AI in Medical Billing: What Works

AI in medical billing: separating useful tools from hype

AI in medical billing covers a wide range of software, from simple rules that flag a missing modifier to models that read a visit note and suggest codes. Some of these tools save real time on repetitive work. Others are promoted with claims that are hard to test, and none of them removes the practice's responsibility for what is billed under its NPI.

This article explains what the tools are used for, where they tend to work, where they fall short and how to judge one before you trust it with claims. It assumes you know the basic claim cycle; if not, start with our medical billing process guide.

Rules-based automation versus AI-assisted tools

Not every automated feature is AI. Automation in medical billing has existed for decades in the form of claim scrubbers, electronic eligibility checks and auto-posting of electronic remittances. These follow fixed rules that someone wrote and can read.

Aspect	Rules-based automation	AI-assisted tools
How it decides	Fixed if-then rules (for example, an edit that blocks a code pair)	Patterns learned from past data or language models that read text
Typical uses	Claim scrubbing, eligibility requests, ERA auto-posting, worklist routing	Coding suggestions from notes, denial risk scores, drafting appeal letters, summarizing payer correspondence
Explainability	Usually clear: the rule that fired is visible	Often partial: a score or suggestion without a full reason
Main risk	Outdated rules if no one maintains them	Confident but wrong output; bias toward patterns in past data

When a vendor says a product uses AI, ask which feature, what it does and what it does not do. Many valuable billing tools are rules engines, and that is not a weakness.

Where AI and automation are used in billing today

Eligibility and benefits

Batch eligibility checks before the schedule day are among the most mature uses. Software sends electronic inquiries for each upcoming visit and flags inactive coverage, plan changes or high deductibles. Newer tools try to read benefit details that payers return as free text, such as visit limits. Staff still need to call when the electronic response is incomplete, which is common for behavioral health and therapy benefits. See eligibility verification for the checks that matter most.

Coding suggestions

Computer-assisted coding tools read documentation and propose CPT, HCPCS and ICD-10-CM codes. They can speed up routine visits and point out documentation that may support an additional diagnosis. They can also suggest a higher-level code than the note supports or miss payer-specific rules. Suggestions should be reviewed by a certified coder before submission, and the clinician's documentation remains the basis for the code.

Claim scrubbing and denial prediction

Scrubbers check claims against code edits, payer rules and the practice's own history before submission. Some tools add a risk score that predicts which claims are likely to deny, based on past outcomes for similar claims. A score is useful for deciding which claims a person looks at first; it is not a reason to change a code.

Payment posting and AR work

Auto-posting of electronic remittance advice is routine. Tools can also sort unpaid claims into worklists by payer, dollar amount and age, or by likely next action. The follow-up itself, such as calling a payer, correcting a claim or filing an appeal, still requires a person who understands the denial. Our guide on how to reduce claim denials covers that work.

Appeals and correspondence

Language tools can draft appeal letters, summarize long payer letters or pull key fields from scanned EOBs. Drafts need checking against the actual record and the payer's appeal requirements before anything is sent.

Limits that still need human judgment

• Garbage in, garbage out. Tools cannot fix wrong demographics, inactive payer enrollment or an incorrect NPI on file.
• Documentation drives coding. A tool can only suggest codes the note supports; it cannot add work that was not documented.
• Payer rules change. Medical policies, edits and contract terms change during the year, and models trained on last year's data can lag behind.
• Errors scale. A wrong rule or suggestion applied automatically can affect hundreds of claims before anyone notices.
• Accountability does not transfer. The practice remains responsible for claims submitted under its NPI, whoever or whatever produced them.

A sensible pattern is to let tools handle sorting, checking and drafting, and to keep code selection, appeals and write-off decisions with trained people who can see the full record.

Compliance and oversight

Any vendor whose tool creates, receives, maintains or transmits PHI for the practice is generally a business associate under HIPAA and needs a business associate agreement before it gets access. HHS guidance on cloud services applies to many billing tools that process ePHI on a vendor's servers.

Beyond HIPAA, build oversight into the workflow:

• Keep an audit trail showing who or what changed a code, posted a payment or wrote off a balance.
• Sample AI-suggested codes regularly and compare them to documentation, the same way you would audit a human coder.
• Watch for drift, such as a gradual rise in higher-level E/M codes after a tool is introduced, and investigate the cause.
• Confirm what data the vendor keeps, where it is stored and whether it is used to train models for other customers.
• Check state law. Texas SB 1188 (89th Legislature, 2025) added requirements for practitioners who use AI for diagnostic purposes, including review of AI-generated records and disclosure to patients; read the enrolled text or ask counsel how it applies to you.

Payers are using automation too

Health plans also use automated review on their side. In traditional Medicare, CMS's WISeR model began in January 2026 in a group of states that includes Texas. It uses technology-enabled review, including AI, for prior authorization of selected services that CMS considers vulnerable to unnecessary use. CMS states that denials of coverage in the model are decided by appropriately licensed clinicians, not by software alone.

For practices this means two things: documentation must stand on its own when a machine reads it first, and denials still deserve a careful appeal, since automated payer review can be wrong too. Our prior authorization page explains how requests are tracked.

Medical billing trends 2026 practices should watch

A few medical billing trends in 2026 are driven by regulation rather than vendor marketing:

• Faster prior authorization decisions. Under the CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F), impacted payers, including Medicare Advantage organizations, Medicaid and CHIP programs and their managed care plans, and qualified health plan issuers on the federally facilitated exchanges, must send prior authorization decisions within 72 hours for expedited requests and seven calendar days for standard requests starting in 2026, with a specific reason for denials.
• Electronic prior authorization interfaces. The same rule requires those payers to build prior authorization APIs, with most API requirements starting in January 2027, which should make status checks and documentation requests more electronic.
• More automated review by payers, as shown by WISeR in Original Medicare.
• Documentation tools at the point of care, such as ambient note drafting, which affect coding because the note is the coding source.
• Continued attention to patient balances, as deductibles and self-pay amounts make up a meaningful part of collections for many practices.

None of these changes removes the basics: accurate registration, clean claims and steady follow-up still decide most of what a practice collects.

Questions to ask an AI or automation vendor

1. Which specific tasks does the tool perform, and which are rules-based versus model-based?
2. Does any output go to a payer without a person reviewing it?
3. How is accuracy measured, on whose data, and can we see results for our specialty?
4. How are payer rule and code set updates loaded, and how quickly?
5. Will you sign a BAA, where is our data stored, and is it used to train models for others?
6. Can we see why the tool made a suggestion, and is every change logged?
7. What happens to our data and workflows if we cancel?

Before rolling out a tool widely, record a baseline for denial rate, first-pass acceptance and days in AR, then run it on a limited set of claims and compare. Our medical billing KPIs article explains how to measure these.

How Fastora approaches technology and review

Fastora's billing workflows are technology-supported, and every claim is reviewed and handled by certified medical billing and coding professionals. If you want a second look at your current claims and denial patterns, a free billing assessment is a no-commitment place to start.

Checklist

Print checklist

• List the billing tasks you want help with before looking at tools
• Ask which features are rules-based and which are model-based
• Sign a BAA before any vendor touches PHI
• Keep human review for code selection, appeals and write-offs
• Record baseline KPIs and pilot on a limited claim set
• Audit AI-suggested codes against documentation on a schedule
• Check state law on AI use and patient disclosure

Frequently asked questions