Automation Teardown #6, dental edition: a claim denial comes back, and someone on the front-office team loses an afternoon working out what it means and writing the appeal from scratch.
The manual reality
Denials arrive by mail, fax, email, and payer portal, and they're written in insurance-speak. Someone has to read the letter, match it to the claim, work out the actual reason (a missing narrative, a frequency limit, a code the payer reads differently), and then write an appeal that answers that specific reason. The writing isn't hard. It just takes concentration the front desk rarely has between phone calls, so appeals pile up and some never get filed.
What the agent does
- One address. The team forwards the denial to a single inbox. For a paper or faxed letter, that means a scan or a phone photo. Forwarding is the only manual step.
- Decodes it in plain English. The agent reads the letter and states the reason in one or two sentences the team can act on. For example: "The claim didn't show why a crown was needed instead of a filling." It pulls out the patient, the claim number, the procedure codes, and the appeal deadline. Anything missing is marked missing, not guessed.
- Drafts the appeal. The draft answers the denial reason directly, with the claim details filled in and placeholders for the clinical evidence: the X-rays, the chart notes, and the dentist's narrative.
- Waits for a person. The decoded reason and the draft land wherever the team already works, like a chat channel or a shared queue, with an Approve button. Once someone approves it, the finished letter is saved as a document ready for attachments.
What stays human
The clinical case belongs to the dentist. The agent never writes a clinical justification it can't support from the chart, never decides an appeal is worth filing, and never sends anything to a payer. A person approves every letter, and the team attaches the clinical evidence themselves.
The guardrails that make it trustworthy
- It knows when it's unsure. A blurry scan or a faxed letter is really just a picture of text, and AI can misread it. When the agent isn't confident it read the letter correctly, it hands the letter to a person with a note instead of drafting an appeal from a guess. This is the most important behavior to ask any vendor about.
- Patient data stays inside a BAA. Denials contain protected health information, so every service that sees the letter is covered by a signed Business Associate Agreement: the inbox, the AI model provider, and the document store. See There's no such thing as "HIPAA certified".
- Every draft is traceable. Each draft is stored next to the letter it came from, so anyone can check the agent's reading against the original.
Honest sizing
If the team spends a few hours a week on appeals, most of that time is reading and first drafts, and that's the part this takes over. Expect to get back most of those hours, plus fewer appeals that never get filed because nobody got to them. It's usually live in a week or two using the email and documents the office already has. It doesn't replace a billing service; it gives the front desk a head start on the denials they already handle.
We've built a working version of this against a made-up practice. Every patient and claim in it is invented, and the pipeline is real.
We run free 20-minute teardowns of workflows like this. Start at /discuss or write to hello@fullspecstudio.com.