Is Claude HIPAA Compliant? What a Bedrock BAA Covers

Short answer: it depends on which door you use to reach the model. Consumer Claude.ai, meaning the Free and Pro plans, is not covered by a business associate agreement, so it is the wrong place for anything that touches patient data; Claude accessed through Amazon Bedrock is a HIPAA-eligible AWS service and sits under the BAA that AWS signs with its customers. That second path is how Curve runs Curve Analyst, the AI chat inside the Curve dashboard. It is also only half of the question. The other half is where your data lives and what the application wrapped around the model is permitted to do with it, and that half gets far less attention than the model's name does.

TL;DR

• Consumer Claude.ai (Free, Pro) is not covered by a BAA. Do not put patient data into it.
• Anthropic makes BAAs available for certain API and enterprise arrangements. Confirm the current terms with the vendor before you rely on that.
• Claude on Amazon Bedrock is a HIPAA-eligible AWS service. AWS signs a BAA covering eligible services, so the model runs inside the same agreement that already covers the rest of an AWS environment.
• A BAA assigns responsibility. It does not make a careless workflow compliant, and it says nothing about whether the application should be allowed to look up individuals or export data.
• Curve Analyst runs on Claude through Bedrock, inside infrastructure covered by Curve's BAA with AWS, and Curve signs a BAA with every customer on every plan. Your data is not sent to a separate AI product and is not used for training.

Model, vendor, and deployment path are three different questions

"Is Claude HIPAA compliant?" cannot be answered as asked, because a model is not the kind of thing that signs contracts. When people ask it, they are usually collapsing separate questions into one.

The model

Claude is a set of weights and the software that runs them. On its own it has no compliance status, the same way a database engine has none. The identical model can be fine in one setting and a reportable problem in another.

The vendor

Anthropic builds Claude and sells access to it in several forms. The consumer product, Claude.ai on the Free and Pro plans, is not covered by a BAA. Anthropic does make BAAs available for certain API and enterprise arrangements. If you plan to go that route, get the current terms in writing from the vendor.

The deployment path

You can also reach Claude without contracting with Anthropic directly. Amazon Bedrock hosts Claude models as a managed AWS service, and Bedrock is one of the services AWS covers under its BAA. Your agreement is with AWS, the model runs inside AWS, and the data handling terms are the ones already in place for the rest of the AWS footprint. For a company whose infrastructure already sits under an AWS BAA, that is a much shorter conversation than adding another vendor to the pile.

So the honest answer to the search query is: the consumer app, no; the Bedrock path, yes at the infrastructure layer; the direct API, ask Anthropic. Then keep reading, because a yes at the infrastructure layer is the first line of the checklist.

What "HIPAA-eligible service" means at AWS

Eligible is a careful word. AWS signs a BAA that covers a defined set of services, and Bedrock is one of them. That means AWS has agreed to take on the obligations of a business associate for PHI processed through Bedrock, under the conditions of that agreement. It does not mean AWS has reviewed your application, your access controls, or your prompts.

Two practical consequences follow. The BAA attaches to the account that holds it. If your vendor runs Claude through Bedrock inside their own AWS account, their BAA with AWS covers that layer, and your BAA with the vendor is what covers you. And eligibility applies to the service, never to whatever gets bolted onto it. If an application reads PHI out of a model response and writes it to a spreadsheet on someone's laptop, the eligible service did its job and the application broke the chain anyway.

Covered is also different from necessary. A marketing analytics tool has no reason to be reading patient records in the first place, and no BAA changes that.

What a BAA covers, and what it does not

A BAA is a contract that assigns responsibility. Under HIPAA, any vendor that handles PHI on a covered entity's behalf is a business associate and needs one. The agreement puts a name on who is accountable for the data while it is in someone else's hands.

What it does

It makes it lawful for a vendor to touch PHI on your behalf, and it passes the same obligations down the chain, which is why Curve's BAA with AWS matters to you even though you never signed anything with AWS. It also gives you someone to hold responsible if the vendor's side of the arrangement fails.

What it does not do

It does not make a bad workflow compliant. A BAA with a chat vendor does not fix a process where a front desk coordinator pastes a patient's intake form into a chat window to get a summary. It does not mean the tool has been configured sensibly, that access is limited to the people who need it, or that the output is going somewhere appropriate. It does not stop an application from doing something it should never have been allowed to do, like answering "which patients booked after clicking the fertility ad?" A BAA is silent on product behavior. That is the application's job, and it is where you should spend most of your evaluation time.

HIPAA is also not the only route to trouble. The FTC has pursued health data disclosures to ad platforms on unfairness grounds, as in the Hims & Hers complaint, and a BAA is no defense against a complaint that your marketing stack shared what it should not have.

The product controls that matter more than the model

If you are evaluating any AI analytics tool for a practice, hospital, telehealth company, or agency that serves them, these are the questions that separate a tool you can defend from one you cannot. The model name does not appear on this list.

Reads, never writes

An assistant that can read reports is a reporting tool. One that can also edit tracking, change which conversions go to Google, or delete a destination is an admin console with a chat interface, and it should be reviewed like one. We built Curve Analyst read-only. It cannot change tracking, edit destinations, or take actions of any kind. If it answers a question wrong, the damage is a wrong answer, and you will see it in the chart before it reaches a slide deck.

No training on your data

Ask the vendor in plain words whether your account's data is used to train or tune any model, theirs or someone else's. Then ask where that is written down. Consumer ChatGPT may use conversations for training unless the user opts out, which is one of the reasons the consumer tier is a nonstarter for healthcare. Curve does not use customer data to train models, and the questions you ask Analyst do not leave the infrastructure covered by the AWS BAA to go to a separate AI product.

No export to a separate AI product

A lot of "AI for your analytics" is an integration that copies your data into someone else's platform so their model can read it. Every copy is a new place PHI can sit. The alternative is to run the model inside the environment where the data already lives and send it only what the current question needs. Analyst works the second way: it queries the account's own data for each answer, and nothing gets bulk exported to make that possible.

No lookups of individuals

This is the control that most directly reflects what a marketing tool is for. Aggregate questions ("how many consult requests came from paid search last month?") are the job. Individual questions ("who submitted the form on Tuesday afternoon?") are not, and a marketing assistant that can answer them is a discovery problem waiting to happen. Analyst does not look up individuals and does not answer questions about a specific person. It also cannot see session recordings or heatmaps, which are the two features in Curve most likely to show one person's behavior.

Honesty about missing data

Nobody puts this on a compliance checklist, and it belongs there. An assistant that fills gaps with plausible numbers is a liability the moment a practice owner repeats one of those numbers to a partner or a lender. Analyst answers only from a query against the account. It never estimates. If a metric is empty it says so, if data is stale, partial, sampled, or still syncing it says that before the number, and it treats "unknown" and "zero" as different answers. Ask which conversion figure it is quoting and it will tell you whether that number is Curve's own attribution or what Google or Meta reported, because those two will not match.

How Curve Analyst is built on this stack

The compliance story for Analyst is mostly the compliance story for Curve, with a model added at the end of the chain. Events from your website and forms go to Curve's US hosted infrastructure first, server side. Curve strips what should not leave, then forwards conversions to Google Ads, Meta, TikTok, Microsoft, LinkedIn, and the rest. Curve signs a BAA with every customer on every plan. If you want the longer version of why that foundation is what makes an AI answer possible at all, read how compliant tracking makes AI answers possible.

Analyst sits on top of that. It runs on Claude through Amazon Bedrock, inside infrastructure covered by Curve's BAA with AWS. When you ask a question, it queries the account's own analytics, goal, funnel, and campaign data, and answers with text plus a rendered chart: a line chart for a trend, a table for a ranked list, a stat card for a total, a step chart for a funnel. It keeps the date range and filters of the screen you opened it from, so "this month" means the month you were already looking at.

What it cannot do is as deliberate as what it can: no SQL, no actions, no patient lookups, no session recordings or heatmaps, and no answers yet about tracking configuration or connector health (that part is planned). Every one of those is a boundary we chose so that the answer to "what is the tool allowed to do with the data?" fits in a paragraph you can hand to a compliance officer. A fuller description of the feature is in what Curve Analyst is.

Choosing Bedrock over a consumer app or a separate AI vendor was a decision about where the BAA chain is simplest to draw. Curve already holds a BAA with AWS for its infrastructure; putting the model under that same agreement means no additional vendor handles your data for the AI feature and there is no new agreement for you to chase.

What to ask Curve Analyst

Once the compliance question is settled, the useful test is whether the tool answers the questions you actually get asked on a Monday morning. Open Analyst from the report you are already looking at and try these, exactly as written:

• How did our goal conversions this month compare to last month?
• Which campaign drove the most consults?
• What were our top landing pages from paid traffic?

Each comes back as text with a chart, and each will tell you whether it is quoting Curve's attributed conversions or the platform's own count. The launch announcement, introducing Curve Analyst, walks through more examples.

Frequently asked questions

Is Claude HIPAA compliant?

Not as a standalone product, and not as a question with a single answer. Consumer Claude.ai (Free and Pro) is not covered by a BAA. Claude through Amazon Bedrock is an AWS service covered by the AWS BAA. Anthropic offers BAAs for certain API and enterprise arrangements; confirm the current terms with the vendor. In every case, the application built around the model determines whether your use is defensible.

Does a Bedrock BAA cover my whole application?

No. It covers the eligible AWS services in the account that holds the BAA, under the terms of that agreement. Your application code, your access controls, what the tool is allowed to answer, and where the output goes are your responsibility or your vendor's. If your vendor runs the model in their AWS account, your protection comes from your BAA with the vendor, so read that one closely.

Can I paste patient funnel data into ChatGPT or Claude.ai to analyze it?

Not on the consumer plans, which have no BAA, and consumer ChatGPT may use your conversations for training unless you opt out. The longer answer is in why you cannot paste your patient funnel into ChatGPT.

Does Curve Analyst send my data to a separate AI product?

No. Analyst runs on Claude through Amazon Bedrock inside infrastructure covered by Curve's BAA with AWS. Customer data is not exported to a separate AI product and is not used to train models. Each answer is produced from a query against your account's own data for that specific question.

What can Curve Analyst not tell me?

It will not identify or describe an individual. It does not see session recordings or heatmaps. It cannot change tracking or destinations. And it does not yet answer about tracking configuration or connector health, which is on the roadmap. If a number is missing, stale, or still syncing, it tells you that instead of estimating.

If you want an AI answer about your marketing that you can show a compliance officer without a preamble, the fastest way to judge it is to watch it run on real data. Book a demo of Curve and bring the question your practice owner asked you last.