HIPAA Aware AI Chatbot for Healthcare

What a healthcare chatbot should and should not do

The useful jobs are narrower than the demos suggest. A clinic chatbot that answers which insurance plans you accept, how to prepare for a colonoscopy, where to park at the imaging center, or how to request records takes real load off the front desk. A bot that helps a patient book, reschedule or request a refill, after confirming who they are, saves phone time on both sides. Tepia starts every chatbot project by listing the questions your staff answer most often and sorting them by risk.

Some jobs the bot should refuse by design. It should not diagnose, interpret a lab result, change a medication plan or reassure someone describing chest pain. Tepia builds detection for emergency language that immediately shows 911 guidance, and for crisis language that surfaces the 988 Suicide and Crisis Lifeline, before any generated text appears. Questions that need clinical judgment go to a nurse line or a staff inbox with the conversation attached, so the patient does not have to repeat themselves.

The channel matters as much as the content. A website widget for anonymous visitors can answer general questions with no sign in, while the same assistant inside your patient portal or app can act on appointments because the patient is already authenticated. Text message chat through Twilio reaches patients who never download an app, but Tepia keeps SMS conversations free of clinical details unless the patient has agreed to receive them that way, since texts sit unencrypted on lock screens and shared phones.

Chatbot use cases ranked by patient data exposure

The amount of protected health information a use case touches decides the depth of engineering it needs. Tepia uses a matrix like this to plan releases, shipping low exposure features first:

A staff facing assistant that searches internal protocols is often the safest first launch, because it builds trust in the retrieval layer before patients ever use it.

Architecture that keeps PHI where it belongs

HIPAA does not ban AI. It requires that any vendor handling PHI on your behalf signs a business associate agreement and that you control access, logging and disclosure. Tepia designs the chatbot so each piece of that chain is covered:

• Model hosting through Azure OpenAI or AWS Bedrock under your cloud BAA, with data retention settings confirmed in writing and no use of your conversations for training.
• Retrieval augmented generation over a vector database of approved content, so answers come from your documents rather than the model’s general knowledge.
• Identity before data: the bot calls scheduling or portal APIs only after the patient signs in, and only with that patient’s scope.
• Minimum necessary prompts: tool calls fetch the single appointment or refill needed, not the whole chart.
• Audit logs of each conversation, retrieval and tool call, stored encrypted and retained under your policy, with identifiers masked in debugging views.

For a broader view of where models fit in clinical products, see AI in healthcare apps, and for the retrieval layer itself, RAG development.

Testing a chatbot before patients see it

Generated answers need a different kind of QA from ordinary screens. Tepia builds an evaluation set from real questions your staff receive, scrubbed of identifiers, plus adversarial prompts that try to extract other patients’ data, coax out a diagnosis or push the bot off topic. Each answer is scored for accuracy against the source document, for refusal where refusal is correct, and for tone.

Clinicians on your side review a sample before launch and after every content change, because a new prep instruction PDF can quietly change what the bot says. Tepia reruns the full evaluation set whenever the model version, the prompt or the content library changes, and tracks drift over time. After launch, unanswered questions and handoffs feed a weekly review that tells your team which content to write next. Our LLM evaluation and testing page covers the method in depth.

Tepia also tracks operational signals after launch: handoff rate, turns per resolved question, and the share of conversations where a patient asked for a human. A rising handoff rate on one topic usually means a policy changed and the content library did not.

How Tepia approaches HIPAA aware chatbots

Tepia runs its six phase process with healthcare specific checkpoints, and a US based project manager owns the plan throughout.

1. Discovery: analysis of call center and portal message logs, review of your content library and existing BAAs, and interviews with front desk, nursing and compliance staff. Deliverables are an Investigation Summary, an Interview Summary and User Stories such as a patient rescheduling after hours.
2. Design: conversation flows, handoff rules, disclaimer wording and a style guide for the bot’s voice, plus wireframes of the chat widget and the staff handoff inbox.
3. Development and Testing: an Alpha with staff only, then a Beta with a small patient group, with test plans covering evaluation scores, PHI leakage attempts, accessibility and load.
4. Training: hands on sessions for staff who receive handoffs and for content owners who update the knowledge base.
5. Launch: staged rollout on the website, portal and app, migration of existing FAQ content, and support reps watching transcripts during the first weeks.
6. Support: monthly evaluation reruns, model version upgrades, bug tracking and new intents as your team requests them.

Typically Discovery takes 1 to 3 months, development and QA 2 to 6 months, and launch about a month. Learn more about our process or talk to Tepia about your use case.

Frequently asked questions