strengths: zero new tooling. The biller already knows your payers and your posting quirks. Judgment calls happen in one head.

Weaknesses: it competes with claims and A/R for the same person’s time, so it slips. Small-dollar credits get deprioritized indefinitely, which is exactly how an aging problem forms. Documentation of why each decision was made lives in a note field, if anywhere.

Custom AI agent + human approval

How it works: an agent reads the credit balance report plus the underlying transactions, classifies the root cause of each credit, and drafts the output — refund request, payer refund letter with claim detail, COB correction task, or posting-error fix. A human reviews a queue and approves.

Strengths: it prepares every line, including the small-dollar tail that never gets worked. Each draft carries a written rationale and the evidence attached, and consistent classification makes the monthly report comparable over time.

Weaknesses: build and governance cost. It depends entirely on posting data quality. It cannot log into a payer portal reliably, cannot cut a check, and should not be allowed to.

The buy option: an outsourced billing company or specialist refund vendor. Many billing partners work credit balances as a contracted line item or a percentage of resolved dollars, and some vendors do nothing else. How it works: you hand over report access (under a BAA), they research, and they return refund packets or submit them on your behalf. Strengths: no hiring, no building, and a specialist vendor has seen your payers’ refund processes more times than your biller has. Weaknesses: you inherit their documentation habits and their prioritization, and percentage pricing creates a pull toward the large, easy credits. Questions to ask: how do they document root cause per line, who physically signs disbursements, what’s their turnaround against your state’s refund deadline, and what happens to credits they can’t resolve. If you have no biller bandwidth and no appetite for building, this is a legitimate answer — sometimes the best one.

Where an AI agent genuinely helps

The useful mental model: this is a classification and packet-prep problem wearing a finance costume. Those are two things current AI agents do well.

Given a credit line plus the remits, charges, adjustments, and patient ledger behind it, an agent can sort into categories like payer duplicate payment, primary/secondary COB sequencing error, patient overestimate at time of service, posting error (adjustment entered as payment), recoupment timing artifact, or genuine prepayment against a future appointment. Then it assembles the artifact a human needs — the refund letter with claim number, DOS, paid amounts and the specific reason, or the corrected posting entries to review.

What it does badly: anything where the truth lives outside your data. If posting is messy, the agent will confidently classify garbage. That’s why payment posting quality is upstream of this project, not parallel to it.

An agent that explains why a credit exists is worth more than an agent that tries to make it disappear.

Building it: an MCP server over your PM data

If you go custom, the plumbing matters more than the model. MCP — the Model Context Protocol — is an open standard for giving an AI assistant governed access to specific tools and data. Instead of pasting ledger screenshots into a chat window, you expose a narrow set of read functions: get_credit_balances(as_of_date), get_transactions(account_id), get_remit(claim_id), get_patient_contact(account_id).

  1. Scope to minimum necessary

    The HIPAA Privacy Rule’s minimum necessary standard, as described in HHS Office for Civil Rights guidance, should shape your tool definitions. Credit balance work needs amounts, dates of service, claim identifiers and payer info. It rarely needs diagnoses or clinical notes — so don’t expose them.

  2. Get the BAA right before any PHI moves

    Whatever model provider and hosting you use must be covered by a business associate agreement, with terms that prohibit training on your data — which varies by provider and by tier. Verify the specific contract, not the marketing page.

  3. Make write access a separate decision

    Start read-only. If you later allow the agent to stage a posting correction, make it a draft a human releases. The permission models for agent access to clinical and PM systems apply here directly.

Transport, auth, audit logging, and how to connect an assistant like Claude to a practice system without leaking PHI are covered in the HIPAA-aware MCP build guide.

Write it as a skill, not a prompt

Encode your credit balance policy as a reusable skill — packaged instructions the assistant loads every time, so the job runs the same way in March as in January. A credit balance skill holds your classification taxonomy, dollar thresholds, refund letter templates per major payer, your state’s refund and escheat timelines, and an explicit escalation list: any government payer, anything over threshold, low agent confidence, patients deceased or in collections.

Be honest about what that leaves. In a Medicare- or Medicaid-heavy practice, routing every government-payer line to human review can send half or more of the monthly list straight to a person on the first pass. The agent still earns its keep there — it classifies, pulls the remit evidence, and drafts the letter, so the human reviews rather than researches. The surface where review is genuinely light is the commercial and small-dollar patient tail. That’s a real win, but it’s narrower than “the agent clears your credits.”

Modeling the payoff with your own numbers

Don’t take anyone’s published savings figure, including a vendor’s. Model it with inputs you can verify in an afternoon:

lines × min/line
Current research time: pull your credit report count and time ten lines yourself
hours × loaded rate
Labor cost of the manual process, at your actual staff cost
$ held ÷ aging buckets
Dollars sitting in credit, split by age — your exposure, not a benchmark

Then add the parts people forget. Recovered biller hours don’t vanish — they move to insurance A/R follow-up, usually higher-yield work, so count the reallocation. Subtract credits that turn out to be real patient credits applied to upcoming visits, which reduce future collection work. And weigh avoided risk honestly: you can’t price a compliance event you didn’t have, so describe it instead of inventing a number.

When the boring answer wins

Be skeptical of building if your monthly list is short enough that a biller clears it in an afternoon; if your PM vendor ships a credit balance module you haven’t turned on; if posting data is unreliable enough that a human re-verifies every output anyway; or if nobody owns this process today. Automation doesn’t create ownership — it amplifies whatever ownership exists.

In the broader job-by-job view of AI in a practice, credit balances are mid-priority. Scribing, scheduling, and denial work usually pay back faster. Credits earn their place when the list is aging, when an audit is plausible, or when you want one well-governed, low-clinical-risk workflow to prove out your MCP and approval patterns before touching anything near the chart. The research-and-classify portion is what agents compress; the payer call, the patient conversation, and the signature on the disbursement stay human.

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