AI & Digital Transformation

HIPAA-Aware AI Patterns That Still Let You Ship

Kevin PlumleeSeptember 21, 20264 min read
HIPAA-Aware AI Patterns That Still Let You Ship

“We can’t use AI, HIPAA” is a sentence that shows up a lot. Sometimes it is wisdom. Sometimes it is a polite way to avoid learning.

HIPAA is real. Penalties are real. “Move fast and break protected health information” is not a strategy. Neither is doing nothing while competitors answer the phone at 2am.

Focus on patterns that respect the regulation and still ship. This is not legal advice. These are operator habits that keep lawyers from becoming your product managers.

Pattern 1: minimize what the model ever sees

The cheapest control is not sending PHI into a model in the first place.

Can the agent do the job with a ticket ID and a callback number, while the clinical detail stays in the EHR your team already uses? Can the knowledge base answer FAQ without patient context? Can you redact names and member IDs from logs the model uses for debugging?

Minimum necessary is not just a compliance phrase. It is good architecture. Less sensitive data in the prompt means less sensitive data in the blast radius when something weird happens.

Pattern 2: BAAs and boring vendors

If a vendor touches PHI, you need the paperwork and the settings that match the paperwork. Business associate agreements. Clear data retention. No training-on-your-data surprises. Region and access controls you can explain.

Be wary of “the model is smart, we’ll figure compliance later.” Later is when the audit starts.

Also be wary of mystery middleware. Know the path: telephony provider, LLM provider, storage, CRM. Draw it on one page. If you cannot draw it, you cannot defend it.

Pattern 3: AI for intake and routing, humans for care decisions

This is a strong shipping pattern because it is useful and defensible.

The agent greets. Collects name and callback. Asks why they called at a high level. Checks for crisis language and escalates. Creates a clean note for the morning team. It does not diagnose. It does not invent coverage rules. It does not invent clinical advice because it read a blog once.

You can ship that slice while the “AI clinician” fantasy waits in the parking lot where it belongs.

Pattern 4: structured notes, not free-form diary dumps

When the agent writes into a system of record, use a schema. Fixed fields. Limited free text. No improvising new clinical categories.

That helps privacy reviews and it helps your staff. Nobody wants to dig through a novel of model prose to find the phone number.

Access controls still apply. The note lands where authorized humans already work. Not in a random shared Google Doc titled “AI dump.”

Pattern 5: human-in-the-loop as a product feature

Review queues are not a failure of automation. They are how you earn trust.

Early on, every conversation gets a human glance. Then you sample. Then you keep sampling forever for high-risk intents. Escalation paths are tested like fire drills, not documented like folklore.

If your team cannot pull up last night’s calls and coach the agent, you do not have a production system. You have a demo with a phone number.

Pattern 6: evals that include “should refuse”

Most teams only test happy paths. In regulated settings, the refusal path is the product.

Build a small suite:

  • crisis language must escalate
  • clinical advice requests must refuse and hand off
  • insurance questions stay inside approved FAQ language
  • the agent never invents a provider name or a policy ID

Run it when you change prompts. Prompt edits are code changes. Treat them that way.

What to ship first

If a healthcare org is choosing where to start tomorrow, do not start with chart summarization across the whole EHR.

Start with after-hours call capture, FAQ deflection for public info, appointment reminder workflows that avoid dumping PHI into the wrong channel, and internal tools that help staff move faster without putting a model in charge of care.

Useful. Bounded. Auditable.

Bottom line

HIPAA-aware does not mean AI-free. It means you respect the data, narrow the job, pick vendors you can contract with, and keep humans on the decisions that change someone’s care or money.

Ship the safe slice. Measure it. Expand only when the boring controls are boringly reliable. Move without pretending the rules are optional, and without using the rules as an excuse to keep losing callers to voicemail.

Kevin Plumlee

Founder, RVNOPS

Kevin Plumlee

15+ years running healthcare SEO, paid media, and operations. Now I advise the teams who do that work, across marketing, intake, technology, and AI.

More about me

Tell me what’s not working.

Stalled growth, spend you can’t trace, systems that fight your team. That’s a good place to start.

Start a conversation