Every chatbot vendor selling into healthcare says it will sign a BAA. Almost none say what changes in exchange.
Two terms first, because the rest of this depends on them. HIPAA is the Health Insurance Portability and Accountability Act, the US law covering how patient information is handled. PHI, or protected health information, is any health detail that can be traced back to a person: a symptom, an appointment, a prescription, a test result, or a name attached to any of those. A BAA, or Business Associate Agreement, is the contract that lets a vendor handle PHI on your behalf.
If visitors can type health details into your chatbot, all three apply to you.
A HIPAA workspace is not the normal product with a contract attached. Content goes in one way. Chats stay in one place. How long chats are kept becomes part of your contract rather than a setting. Most teams learn this after signing, which is too late to plan around.
iShort answer
A chatbot is HIPAA compliant when three things are true. The vendor signs a BAA. The service meets HIPAA Security Rule safeguards. And patient data stays in chats, out of the content you train the bot on. The first is a contract, the second is the vendor's job, and the third is yours. On SiteGPT the BAA comes with the Enterprise plan only. In a HIPAA workspace you upload files instead of syncing them from a cloud drive, and chats run in SiteGPT's own widget. Cloud drives and helpdesk tools are off by default, and can be switched on for any vendor you hold your own BAA with.
Five things to get right, in the order you will meet them.
A signed BAA
The contract that lets patient data reach the vendor at all. No BAA means no patient data, however good the security is.
Security Rule safeguards
Encryption in transit and at rest, access controls, audit logs, and a breach process with a deadline attached.
Training content with no patient data
Pages, files, snippets, Q&A entries and instructions stay clean. This is a contract rule, not a HIPAA one. It matters because anyone who can chat with the bot can reach that content.
How long chats are kept
Set in your order form, not in a settings menu. Work out the number with your lawyer before you sign.
One way to add content
You upload files yourself instead of syncing a cloud drive, and chats run in the widget. Any source can be switched on if you hold your own BAA with that vendor.
Key takeaways
| Question | The short answer |
|---|---|
| Is there a HIPAA certification? | No. No government body certifies anyone. Vendors state their own compliance, sometimes backed by a SOC 2 Type II audit. |
| What makes it legal? | A signed BAA. Everything else is a security control, not permission. |
| Where can patient data live? | In chats. Keeping it out of training content is a rule you agree to in the BAA, not a HIPAA requirement. It holds because anyone who can chat can reach training content. |
| Who keeps it out of training content? | You do. Nothing detects it automatically, which is why you upload each file yourself and check it first. |
| Which SiteGPT plan? | The Enterprise plan, at custom pricing. Not Starter, Growth, or Scale. |
| How long are chats kept? | Chat text is removed 7 days after the last message, and you can change that number in your order form. Leads stay until you delete them. |
The first question: will the vendor sign a BAA?
Any vendor that handles patient data for you is a business associate under HIPAA. The BAA is the contract that makes that legal. Without one, sending patient data to that vendor breaks the law on your side, however well built their product is.
Ask this first. It is not the most interesting question, but it is the only one that rules a vendor out in a single email.
One more thing to know before you read any vendor page: there is no such thing as HIPAA certification. No government body certifies anyone. Every "HIPAA certified" badge is the vendor saying so itself. Sometimes an independent audit such as SOC 2 Type II or HITRUST sits behind it. Sometimes nothing does.
So ignore the badge. Ask for two documents instead: the BAA, and the security documentation behind it. A vendor that will show you neither before a sales call has told you something.
SiteGPT publishes both. The HIPAA program page states what is available, and the standard BAA is published in full so your lawyer can read it before you talk to anyone. The limit comes in the same breath: the BAA is available on the Enterprise plan only, not on Starter, Growth, or Scale.
What HIPAA actually asks of a chatbot
The HIPAA Security Rule is not a technology checklist. It groups safeguards into categories and expects you to apply them sensibly for your size and risk. That is why two honest vendors can build very different systems and both be fine.
Here is what each category means for a chat widget.
| Safeguard category | What the Security Rule asks | What that means for a chatbot |
|---|---|---|
| Administrative | Risk checks, staff training, access management, incident response | Someone decides what the bot may read, staff know patient data never goes into training content, and there is a named process when something leaks |
| Physical | Building access, workstation and device security | This lands on the vendor's hosting provider and reaches you through their subprocessor agreements |
| Technical | Access control, audit logs, data integrity, secure transmission | Encryption in transit and at rest, per-user access to transcripts, logs of who read what, and a clear deletion path |
| Organizational | Business associate contracts, passing terms down the chain | The BAA itself, plus a rule that every vendor further down the chain signs its own |
Two of these four are the vendor's job. One is mostly yours. The administrative row is where chatbot projects usually fail, because it is the only row you cannot solve by buying something.
For the source, HHS publishes a Security Rule summary and guidance on business associates.
Where patient data is allowed to live
A chatbot holds two separate kinds of data. Mixing them up is the most common mistake here.
Training content is what you feed the bot: website pages, uploaded files, text snippets, Q&A entries, instructions, and custom responses. Chats are what visitors type and what the bot replies.
Patient data belongs in chats and stays out of training content. It is worth being clear about where that rule comes from: it is a contract term, not the law. Nothing in HIPAA says patient data cannot sit in a knowledge base. Almost every chat vendor asks customers to keep it in chats only, and you agree to that in the BAA.
The rule is still a good one, for a simpler reason. Training content is what the bot pulls from when it answers. So anyone who can chat with the bot can reach that content indirectly. A chatbot is usually open to every visitor on your site, which makes a knowledge base the wrong place for one person's medical details, whatever the contract allows.
| Where | Patient data allowed? | Why |
|---|---|---|
| Chats with visitors | Yes, this is what the BAA covers | A patient asking about their own care is the reason the chatbot exists |
| Website pages and uploaded files | No | The bot pulls from this content to answer, so any visitor can reach it |
| Q&A entries and custom responses | No | A saved reply quoting a real patient case counts, even though it never came from a medical record |
| Instructions and prompts | No | Instructions are part of what the bot reads, not a private note |
| Lead form fields | Avoid by design | Do not build fields that invite people to describe a condition, and delete any lead that contains one |
That last row is a design choice, not a policy one. A free-text box labeled "How can we help?" on a clinic site is an invitation to type a diagnosis. Rename it or limit it.
If patient data does land in a lead, it is not a crisis. Leads are never deleted on a schedule and stay until you remove them, so the job is noticing and clearing it. Lead settings are stricter on a HIPAA workspace for the same reason: ready-made form templates are hidden so every field is a deliberate choice, and the builder warns you when you add a free-text field.
What changes in a HIPAA workspace, and what you use instead
This is the part most vendors leave out, and it is the part that decides whether the tool fits your team.
A HIPAA workspace has one way to add content. You upload the file, paste the text, or point the bot at your own website. You do not connect Google Drive, Notion, Dropbox, OneDrive, Box, or GitHub and let them sync.
The reason is the whole design. A sync copies whatever sits in the folder, and nobody re-reads every file in a shared drive before connecting it. One old discharge summary in a subfolder is enough to put patient data where every visitor can reach it. An upload is one file at a time, chosen on purpose. That is the moment a reminder can appear telling you the file must be free of patient data.
Chats work the same way. They run in SiteGPT's own widget instead of flowing into Zendesk, Slack, or Crisp, so the transcript stays where the BAA covers it.
The test is narrower than it looks. It is not whether the vendor is HIPAA compliant, because almost every large vendor is. It is whether you have a BAA with them. Without one, that vendor is entitled to assume your data holds no patient information, which leaves it free to store and process those chats through any supplier it likes. With one, the chain holds and the integration can be turned on.
Here is the same thing as a table, including what you would do if HIPAA were not a concern.
| What compliance needs | What SiteGPT Enterprise gives you | If HIPAA is not a concern |
|---|---|---|
| A signed contract before any patient data arrives | A published BAA on the Enterprise plan, amendable during onboarding | No BAA needed, any plan works |
| Every system touching patient data covered by its own contract | Upload files, paste text, or crawl your own website. A cloud drive can be switched on if you hold a BAA with it | Connect Notion, Google Drive, Dropbox, OneDrive, Box, or GitHub and let them sync |
| Chats kept inside one covered channel | SiteGPT's own chat widget, with human takeover in the dashboard | Route chats into Zendesk, Slack, Crisp, Messenger, Google Chat, Freshchat, or Zoho SalesIQ |
| Patient data kept no longer than you need it | Chat text removed 7 days after the last message, and the number is set in your order form | Chats kept until you delete them |
| Training content free of patient data (a contract rule, not the law) | The same, with uploads as the only way in so each file gets checked | Train on whatever content you like |
| Nothing specific about branding | White-label, though the "Responses are AI-generated" line stays | Full white-label, branding removal at $39 a month |
Most of what a healthcare deployment needs still works: crawling your website, uploading files, adding text snippets, Q&A entries and custom responses, white-label branding, and answering in 95 or more languages. Human handoff is unaffected. Chat history, takeover by a human agent, and chat modes all work normally in the dashboard.
One migration detail: if you sign a BAA on an account that already has sources connected, those syncs are switched off when HIPAA is enabled. If your team relies on a Drive or Notion sync today, decide beforehand whether to upload that content directly or produce your own BAA with that vendor.
How long chats are kept
In a HIPAA workspace, chat text is removed seven days after the last message in that chat. You can change that number, but you change it in your order form, which makes it part of your contract rather than a setting an admin can edit later.
Leads work differently on purpose. They stay until you delete them. SiteGPT's reasoning is that deleting them on its own schedule could destroy records you are legally required to keep. Exporting and deleting them is your call.
The HIPAA workspace documentation covers how this works day to day, including the deletion default and the lead rules.
What SiteGPT offers, stated plainly
Pros
- The BAA is published in full at sitegpt.ai/legal/baa, readable before you talk to sales
- That published BAA is the standard document, and reasonable changes are negotiated during onboarding rather than refused
- Breaches reported within five business days of discovery, plus security incident reporting
- Every vendor further down the chain signs its own BAA
- Patient data never used to train AI models, backed by zero-data-retention endpoints at the AI layer, including OpenAI
- Cloud drives and helpdesk tools are off by default rather than refused, and can be switched on for any vendor you hold a BAA with
- SOC 2 Type II, with the trust portal and security documentation published
- Encryption in transit and at rest, plus access controls, promised in the contract rather than only on a marketing page
Cons
- The BAA comes with the Enterprise plan at custom pricing; Starter, Growth, and Scale do not include it
- A self-serve HIPAA plan is planned and is not available today
- Cloud drives (Notion, Google Drive, Dropbox, OneDrive, Box, GitHub) are off unless you hold your own BAA with that vendor
- Helpdesk and chat tools (Zendesk, Slack, Crisp) are off unless you hold your own BAA with that vendor
- Sources already connected to an account are switched off when HIPAA is enabled
- Keeping patient data out of the content you supply is your responsibility, with no automatic detection
- The 'Responses are AI-generated' line stays on the widget, even with white-label branding
The security behind the BAA, including SOC 2 Type II, is documented on the SiteGPT security page. Every other plan has public pricing. Enterprise is custom because the BAA and the retention window are agreed with you rather than sold off a page.
Best forUS healthcare organizations that need a signed BAA and can run a chatbot on content they upload themselves.
If you are still choosing between vendors, the roundup of HIPAA compliant AI chatbots compares the options side by side. This guide sits underneath it: what the requirements are once you have picked.
The checklist to hand your compliance officer
Five questions, in the order that rules vendors out fastest. Ask them in writing.
| Ask the vendor | A good answer | A warning sign |
|---|---|---|
| Will you sign a BAA, and on which plan? | A named plan and a clear yes or no | "Contact sales" with no plan named, or a HIPAA badge and no mention of a BAA |
| Will you share the BAA before a sales call, and will you amend it? | A published document, plus willingness to negotiate reasonable changes | The BAA appears only after you sign, or is strictly take it or leave it |
| What changes in a HIPAA workspace, and what do we use instead? | A specific list, each with its alternative | "Nothing changes", which usually means the program is new |
| How long are chats kept, and is that in the contract? | A stated number, and whether it can move | A vague "you control your data" with no number |
| Is patient data ever used to train models, by you or your suppliers? | A written commitment that covers the suppliers too | A promise covering the vendor but silent on its AI providers |
Three more belong on your own side of the list, because no vendor can do them for you:
- Name an owner for training content. One person approves what the bot is allowed to read.
- Check your lead form fields. Remove or limit anything that invites someone to describe a condition.
- Write down the handoff path. Decide where a human picks up the chat and how quickly.
Can visitors type health information into the widget?
- If yes→You need a BAA before launch, on an Enterprise agreement
- If no, and the form fields prevent it→A standard plan may be sufficient, revisit if scope changes
Does your knowledge base live in Google Drive, Notion, Dropbox, OneDrive, Box, or GitHub?
- If you hold a BAA with them→Ask to enable that specific source during scoping
- If you do not→Upload the same files directly, which is the default path
Do escalations need to land in Zendesk, Slack, or Crisp?
- If you hold a BAA with them→Ask to enable that integration during scoping
- If you do not→Use the chat widget, where takeover works normally
Do you have a retention obligation longer or shorter than 7 days?
- If yes→Set the window in the order form, it is a contract term and not a settings change
- If no→The 7-day redaction default applies from last activity
Frequently asked questions
What makes a chatbot HIPAA compliant? Three things together. First, the vendor signs a Business Associate Agreement (BAA), the contract that lets patient data reach them at all. Second, the service meets HIPAA Security Rule safeguards: encryption in transit and at rest, access controls, audit logs, and a breach notification process. Third, the setup keeps patient data inside chats and out of training content, prompts, and lead forms. A vendor can do the first two and still leave you non-compliant if the third goes wrong, because that part is yours to configure.
Is there such a thing as a HIPAA certified chatbot? No. HIPAA has no certification program and no official certifying body, so no chatbot can be certified HIPAA compliant in any binding sense. Vendors state their own compliance, sometimes backed by an independent audit such as SOC 2 Type II or HITRUST. Treat a 'HIPAA certified' badge as marketing and ask for the two documents that carry legal weight instead: the BAA and the security documentation behind it.
Do you need a BAA for a website chatbot? If the chatbot handles patient data on your behalf, then yes. The vendor is a business associate, and you need a BAA before any patient data reaches them. What decides it is what visitors can type, not what you built the bot for. A chatbot on a clinic website will eventually be asked about a specific appointment, symptom, or prescription, so most healthcare deployments need one even when the bot was only meant to answer general questions.
Which SiteGPT plan includes a HIPAA BAA? The Enterprise plan only, at custom pricing. BAAs are not available on Starter, Growth, or Scale. A self-serve HIPAA tier is planned but is not available today, so any HIPAA deployment currently runs through an Enterprise agreement. The authoritative source is the SiteGPT HIPAA program page, and the standard BAA is published in full for legal review before you talk to anyone.
What changes in a HIPAA covered SiteGPT workspace? Content arrives one deliberate way: you upload files, paste text, or point the bot at your own website, rather than syncing from Notion, Google Drive, Dropbox, OneDrive, Box, or GitHub. Conversations run in SiteGPT's own chat widget rather than flowing into Zendesk, Slack, or Crisp. The "Responses are AI-generated" disclosure stays pinned even on white-labeled widgets. Each of these has an alternative rather than being a dead end, and the cloud drive and helpdesk sources are off by default rather than permanently blocked.
Can you enable Zendesk or Google Drive on a HIPAA workspace? Yes, if you hold your own Business Associate Agreement with that vendor. Tell SiteGPT during scoping and it can enable the specific sources you are covered for. The test is not whether the vendor is HIPAA compliant, since almost every large vendor is, but whether you have a BAA with them. Without one, that vendor is entitled to assume your data contains no PHI, which leaves it free to store and process the content through any subprocessor it chooses. The default is off because the safe assumption is that no such agreement exists.
Can PHI be used to train a chatbot? It should not be, though it is worth being precise about why. Nothing in HIPAA forbids patient data in a knowledge base. Confining PHI to conversations is a contract term you agree to in the BAA, and almost every chat vendor asks for it. The practical reason is stronger than the contractual one: training content is what the bot retrieves from when it answers, so anyone who can chat with the bot can reach it indirectly. A chatbot open to every visitor is the wrong place for one person's medical details regardless of what the contract permits.
How long does a HIPAA chatbot keep conversation data? On SiteGPT, chat text is removed seven days after the last message by default, and you can change that number in your order form. Leads are different and stay until you delete them, because deleting them on a vendor's schedule could destroy records you are legally required to keep. Since the number lives in your contract rather than a settings menu, decide it before you sign.
Does SiteGPT detect or scrub PHI automatically? No. Nothing scans the content you supply, and no chatbot vendor should be trusted to catch it for you. Keeping patient data out of training content, instructions, and lead forms is a review step your team owns. Build it into whoever approves content before it is uploaded, because the bot will read whatever it is given.
What does the SiteGPT BAA actually commit to? HIPAA Security Rule safeguards including encryption in transit and at rest and access controls, breach reporting within five business days of discovery plus security incident reporting, subcontractor flow-down so every vendor in the PHI chain operates under its own BAA, and a commitment that PHI is never used to train AI models, backed by zero-data-retention endpoints at the AI layer including OpenAI. The full text is published at sitegpt.ai/legal/baa so counsel can review it before any sales conversation starts. That published document is the standard BAA, and reasonable amendments are worked in during onboarding rather than refused.
What should you ask a chatbot vendor before signing? Ask five questions in this order. Will you sign a BAA, and on which plan? Will you share the BAA before a sales call, and will you amend it? What changes in a HIPAA workspace, and what do we use instead? How long are chats kept, and is that in the contract? Is patient data ever used to train models, by you or by your suppliers? A vendor that cannot answer the second and third questions in writing has a newer program than its marketing page suggests.
Sources
- SiteGPT HIPAA program page for BAA plan gating and program status, verified 4 August 2026
- SiteGPT standard BAA for breach reporting timelines, subcontractor flow-down, and the model training commitment
- SiteGPT HIPAA workspace documentation for disabled integrations, training content rules, redaction defaults, and lead handling
- SiteGPT security page for SOC 2 Type II and the trust portal
- SiteGPT subprocessor list for the AI subprocessors and zero-data-retention endpoints
- HHS HIPAA Security Rule for the safeguard categories
- HHS business associate guidance for when the relationship is triggered
- HHS breach notification rule for notification obligations
Last updated: August 2026. Program details and plan gating verified against SiteGPT's published HIPAA pages as of 4 August 2026.