How Counterclaim writes a winning insurance appeal in under a minute

Extracts the facts from your denial.

The Reader agent ingests the OCR text of your denial letter or Explanation of Benefits and pulls out the structured facts that the rest of the pipeline depends on: insurer name, plan type, member ID, claim number, denial code (CO-50, CO-11, PR-1, and so on), CPT/HCPCS codes, ICD-10 diagnosis codes, billed amount, place of service, treating physician, denial date, and the deadline by which an internal appeal must be filed.

Reader runs on the same model as the rest of the pipeline but with a tightly scoped extraction prompt and a strict JSON schema for output. It also flags inconsistencies - for example, a CPT code that doesn't pair with the diagnosis on the claim, or a denial reason that doesn't match the cited code - so the Writer downstream knows where the insurer's own paperwork is weak.

Pulls the law and clinical evidence.

Running in parallel with Reader, the Researcher agent identifies the legal and clinical authority a winning appeal will need. That includes the relevant section of your state's insurance code (e.g., Texas Insurance Code § 1259, California's Knox-Keene Act, New York Insurance Law § 4904), federal law where applicable (the ACA's external review provisions, ERISA § 503, MHPAEA, the No Surprises Act), peer-reviewed clinical literature supporting medical necessity, and your insurer's own published clinical coverage policy.

Reader and Researcher are independent - Reader reads the patient's documents, Researcher reads the world - so they execute concurrently via Promise.all. Each writes its output to the database the moment it resolves, so the live pipeline visualizer ticks them off independently rather than at a single artificial moment.

Drafts the formal appeal letter.

Once Reader and Researcher both complete, the Writer agent receives both outputs and synthesizes them into a three-paragraph formal appeal letter formatted for USPS Certified Mail. Paragraph one establishes the patient, the claim, and the denial being contested. Paragraph two presents the medical and legal basis for the appeal, citing both clinical evidence and the specific statutes or policy provisions the insurer has misapplied. Paragraph three states the requested relief and the deadline for the insurer's response.

The Writer is constrained by a system prompt that requires every assertion to be supported by a citation supplied by the Researcher - no hallucinated case law, no invented statute numbers. If the Researcher didn't surface authority for a particular argument, the Writer omits that argument rather than fabricating one.

Attacks the draft as the insurer's lawyer.

The Adversary agent is given the Writer's draft and instructed to act as the insurance company's appeals attorney. Its job is to find every weakness: vague citations, claims unsupported by the cited evidence, factual statements the insurer could rebut with their own coverage policy, missing exhibits, and procedural mistakes that would let the insurer dismiss the appeal on technical grounds.

This adversarial step is where most single-shot AI appeal letters fall down. A draft that looks confident in isolation often crumbles when read by someone whose job is to deny it. By pre-litigating those weaknesses internally, Counterclaim forces the Editor downstream to either fix them or remove them.

Hardens the citations and ships the final letter.

The Editor agent takes the Writer's draft plus the Adversary's critique and produces the final, print-ready letter. It tightens citations, removes any argument the Adversary identified as weak, adds the document checklist a patient will need to attach, and emits clean print-ready HTML that renders to a PDF on download with no UI chrome - just the letter, formatted for certified mail.

Editor also runs a citation verification step: every statute, regulation, or guideline cited in the final letter is checked against the Researcher's source material and flagged with a verification status. Citations the system can't verify against a real source are removed rather than passed through as plausible-looking filler.