A practice adds an AI tool that suggests evaluation and management (E/M) levels and diagnosis codes from visit notes. Within months, its coding pattern shifts. At the same time, a major payer starts reducing some higher-level E/M claims automatically. The practice is now part of what a 2025 npj Digital Medicine article, cited by MedCentral, called a "coding arms race," and it has to prove each code is supported.
In brief: AI medical coding can reduce manual work, but it does not make a claim correct. Practices stay defensible when every AI-suggested code is traceable to documentation, a qualified coder or clinician reviews what the tool produces, and leaders monitor coding patterns before payers do. The goal is accurate coding, not higher coding. Practices should also establish clear rules for reviewing AI-generated outputs, as explained in our guide to AI policies for medical practices
Why Is AI Medical Coding Under Scrutiny Now?
Because adoption is rising and payers are pushing back. In an October 20, 2026 MGMA Stat poll, 63% of medical group leaders said their revenue cycle operations are leaning toward more automation, with coding the most cited area, according to MGMA. The same article cites an April 2026 poll in which 33% of practices said they use AI for coding, charge review or both.
Payers see a cost problem. A Blue Cross Blue Shield Association (BCBSA) analysis released in September 2026 estimated about $942 million in added spending over two years tied to more complex billing in Blue System hospital inpatient claims, Healthcare Dive reported. BCBSA said diagnoses rose without a matching rise in treatment, suggesting AI is "identifying more billable conditions, not sicker patients." Coding vendors quoted in the same article said the tools mainly capture services that were previously missed.
BCBSA's estimate is also disputed. Becker's Payer Issues reported on October 14, 2026 that the American Hospital Association said the analysis did not review medical records or identify which claims involved AI tools. Becker's also notes BCBSA did not measure AI use at individual hospitals, and the study covered inpatient claims, not physician practices. Practices evaluating AI adoption should also understand how healthcare AI development incorporates integration, monitoring, validation and human oversight into existing healthcare workflows.
Federal leaders are watching too. CMS Administrator Dr. Mehmet Oz said AI will be "inflationary" in the short term because it will "turbocharge the ability of the current billing systems to work more effectively," Healthcare Dive reported on September 24, 2026.
For practices, the takeaway is simple. Coding changes after an AI rollout will be noticed, and the burden of proof sits with the documentation. Establishing appropriate AI governance for healthcare helps practices introduce oversight, monitor AI performance and maintain accountability as these tools become part of their workflows.
What Is Payer Downcoding, and How Are States Responding?
Downcoding is when a payer reduces the code level on a submitted claim, usually lowering payment. As MedCentral reported on August 21, 2026, Cigna announced a policy, effective October 1, 2025, to automatically reduce many level 4 and 5 E/M claims by one level unless documentation supported the higher level.
States have started to set guardrails, according to the same MedCentral report:
- Indiana (HB 1271): enacted in March 2026. Many health plans may not use AI as the sole basis to downcode a claim, and clinicians must keep humans in the loop when submitting claims.
- Illinois (Transparency in Coding Act): enacted in July 2026, effective 2028, limiting downcoding by automated tools alone.
- California (AB 2431): would have limited downcoding decisions to licensed professionals; it stalled in the 2026 session.
MedCentral also cites an Illinois State Medical Society survey in which 77% of physicians reported more automatic downcoding over five years. State laws generally do not reach self-funded employer plans, so protections vary by payer. This is not legal advice; check how these rules apply to your contracts with counsel.
How Accurate Are AI Coding Tools?
The honest answer is that independent evidence is thin. The U.S. Government Accountability Office's July 16, 2026 Science & Tech Spotlight states that "there are few independent studies evaluating the accuracy of these tools." GAO notes that errors could lead to over- or under-payment, and that the overall effect on health care spending is uncertain.
GAO also flags a practical access gap: under-resourced providers, including some small practices, may face cost, technical support and training constraints. Vendor accuracy figures exist, but GAO presents them as developer claims.
That means each practice needs its own evidence. Accuracy in your specialty, with your templates and your payers, is what matters.
How Can a Practice Keep AI-Assisted Coding Defensible?
Treat the AI tool as a suggestion engine inside a controlled workflow, built in these steps.
1. Set a Baseline Before Go-Live
Record your E/M level distribution, top diagnosis codes and denial rates by payer for the months before the tool goes live. Without a baseline, you cannot explain a change later.
2. Require Traceability for Every Suggested Code
Each code the tool suggests should point to the specific documentation that supports it. If a reviewer cannot see why a code was suggested, it should not be submitted.
3. Keep Qualified Humans in the Loop
Define which claims a certified coder or the treating clinician must review. Start with all higher-level E/M claims and any new diagnoses, then adjust based on audit results.
4. Audit a Sample Every Month
Pull a random sample of AI-assisted claims and have a coder review them blind. Track agreement, overcoding and undercoding separately. Both are accuracy problems; overcoding carries more audit and repayment risk, while undercoding loses earned revenue.
5. Watch Distribution Shifts by Provider and Payer
A sudden rise in level 5 visits or in secondary diagnoses deserves review before a payer asks. Compare each provider against their own baseline and against peers in the same specialty.
6. Build a Downcoding Response File
When a payer downcodes, capture the claim, the payer's stated reason and the supporting documentation. Patterns across claims help with appeals and contract discussions. Our article on front-end claim denial prevention covers the registration and authorization side of the same revenue problem.
Illustrative scenario: an eight-provider orthopedic group pilots AI coding suggestions for two months. Its monthly audit shows strong agreement on procedure codes but frequent extra secondary diagnoses that the notes do not fully support. The group keeps automation for procedures, requires coder review for added diagnoses, and updates its note templates. This is a hypothetical example, not a client result.
What Should Practices Ask an AI Coding Vendor?
Use this checklist before signing or renewing.
Question | Why it matters |
|---|---|
Is accuracy measured independently, in our specialty? | GAO found few independent accuracy studies |
Does every code link to supporting text in the note? | Traceability is the basis of any appeal or audit |
Can we set review rules by code type, level or payer? | Higher-risk claims need human review |
What reports show coding distribution changes? | Lets you spot shifts before payers do |
Where is PHI stored, and for how long? | GAO notes retention practices vary among AI scribe vendors; ask coding vendors too |
Can we export audit logs of suggestions and edits? | Needed for internal audits and payer disputes |
How does the tool handle payer-specific rules? | Payer policies on E/M levels differ |
If your practice has no written AI governance yet, define who approves tools like this, who reviews their output and who monitors them. Clear AI governance for healthcare makes these vendor answers easier to act on.
Where Does AI-Assisted Coding Go Wrong?
- Coding intensity without care change: if codes rise while care patterns stay the same, expect payer questions.
- Template drift: AI scribes and templates can add detail that changes code suggestions. Review templates when coding patterns move.
- Over-reliance: staff may stop questioning suggestions. Keep blind audits in place.
- Data fragmentation: coding, claims, denial and remittance data often sit in separate systems, which makes monitoring slow.
- Contract terms: check payer contracts for policies on code-level adjustments and appeal windows.
How Softnotions Can Help
Defensible AI coding depends less on the model and more on the data and controls around it.
Softnotions can support practices, billing organizations and digital health teams with:
- Connecting EHR, coding, claims and remittance data through data platform and integration engineering, so coding patterns and downcoding can be monitored in one place.
- Building audit sampling, traceability and review workflows around AI coding suggestions.
- Setting up governance controls, logging and access rules so AI-assisted work can be reviewed later.
We focus on the engineering. Coding policy and compliance decisions stay with your coders, clinicians and counsel.
AI coding tools are spreading quickly, and payers are responding with scrutiny and automated downcoding. Practices that set a baseline, require traceable codes, keep qualified reviewers in the loop and monitor patterns will be in a stronger position in audits and appeals.
Want to check whether your coding and claims data can support that kind of monitoring? Schedule a discovery call and we will review where the gaps are.
