Downcoding Appeals: AI Agent vs Biller vs Write-Off

By Jude Lee · · Workflow

Office manager and biller reviewing a remittance advice showing a downcoded claim on a computer screen

Downcoding isn’t a denial, and that’s why it goes unworked

Most billing workflows are built around rejection. A claim bounces, it lands in a work queue, somebody fixes it. Downcoding doesn’t trigger any of that. You bill a level 4 office visit, the payer processes it as a level 3, money arrives, the claim closes as paid. The difference shows up only as a line on the remittance advice that nobody reads line by line.

That’s also what separates it from the two adjacent problems. A denial is a non-payment you can route and resolve — see our comparison of AI denial triage versus scrubber rules and outsourced RCM. An underpayment is the payer paying the wrong rate for the code you billed, which is a contract-compliance question covered in payer underpayment detection. Downcoding is the payer paying the correct rate for a different code than the one you submitted. Same ERA, three different investigations.

The structural point for 2026: payers are running automated review at scale, and practices are still reviewing by hand. We wrote about that asymmetry in what payer AI agents mean for practices. Downcoding is the clearest place where it shows up on your bank statement.

What the three options actually look like

AI agent over your ERA + EHR data
Detection coverage: every 835/ERA line, every remit — flags any claim where the paid procedure code differs from the submitted one and groups by payer and provider. Judgment quality: poor on its own; it can draft but not decide. Cost: real setup effort, ongoing maintenance, and a BAA with any vendor touching the data. Failure mode: confident, well-formatted appeals built on documentation that doesn’t actually support the level.
A biller working it manually
Detection coverage: partial — usually high-dollar claims or a provider who complains; cross-payer patterns rarely get aggregated. Judgment quality: strong, assuming a certified coder. Cost: finite hours that compete with aging A/R on a harder deadline. Failure mode: the leak stays invisible because nobody had time to look at the remits line by line.

The third option — writing it off — deserves the same four fields, not just a sentence of prose:

A downcode you never detected isn’t a business decision. It’s a blind spot you’re calling a business decision.

The agentic workflow, step by step

  1. Detect on every remit, automatically

    Compare submitted CPT/HCPCS against adjudicated code on each ERA line. This is deterministic matching — a rule, not a language model. Flag mismatches, tag the payer, provider, date of service, and dollar delta.

  2. Aggregate before you appeal

    One downcoded visit is noise. Forty from one payer against one provider is a conversation with your network rep — and possibly a contract or policy issue rather than a claim-by-claim fight. The agent’s real advantage is seeing the pattern you can’t.

  3. Assemble the evidence packet

    Pull the encounter note, the problem list, any relevant orders, and the payer’s published E/M policy. An AI assistant connected to your EHR and practice-management system can draft a structured summary mapping documented medical decision-making elements (or total time) against the level billed. This is also where language models fail most convincingly: they will cite elements — a reviewed lab, a discussed risk, a stated time total — that aren’t actually in the note. Require the coder to verify every element the draft cites against the chart itself, not against the summary.

  4. Route to a human who can say no

    A certified coder or the rendering provider reviews the verified draft and answers one question: is the originally billed level genuinely supported? If not, the correct action is to drop the appeal — and to flag the documentation pattern for education.

  5. File and track the clock

    Appeal windows are short and payer-specific. For Original Medicare, CMS describes a five-level appeals process beginning with a redetermination by the Medicare Administrative Contractor, which must be filed within 120 days of receiving the initial determination notice. Confirm the current window and submission requirements on the CMS fee-for-service appeals page before you build any deadline logic. Commercial windows vary by contract.
Submitted code ≠ paid code
The single ERA comparison that drives the whole workflow
Workflow design
5 levels / 120 days
Original Medicare appeals process and redetermination filing window — verify current timelines
CMS
hours × rate
Model recovery with your own volumes and loaded labor cost — don't trust a vendor's headline number
Example method

Modeling whether it’s worth it

Don’t take anyone’s ROI claim here, including ours. Build the number yourself with four inputs you already have:

Recovered dollars = volume × delta × win rate. A worked example with entirely made-up numbers, purely to show the shape of the arithmetic: assume 40 downcoded claims a month, a $35 average allowed-amount delta, and a 50% win rate — that’s 40 × $35 × 0.50 = $700 a month in recovery. Now subtract cost: if a packet takes 15 minutes of coder time at a loaded $40/hour, that’s 40 × 0.25 × $40 = $400, leaving $300 before tooling. None of those figures are benchmarks; swap in your own four and the answer may flip entirely. Then add the second-order value: the documentation education that comes from seeing the pattern, and the biller hours freed from manual ERA review that get reallocated to aged A/R. That reallocation may be the larger line, but you have to actually redeploy the hours for it to count.

Build it, buy it, or just query it

The detection layer is often buyable. Many clearinghouses and practice-management systems can already report code-level adjudication variances; ask your vendor before you build anything. Some denial-management point tools add downcode flags as a feature.

Custom becomes worth it when the packet assembly — pulling the note, mapping documentation to criteria, formatting per payer — is the bottleneck, because that’s the part off-the-shelf tools rarely reach into. The practical pattern as of 2026: connect an assistant like Claude to your PM and EHR through MCP with HIPAA guardrails, expose only the minimum-necessary fields for appeal work, and package the payer-specific appeal format as a reusable skill so every packet comes out identical. Any vendor or model provider processing PHI needs a business associate agreement; HHS Office for Civil Rights guidance on business associates and the minimum necessary standard is the primary source to work from, and your privacy officer should review the data flow before it touches a single chart.

On the “30% rule” you’ll see floated in AI discussions — there’s no such standard from CMS, OCR, or any professional body, and people use it to mean wildly different things. Set your own human-review threshold based on risk: for coding appeals, our opinion is that it’s 100% human sign-off until you have a long track record, and even then the review stays.

What happens to the billing job

The worry behind “which medical jobs survive AI” has a boring answer in revenue cycle: the work that disappears is the scanning, the sorting, and the retyping. The work that remains — and gets more valuable — is judgment under ambiguity. Deciding whether a note supports a level 4. Deciding whether forty downcodes are a claim problem or a contract problem. Talking to the payer rep. A practice with an agent catching every downcode needs its coder’s judgment more than before, because now there’s a queue of real decisions instead of a queue of paper.

Start narrow: one query, ninety days of remits, one payer. If the number is small, write it off on purpose and move on to a bigger leak. If it isn’t, you now have the business case to build the rest.

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