90837 vs 90834: What to Do When Your Insurer Downcodes Your Therapy Claim with AI

If you bill insurance for psychotherapy, there’s a quiet revenue leak you may not have noticed. You bill CPT 90837 — a 60-minute session, properly documented, medically necessary. The claim doesn’t get denied. It gets paid. But when you read the EOB carefully, you see something strange: the payer paid you the 90834 rate. The 45-minute rate.

This is downcoding. It’s been around for years, but in 2026 it has a new accelerant: AI. Major payers are deploying machine learning systems that flag 90837 claims based on session frequency, diagnosis patterns, or documentation features — and either auto-downcode them or pend them for review. The result is the same. You did the work. You documented it. You billed correctly. And you got paid less.

What downcoding actually is

Downcoding is when an insurer processes your claim at a lower CPT code than what you submitted. It’s distinct from a denial — the claim is paid, just at a reduced rate. For therapists, the most common pattern is 90837 paid as 90834, but it also appears in:

• 90847 (family therapy with patient) being paid at the 90846 (family therapy without patient) rate
• 90791 (diagnostic evaluation) being paid at the 90832 (30-min therapy) rate when the documentation is read as “primarily therapeutic”
• Add-on codes (90785 for interactive complexity) being silently dropped from payment

The reason this matters more than denial-based revenue leakage: you can’t see it on a denial report. The claim shows as paid. The patient’s chart shows as billed correctly. The only way to detect it is to compare what you billed to what was paid, line by line.

Why AI made this worse in 2026

Three things happened in parallel.

First, payers deployed AI claim-review at scale. Cigna, Aetna, UnitedHealth, and several Blues plans have publicly disclosed AI tools that review claims before payment. The systems are pitched as “fraud and abuse detection,” but in practice they catch a much wider net — any claim that statistically deviates from the system’s expected pattern gets flagged.

Second, 90837 has a long history of being a payer target. Even before AI, 90837 was the most-audited mental health code. The OIG has flagged it multiple times. Some payers historically required prior authorization for any therapist whose 90837 utilization exceeded a threshold. AI just automated what auditors used to do manually — and made it cheaper to apply to every claim, not just statistical outliers.

Third, the legislative response is just beginning. The American Academy of Family Physicians has called on the federal government to investigate AI downcoding, and lawmakers in multiple states have introduced 2026 bills that would either require insurers to disclose AI use or ban automated downcoding outright. Until those bills pass — and most haven’t — the practice continues.

How to detect downcoding in your own practice

You need to compare two things: the CPT code you submitted on the claim, and the CPT code the payer actually paid. The data lives in your ERAs (Electronic Remittance Advice) or EOBs.

The 30-minute audit

1. Pull your last 90 days of ERAs for one payer at a time. Most EHRs let you export this.
2. Filter for any line item where you billed 90837. There should be a “billed CPT” field and a “paid CPT” field.
3. Count the rows where billed CPT ≠ paid CPT. That’s your downcoding rate.
4. If it’s above zero — and especially if it’s above 10% — you have a pattern worth appealing.

Run this for each of your top 3-5 payers. You’ll often find that one payer is responsible for most of the downcoding, while others pay clean. That’s actionable: you can renegotiate, appeal in batch, or — in extreme cases — drop the panel.

How to appeal a downcoded 90837

Most payers have a 60-180 day appeal window from the date of the EOB, depending on your contract. The appeal needs three things to win:

Single-claim appeals are time-consuming and rarely worth it on their own. The leverage comes from batch appeals — submitting 20-50 downcoded claims at once with a cover letter that documents the systematic pattern. Payers don’t like systematic-pattern appeals because they signal regulatory complaint risk.

What this means for the next 6 months

The legislative landscape is moving fast but unpredictably. A few of the proposed state bills would require payers to disclose every claim where AI was used in the adjudication decision. Others would create a private right of action — meaning a therapist could sue a payer directly for systematic downcoding. None have passed yet.

What’s more likely in the near term:

• State insurance commissioners opening investigations after enough complaints. Maryland and California have hinted at this.
• Class action lawsuits from billing companies. Several have already been filed against major payers in 2025-2026.
• Quiet payer policy revisions. When pressure builds, payers tend to update their AI thresholds without announcement. Watching reimbursement patterns over time becomes the only reliable detection method.

None of this helps a solo therapist getting downcoded today. What helps today is detection, batch appeals, and choosing your panels carefully.

The bigger picture

AI downcoding sits in a larger pattern that hit independent practices in 2026: payers using automation to extract margin without changing the contract. It’s the mirror image of the AI-driven hospital upcoding story, where health systems use AI to push claims toward higher-paying codes. The technology is the same. The direction of force is opposite. And the people caught in the middle — small practices without billing departments — pay the cost both ways.

The defense isn’t more billing software. It’s regular monitoring of your reimbursement patterns, claim-by-claim, payer-by-payer. The therapists who catch this are the ones who treat their ERA data as something to read, not just file.