Why this matters
A denied prior authorization costs a clinic twice: the visit or scan waits, and someone in billing writes a one-page appeal. Staff have found that AI chat tools write those letters fast, so chart notes, names and MRNs are being pasted into tools nobody approved. One pasted MRN can be a reportable privacy incident.
The letter does not need to know who the patient is. It needs the clinical story: the diagnosis, what was tried, for how long, why it failed and why the service is needed now. In this module you build an Airia agent that works from those facts only and writes [PATIENT], [MRN] and [MEMBER ID] where the identifiers go, so staff fill them in inside the EHR. Then you paste identifiers in on purpose and see the honest limit of a prompt rule.
This lesson is training, not legal advice. A guardrail does not make a tool HIPAA compliant on its own: your organization also needs a Business Associate Agreement with any vendor that handles PHI, its own policies, access controls and staff training. Ask your compliance officer before any real patient data goes near an AI tool. Every patient in this lesson is made up (Pat Example, MRN 000-TEST).
What you’ll build
An Airia agent, Appeal Letter Drafter, whose instructions allow only clinical facts and replace every identifier with a placeholder.
A tested appeal letter from a made-up MRI denial, and a second test that shows what happens when someone pastes a name and MRN anyway.
What you need
- Nothing to start the practiceThe practice is built from real Airia screens and touches no account.
- An Airia account, to do it for realCheck airia.com for plans and agent limits. Use made-up patients only until your compliance officer approves real use.
- One denial with made-up factsThe denied service, the denial reason and the clinical facts. No names, dates of birth or MRNs.
Concepts
2 ideas you’ll use in the practice.
Minimum necessary, written into the prompt
HIPAA’s minimum necessary idea is simple: use and share only the information the task needs. For an appeal letter that is the clinical story, not the identity.
The instructions do three jobs: list what the agent may use, list what it must never write, and give the exact placeholders ([PATIENT], [DOB], [MRN], [MEMBER ID], [PROVIDER], [DATE]). They also say what to do when someone pastes identifiers anyway: leave them out and say so.
What a prompt rule cannot do
In our test the letter came back with placeholders and a note that identifiers were removed. The Run progress Input log still showed the whole message, name, date of birth and MRN included. The model received them.
A prompt rule protects the output. A guardrail that redacts identifiers before they reach the model is the stronger control. Until that is in place, the staff rule is the real protection: clinical facts only.
Practice
Do every step yourself in a practice copy of the real tool. It checks your work, and “Do it for me” finishes any step you get stuck on.
Do it for real
Now repeat it in your own account. Tick each step as you go.
Quick check
3 questions to check you’ve got it.
1. What does an appeal-letter agent need from the chart?
- The clinical facts of the denial
- The patient’s name and MRN
- The whole chart note
Show the answer
The clinical facts of the denial. The letter argues medical necessity. Staff add the identifiers in the EHR, where they already are.
2. Someone pastes a name and MRN anyway. What did the Input log show?
- The model received them, even though the letter used placeholders
- Airia removed them before the model
- Nothing was sent
Show the answer
The model received them, even though the letter used placeholders. A prompt rule shapes the answer. It cannot stop what is pasted from reaching the model.
3. Which control is stronger than a minimum-necessary prompt?
- A guardrail that redacts identifiers before the model
- A longer prompt
- A bigger model
Show the answer
A guardrail that redacts identifiers before the model. Only a control that runs before the model keeps identifiers away from it.
Airia: Sort Patient Portal Messages Fast Without Over-sharing
Next, the same minimum-necessary idea on the front desk: sort portal messages without repeating anything about the patient.
Start module 2 →