Billing Question Call Deflection: Automate the Boring Half

A patient calls, reads out a number from a statement, and says they do not understand it. Your biller stops what they were doing, pulls the account, opens the ledger, and spends 4 minutes explaining a line item. Then the next one calls.

That call is deflectable, mostly. But only the boring half, and only if one thing is true first. J.P. Morgan's 15th annual Trends in Healthcare Payments report, published 21 May 2025, found that just 22% of consumers always know what they owe for a visit beforehand. The phone is where that gap gets settled, and the statement is what decides whether a machine can settle it.

This post splits the billing call into the 4 kinds it actually is, names the 2 you can automate end to end, explains why the other 2 need a person, and gives you the privacy and payment guardrails before any of it touches a patient line.

Why the statement decides whether the call can be deflected at all

Deflection is downstream of the statement. Automation can only read back what the record already holds, so if the bill does not carry a plain-language line item, the date of service, what the payer allowed, and what the patient owes and why, then nothing on the phone can explain it.

Consumer sentiment from 2022, not your call log. Each bar maps to a different kind of call.

HFMA's Patient Friendly Billing Project puts the standard plainly: statements should be clear, concise, correct, and written so an average reader can follow them. That is a design brief for the document, not a slogan for the website. Most practice statements fail the "concise" test and the "clear" test at the same time, by listing procedure codes and an insurance adjustment with no sentence attached.

The gap shows up in what patients say bothers them. In a 2022 survey of 2,026 US adults run by YouGov for AKASA, 38% called medical bills confusing, and the single biggest frustration, at 29%, was simply not understanding what they were being billed for. That is old, self-reported consumer sentiment rather than your call log, but it lines up with what billers hear every day.

There is a cheap test. Hand your own statement to somebody who does not work in healthcare and ask them to tell you what they owe and why. Whatever they cannot answer is a call you are going to take.

The four kinds of billing call, and which two are safe to automate

Billing calls are not one job. They are 4, and they split down the middle. Two are bounded: the answer already sits in the statement and the ledger, and the caller wants it read back or wants a way to pay. The other two need something automation does not have.

Two of the four billing calls are bounded. The other two need judgement or an investigation.

The first is statement explanation: "what is this charge?" The second is payment logistics: "how much, and where do I pay it?" Both are closeable without a person, assuming the caller is verified and the statement carries the answer.

The third is a coverage dispute: "my insurance should have covered this." That is a claims investigation, not a question. The fourth is hardship: "I cannot afford this." That is a judgement call with your financial policy attached, and the authority to make it does not sit with software.

Keep this separate from outbound work. Chasing money already owed runs on different timing and different consent rules, which we cover in AR follow-up and collections automation. If you have not scoped this kind of system before, start with what an AI receptionist does and where it stops.

Why the coverage dispute and the hardship call need a person

A denial call is not a billing question in disguise. The patient is asking you to explain a decision a third party made, for a reason often written down nowhere you can read. Answering it means opening the explanation of benefits, reading the payer's codes, and sometimes calling the payer.

The scale is real. KFF, publishing on 2024 data in March 2026, found that ACA marketplace insurers denied 19% of roughly 451 million in-network claims, with rates ranging from 3% to 36% across 157 insurers. Marketplace plans are not every payer, but the pattern travels.

The reason mix is what makes this undeflectable. Of those in-network denials, 36% were logged as "other" with no stated reason, 25% were administrative, 13% an excluded service, 9% a prior authorization or referral problem, and 5% medical necessity. More than a third carry no explanation at all, so nobody, human or machine, can answer that call from the statement alone.

It gets more expensive after the call ends. CAQH CORE priced a single manual claim status transaction at $15.96, drawing on the 2023 CAQH Index, and estimated that moving from a partly or fully manual process to a fully electronic one saves an average medical provider about 17 minutes per transaction. The patient's 4-minute call is the cheap part. The investigation it starts is not.

One more trap sits inside the estimate. Under the No Surprises Act, an uninsured or self-pay patient who schedules at least 3 business days ahead is entitled to a good faith estimate, and can start a dispute if the final bill is at least $400 above it. Thresholds and rules change, so confirm the current ones with your own counsel rather than treating any summary as advice.

What a deflected billing call actually looks like

The safe version of this call has a fixed shape, and the order matters more than the script. Verify the caller, confirm which statement they hold, explain only what the record says, then take payment through a compliant page or route with a structured handoff. Log a reason code every time.

Verification first, payment through a compliant page, a reason code on every call.

Verification comes first because everything after it is protected information. Match at least 2 identifiers to the account before a balance is spoken aloud, and treat a failed match as a routing decision rather than a hurdle to talk the caller past.

Confirming the statement sounds trivial and is not. A patient with 2 dates of service and 1 envelope is the ordinary case, and answering about the wrong visit is worse than not answering at all. Name the date of service and the statement date back to them.

Then explain only what the record holds: the charge, what the payer allowed, and what remains. When the caller wants to pay, hand off rather than improvise. When the call turns into a dispute or a hardship conversation, route it with the account, the date of service, and what the caller actually said, so the biller does not start from zero.

The privacy and payment guardrails this needs

Two separate rulebooks apply to the same 4-minute call. Billing detail is protected health information, because a balance tied to a date of service and a procedure identifies both a person and their care. Card data sits under payment card rules instead, which are stricter about storage than most practices expect.

On the health side, any vendor that hears or stores billing detail is a business associate and needs a signed Business Associate Agreement before it touches a live line. "HIPAA compliant" describes a configuration plus that contract, and "HIPAA certified" is not a thing that exists, whatever a vendor page says.

The practical answer is to never have the number in the first place. Send a payment link by text, hand the caller to a hosted page, or use a masked capture that keeps the digits out of the transcript and the recording. An AI receptionist should route to payment, not take payment.

Field note

Pre-flight before a billing line goes live

• Confirm a signed BAA covers the vendor that hears balances and dates of service.
• Confirm card numbers never enter the transcript, the recording, or the call log.
• Test the flow against a failed identity match, and check where that call lands.
• Have someone outside the practice read your statement and say what they owe.

Tag every billing call with a reason code

Deflection saves minutes. Reason codes stop the call from happening. Tag every billing call with why it came in, including the ones that routed to a human, and you turn a phone log into a list of upstream defects. The same 5 tags will explain most of your volume within a month.

Read the tags monthly and act on the biggest column, not the loudest complaint. A spike in "charge unclear" is a statement wording job that a designer can fix in an afternoon. A spike in "balance surprise" is not a billing problem at all, it is an eligibility check that is not happening at the front desk. This one call type sits inside a wider phone picture, and the front desk call metrics worth tracking put it in context.

Start with the two calls you can already answer

Pilot on statement explanation and payment logistics only, and leave disputes and hardship on the human path from day one. Narrow scope is what makes this safe, and it is also what makes the first month readable, because you are testing 2 known call types rather than an open-ended promise.

Route by what the call is, not by what the caller opened with. Every call ends tagged.

Walk the flow once. An unverified caller hears nothing about a balance. A coverage dispute goes to a biller with the benefits statement open. A hardship call goes to whoever can apply your financial policy. Everything else is a line item explained and a payment link sent.

Read the first 2 weeks of transcripts end to end, not a sample, because that is where you find the statement wording that confuses everyone and the question your knowledge base cannot answer. Then re-measure with your own reason codes after 30 to 60 days rather than a vendor's before-and-after slide. If you would rather see the size of this leak before anyone talks tools, the free Growth Leak Audit sizes it from your own numbers.