Secondary Claims & COB: AI Agent vs Clearinghouse vs Biller

By Jude Lee · · Workflow

Two staff members in a small medical practice reviewing insurance claim denials on dual monitors

The four ways secondary billing actually breaks

If you pull your aging report and filter for balances sitting in a secondary bucket for 60+ days, the causes almost always sort into four piles:

  1. COB not on file with the payer. The patient has two plans, but the payer’s records say otherwise, so the claim denies with a coordination-of-benefits reason code (CARC 22 — “this care may be covered by another payer per coordination of benefits” — per the Claim Adjustment Reason Code list maintained by X12). Usually the patient has to call the payer to update COB, but not always: some payers accept a COB update submitted by the provider, and many will complete it on a three-way call with the patient on the line. That three-way call is a practical front-desk tactic worth scripting, because it converts a two-week wait into a five-minute task.
  2. Wrong payer sequenced as primary. Commercial vs. Medicare, auto/workers’ comp, a spouse’s plan, a child covered by two parents. Getting this wrong produces a denial that looks like a billing error but is really an intake error.
  3. The secondary claim never left the building. Either you assumed Medicare crossover handled it, or the primary’s remittance posted and no one queued the secondary. Per CMS’s Coordination of Benefits & Recovery program, Medicare will automatically cross claims over to a supplemental payer when that coverage is on file with the Benefits Coordination & Recovery Center — but only then. Verify the specifics for your MAC with CMS directly rather than assuming every supplemental plan crosses over.
  4. The secondary claim went out incomplete — or went out too late. A secondary claim has to carry the primary payer’s adjudication detail: paid amounts, adjustment reason codes, patient responsibility. When that arrives as a paper EOB or scanned PDF instead of an 835 remittance, somebody has to retype it. And many payers start the secondary timely-filing clock at the primary’s remittance date, not the date of service, which is why a balance that looks young by DOS can already be unappealable. Check each payer’s filing rule in its own provider manual.

Piles 1 and 2 are prevention problems. Piles 3 and 4 are the ones automation genuinely moves.

Option one: clearinghouse edits and crossover hygiene

The cheapest fix is usually not AI. Most clearinghouses let you build custom edits that stop a claim before submission — for example, flagging any claim where the patient record shows two active policies but no COB segment populated. Combine that with an annual audit of which of your payers actually auto-cross and which don’t, and you will close a meaningful share of the leak with configuration alone.

What rules can’t do: read a scanned EOB from a plan that doesn’t send electronic remittance, decide whether a denial is a sequencing problem or a genuine non-covered service, or write the appeal. Rules are deterministic, auditable, and cheap. They are also brittle the moment the input stops being structured. That trade-off is the same one we walk through in AI agents vs RPA vs rules.

Option two: an AI agent working the secondary queue

An AI agent is different from a chatbot: it carries out a multi-step task across your systems and takes actions, within limits you set. For secondary claims, a realistic scope looks like this:

That last item is underrated. The bottleneck in COB is often a two-minute patient phone call nobody has time to prompt. An agent that drafts fifteen personalized, accurate outreach messages a week is doing real revenue work with near-zero clinical risk.

The agent’s job isn’t to submit claims. It’s to make sure every claim that should be submitted is sitting in front of a human, already assembled.

Where agents break: inconsistent EOB layouts from small payers, payer portals that change without notice, and any judgment call about medical necessity or write-off authority. Treat extraction confidence as a first-class signal — low-confidence reads go to a person, not into a claim.

Option three: more human biller hours

An experienced biller or an outsourced RCM partner will clear a secondary backlog, and they bring something no agent has: working relationships with payer reps and pattern memory about which plans misbehave. The honest downside is that secondary claims are low-dollar, high-touch work — exactly what gets deprioritized when the day fills with high-dollar denials. If your billing is outsourced on a percentage of collections, ask explicitly how secondary balances are prioritized; many arrangements rationally favor the bigger checks.

Clearinghouse rules
Deterministic, auditable, cheap to maintain. Catches the same error every time. Fails on scanned EOBs, ambiguous sequencing, and anything requiring a drafted explanation. Best for prevention at submission.
AI agent
Handles unstructured inputs and drafts work product. Needs confidence thresholds, logging, and human approval. Best for cleanup, triage, and outreach drafting — not for autonomous submission.

The third column — human biller or outsourced RCM: strongest on payer-rep relationships, appeals that need a phone call, and institutional memory about plan quirks no rule encodes. Weakest on consistency under load, because low-dollar secondary balances are the first thing to slip when the queue gets heavy — and that risk grows if the fee structure rewards large recoveries.

What a custom build actually involves

If you go the agent route and your practice-management system doesn’t ship this feature, the modern pattern is to connect an assistant like Claude to your systems through MCP (the Model Context Protocol — an open standard for giving an AI governed access to specific data and tools). You’d expose a deliberately narrow set of operations: list_encounters_with_secondary_balance, get_remittance_detail, get_coverage_records, create_draft_claim. Nothing that submits. Nothing that adjusts a balance.

Then you define a skill — a packaged, reusable instruction set covering how to sequence payers, which adjustment codes mean “rebill” vs. “appeal” vs. “patient responsibility,” your write-off threshold, and when to escalate. Mechanically, a skill reduces run-to-run variation by pinning the same instructions to every execution; it doesn’t guarantee correctness, which is why the review queue stays.

Be clear-eyed about the cost side. This is a software project: engineering hours to build and test the tool definitions, then ongoing maintenance every time your PM system changes its API, your clearinghouse changes a file format, or a payer changes its portal. Someone has to own that — and if the person who built it leaves, an undocumented MCP server becomes a liability rather than an asset. If you can’t name the owner and the maintenance budget, buying an off-the-shelf feature or more biller hours is the more honest call. The safety mechanics — BAA coverage, minimum-necessary scoping, audit logging — are covered in our HIPAA-aware guide to connecting Claude to an EHR via MCP. Confirm your specific PHI-handling design with a qualified compliance advisor before it touches live data; HHS OCR’s HIPAA guidance is the governing reference, and your vendor’s BAA is the contractual one.

Modeling the payoff without making numbers up

We have no benchmark for what this saves your practice, and you should distrust anyone who quotes one. Build it yourself:

(claims/month) × (minutes each)
Current manual secondary workload
hours saved × loaded hourly rate
Labor value recovered — plug in your own rate
aged secondary balances × your own recovery rate
Derive the rate from your last 12 months: secondary dollars collected ÷ secondary dollars billed

The third line usually dominates, and it’s the one practices forget — but only count balances still inside each payer’s filing window from the primary’s remittance date. Recovered staff hours only matter if they get reallocated to revenue work: scheduling, treatment follow-up, patient collections. If the hours just evaporate into the day, your real return is captured revenue plus errors avoided. Pair this with your denial triage workflow and payment posting process — secondary billing sits downstream of both, and fixing posting first often shrinks the queue on its own.

How much of the queue to hand over

There’s no authoritative rule for what share of a workflow to automate, despite how often one gets quoted. The useful heuristic, stated plainly as our opinion: scope an agent to the portion of a queue where inputs are consistent and the downside of an error is rework, not a compliance event. For most practices that’s a minority of the queue at first, and it should grow only as your review logs show the agent is right on the cases you let it keep.

Where this fits your broader AI plan

Asking what the “best AI for a medical practice” is has a job-by-job answer, not a platform answer — our buyer’s map by job lays that out. Secondary claims is a good early candidate precisely because the output is a draft, the reviewer is already a billing expert, and success shows up in collected dollars within one or two cycles. Nothing here replaces a biller; it replaces the part of their day spent retyping an EOB.

  1. Quantify the leak

    Pull 90 days of secondary-responsibility balances, sort by the four failure modes, and mark which ones are still inside the payer’s filing window measured from the primary’s remittance date. That recoverable subset is your business case.

  2. Fix prevention first

    Add clearinghouse edits and a front-desk COB question. Audit which payers actually auto-cross.

  3. Pilot read-only

    Give the agent read access to remittances and coverage for one payer. It drafts; a human reviews everything. Log every extraction.

  4. Measure against the human baseline

    Compare the agent’s drafts to what your biller would have produced on the same 50 claims. Keep the scope where it wins.

  5. Expand by payer, not by permission

    Add payers before you add write access. Submission authority is the last thing to delegate, if ever.

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