Insurance AR Follow-Up: AI Agent vs Tools vs Biller

By Jude Lee · · Comparison

Office manager and biller reviewing an aged insurance accounts receivable worklist on a monitor in a small medical practice

The work hiding inside “AR over 30 days”

Run your aged AR report and the denials are the easy part: a denial carries a reason code, and you can route it. The harder pile is everything in limbo. A claim went out clean, the payer acknowledged it, and then… nothing. Is it pended for records? Suspended for COB? Applied to a deductible and awaiting patient responsibility? Sitting behind a prior-auth mismatch? Lost after a payer system migration?

Working that pile is a loop: pull the aged bucket, check status, interpret an often-vague response, decide (wait / call / resubmit / corrected claim / appeal / write off), do the thing, document it, and set a follow-up date. Most of those steps are reading and judgment — which is exactly the shape of work large language model agents handle better than traditional automation, and exactly the shape of work that needs guardrails.

The clock matters. Per CMS, Medicare fee-for-service claims must generally be filed within one calendar year of the date of service; commercial payer deadlines vary by contract and are frequently shorter. Verify every deadline against your actual payer contracts and current CMS guidance rather than a spreadsheet someone built in 2019.

A claim that ages quietly past its filing or appeal window isn’t a delay. It’s a write-off you chose by not looking.

Option one: rule-based status checking

Under HIPAA Administrative Simplification, HHS adopted the ASC X12N 276/277 health care claim status request and response as a standard electronic transaction — which is why most clearinghouses and practice management systems can automatically ping payers for status and drop results into a worklist. If your PM system supports scheduled batch status checks, turn them on before you buy anything. This is free-ish, deterministic, and auditable.

Its limit is the response itself. A 277 can come back with a status category that means, functionally, “we have it and we’re thinking about it.” Rules can bucket that; they can’t tell you why, and they can’t read the three-line note a rep left in the payer portal. This is the same rules-versus-agents boundary we walk through in AI agents vs RPA vs rules: if the input is structured and the decision is a lookup table, you do not need an LLM.

Option two: a human biller working the queue

A good AR specialist is still the benchmark, especially for payer phone calls, escalations, and anything requiring relationship or persuasion. The constraint is throughput and coverage. AR follow-up is the first thing to slip when the same person is also posting payments, handling patient calls, and covering the front desk — and it slips invisibly, because nobody complains about a claim that’s merely late.

Outsourced RCM solves coverage and adds scale, at the cost of visibility into which claims got real attention. If you’re already comparing that route, the tradeoffs overlap heavily with AI denial triage vs scrubber rules vs outsourced RCM.

Option three: an AI agent working the messy middle

An agent here is not a chatbot. It’s a system that takes a multi-step task — “work the 45–60 day commercial bucket” — and executes it: pull the claims, retrieve status, read the response and the payer’s remark text, compare against the original claim and the encounter, classify the situation, and produce a recommended next action with the reasoning and a draft artifact.

One thing it does not do: make the call. Payer IVR trees and rep conversations stay fully human in this design. The agent drafts the script and the reference numbers to read out; a person dials, waits on hold, and negotiates. Do not budget as though phone follow-up disappears.

  1. Pull the bucket

    Agent queries the PM system for open claims by age, payer, and dollar band — read-only, minimum necessary fields.

  2. Gather evidence

    It fetches the claim status response, remittance history, prior notes, and relevant encounter data. Prefer API and X12 EDI paths for payer data; before pointing any automation at a payer web portal, check that payer’s terms of use, because portal scraping commonly violates them and usually depends on sharing named-user credentials.

  3. Classify and explain

    It assigns a disposition (pend/resubmit/corrected claim/appeal/patient responsibility/call) and states, in plain language, what evidence drove it.

  4. Draft the action

    It prepares the artifact: corrected claim delta, appeal letter with dates and codes filled in, or a call script naming the exact reference numbers.

  5. Human approves

    A biller reviews the queue, approves or corrects. Nothing is submitted to a payer or written to the chart without that click.

  6. Close the loop

    Approved actions post back with a follow-up date, and every disposition is logged so you can measure agreement rate over time.

Assume specific failures, because these are the ones that show up. An agent can fabricate a payer reference number or a call-log ID that looks plausible and does not exist — so never let a generated reference reach a payer unverified. It can read a stale portal note and recommend a resubmission on a claim that was already reprocessed, creating a duplicate. It can misclassify a COB pend as patient responsibility and send a bill the patient does not owe. And it can drift silently: a payer rewords its remark text, classification quality degrades, and nothing alerts you because the agent keeps producing confident output. Sample the queue weekly.

Buy the module or build the agent?

As of January 2026, well-funded vendors are actively shipping AR-agent and claim-status modules into this exact workflow, and the landscape shifts quarterly — re-check who exists and what they support before you commit. Funding and momentum are not proof that a product works in your payer mix, but they do mean “we’ll build it ourselves” deserves a harder justification than it did two years ago.

Off-the-shelf AR agent module
Three real categories: clearinghouse-native status modules that already hold your 276/277 connectivity, AI add-ons from your PM/EHR vendor that sit directly on your claims data, and specialist RCM agent vendors with deep payer-specific logic. Fastest path — payer connectivity and status parsing are built, and the vendor absorbs payer-format churn. Tradeoffs: your follow-up logic bends to their model, you inherit their queue UI, and switching later means re-teaching everything. Best when your payer mix is common and your process isn’t unusual.
Custom agent over your own systems
You connect an assistant like Claude to your PM system through MCP — an open protocol for giving an AI governed access to specific tools and data — and package your follow-up playbooks as reusable skills. Tradeoffs: real engineering, real maintenance, real security review, and you own payer-format churn yourself. Best when you have payer-specific quirks that generic tools flatten, or multiple systems that need to be reasoned across.

A middle path many practices underrate: use off-the-shelf tools for connectivity and a narrow custom layer only where your logic is genuinely differentiated. If you want the mechanics of the connection, connecting Claude to your EHR via MCP covers the guardrails in detail.

Choose per job, not per tool

People search for a single best tool. There isn’t one, and AR follow-up is a good illustration of why. General-purpose assistants (ChatGPT and Claude are both general-purpose large language model assistants, not medical software) are strong at reading and drafting but have no native connection to your claims data and no BAA by default. Your PM or EHR vendor’s built-in AI has the data but usually a narrower feature set. A specialist RCM agent vendor has payer depth but lives outside your workflow. The right answer is job-by-job — which is how we’d suggest mapping it across the whole practice in the job-by-job buyer’s map.

There is no industry-standard “30% rule” for AI coverage

This phrase circulates online, and it’s worth being blunt: there is no official or industry-standard “30% rule in AI.” It’s used loosely — sometimes to mean a model handles roughly a third of cases unaided, sometimes as a budgeting heuristic. Treat it as folklore, not a benchmark.

The useful reframe, and this is our opinion rather than a measured finding: before rollout, decide what share of your AR queue the agent must close with human approval and no rework for the project to be worth it, then measure the actual rate for 60 days. That number is yours alone — it depends on your payer mix, claim volume, and how clean your upstream data is.

Modeling the money without inventing it

Do this on your own numbers. The inputs are formulas, not findings:

claims × minutes ÷ 60
Weekly follow-up hours today
hours × loaded hourly rate
Labor currently in the queue
measure, don't assume
Share of queue the agent closes with approval only

Worked example, all assumptions invented for illustration — substitute yours. Say 400 open claims in the bucket, 6 minutes per touch, and a $32 loaded hourly rate: 400 × 6 = 2,400 minutes, or 40 hours per pass; 40 × $32 = $1,280 of labor per pass through the queue. If review-and-approve cuts the per-touch time to 2 minutes on the share the agent handles cleanly, recompute with that number — and only after you’ve measured it, not before.

Then add the two effects that usually dwarf labor savings. First, captured revenue: claims worked before a filing or appeal deadline instead of after. Pull last year’s timely-filing write-offs — that’s a real number you already have. Second, reallocation: hours returned to the biller don’t vanish, they move to higher-yield work like appeals and underpayment review, so count the value there, not as a headcount cut you’re not going to make.

If the labor math alone justifies the project, be suspicious — the assumptions are probably too generous. The practice automation ROI and HIPAA vendor checklist walks through the same discipline for vendor selection.

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