trunkline-payer-call

Trunkline payer call

A medical practice's billing office spends its day on hold with health plans asking four questions: what happened to this claim, why was it denied, where is this authorization, and is this patient covered. This skill compiles one of those questions into a CALL-E call task and checks the answer that comes back.

Two rules make it different from a general "call and ask" skill, and both are enforced before and after the call rather than requested in the prompt:

1. Minimum necessary. Each workflow declares the only patient identifiers it may disclose. Eligibility gets a member id and a date of birth. It does not get a name, because it does not need one.
2. No answer without evidence. Every field comes back with the representative's own words. If that quote is not in the transcript, the field becomes unknown and a person looks at it.

When to use

Use this skill when a provider's billing office needs an administrative answer from a payer about a claim it has already submitted, and the answer is only available by phone.

When not to use

Do not use this skill to call a patient, to file or withdraw an appeal, to agree to any amount or adjustment, to discuss clinical care, or to place a live call without explicit approval and a written authorization for that destination. Trunkline gathers and structures; people decide.

Workflow

1. Choose the workflow. One of claim_status, denial_reason, prior_auth_status, eligibility. Read references/schemas.md for the closed result schema each one returns.
2. Assemble the request. Practice NPI, payer name and E.164 number, and one entry per claim: claim number, date of service, and a patient object holding that patient's identifiers.
3. Build the plan and read it. Run the preview below. It applies the disclosure envelope, drops every field the workflow may not disclose, and prints the exact task text and schema. It places no call.
4. Confirm intent. Show the operator the masked destination, the claims, and which identifier field names will be disclosed. Get explicit approval before dialling.
5. Place the call with run_call, or POST /v1/calls with the printed task and result_schema and an Idempotency-Key.
6. Check the answer with the check command. It validates the result against the schema locally and confirms every evidence_quote appears in the transcript. Anything it cannot confirm is reset to unknown.
7. Scrub before storing. Remove the patient identifiers from the transcript before it is written anywhere. Do not store a raw payer transcript.