Is Claude HIPAA Compliant in 2026? Not on the Plan You're Probably Using

Every few weeks someone asks me a version of the same question, usually a little sheepishly, usually after the fact: we've been using ChatGPT to draft patient follow-ups — is that a problem?

The internet answers this badly. Search it and you'll get a wall of confident yes/no verdicts, most of them written by companies selling a "HIPAA-compliant AI wrapper," most of them skipping the one distinction that actually decides your answer. So here it is first, before anything else.

No model is HIPAA compliant. A contract is.

HIPAA compliance is not a property that software has. There is no certification body that inspects a language model and stamps it. What exists is a Business Associate Agreement — a contract in which a vendor handling protected health information on your behalf accepts legal obligations for safeguarding it.

So "is Claude HIPAA compliant?" is the wrong shape of question. The right one is: will this vendor sign a BAA covering the specific product I am about to paste a patient's name into? That question has a precise, checkable answer, and it changes tier by tier within the same brand.

Which is why the yes/no articles are so useless. The answer for Claude is genuinely both. It depends entirely on which Claude you're using — and almost certainly not the one you're using.

What Anthropic will and won't sign for

Anthropic's published BAA documentation draws the line in two places:

Two details in that table do real damage if you miss them.

Self-serve Enterprise is not the same product as sales-assisted Enterprise. Reaching the top of the pricing page with a credit card does not put you in BAA territory. The HIPAA-ready plan is the one you get through a conversation with their sales team, and buying the expensive tier yourself is not a shortcut to it.

And HIPAA-ready Enterprise is not automatic once you have it. Anthropic requires the organisation's Primary Owner to actively switch on HIPAA compliance in the workspace's Data and privacy settings and accept the BAA. A standard Enterprise organisation is not covered until somebody does that. If you're on Enterprise today and nobody can tell you who accepted the BAA or when, assume it hasn't happened.

One more piece of housekeeping worth checking if you've been around a while: Anthropic has stated that BAAs signed before 2 December 2025 cover API usage only and do not extend to the HIPAA-ready Enterprise plan. An old agreement in a drawer may not cover what you think it covers.

Is ChatGPT HIPAA compliant? Same answer, different tier names

OpenAI draws the line in the same conceptual place, and the tiers that make it across are the enterprise-and-API ones:

• Covered by a BAA: ChatGPT Enterprise (the sales-managed one), OpenAI's healthcare-specific enterprise products, and the API Platform — and notably the API does not require an enterprise agreement to get a BAA, which matters enormously for small practices.
• Not covered: Free, Plus, Pro, Team, and self-serve ChatGPT Business.

The naming here is genuinely hazardous. OpenAI ships a consumer-facing health product whose name contains the word "health" and which is not BAA-covered, sitting in a catalogue alongside enterprise healthcare products that are. If your compliance reasoning is "well, it's called Health, so presumably…" — stop there. The name is marketing; the BAA is the fact. If you're weighing the two vendors more broadly, Anthropic vs OpenAI for business covers the rest of the comparison.

The plans that trip everyone up: Claude Team and ChatGPT Business

If you take one thing from this page, take this one, because it is the mistake I see most and it is an expensive one.

Claude Team and ChatGPT Business are not HIPAA-eligible.

These are precisely the plans a five-person practice talks itself into. They cost real money. They say team and business on the label. They come with admin controls, shared workspaces, and a contractual promise not to train on your data — and that last one is what does the damage, because "they don't train on our data" feels like the compliance box being ticked.

It isn't. Training policy and BAA coverage are two unrelated things. You can be on a plan that will never train on your inputs and still have no agreement in place governing protected health information, which means every PHI-containing prompt is a disclosure to a vendor who never accepted responsibility for it. Upgrading from Pro to Team moved you into better admin tooling and left your HIPAA position exactly where it was. (I've written separately about what the Claude Team plan actually buys you — it's a good plan. It is not this.)

Signing a BAA doesn't make you compliant either

The other half of the trap sits on the far side of the contract. A BAA is a shared-responsibility document. It covers the vendor's obligations. It says nothing about yours.

Still entirely on you, after signature:

• Access control — who in the practice can open the workspace, and what happens to that access the week someone leaves.
• Minimum necessary — whether the prompt needed the patient's full name and date of birth to do the job, or whether that was just how the front desk happened to paste it.
• Staff training — because the actual breach vector is somebody using their personal ChatGPT on their phone at lunch, not your carefully configured enterprise workspace.
• Breach detection and notification — you need to be able to tell that something went into the wrong tool, and you're on the clock once you know.
• Documentation — a written policy naming which tools are approved for what. If it isn't written down, it doesn't exist in an audit.

A signed BAA is a starting gun, not a finish line.

The cheapest legal answer is usually "don't send PHI at all"

Here's the part the compliance-wrapper vendors would rather not lead with. A large share of what a practice actually wants AI for does not require protected health information in the first place.

Drafting a recall campaign. Explaining what a benefits breakdown means in plain English. Writing a procedure explainer for the website. Turning a messy claims narrative into a reusable template. Rewriting your no-show policy so it sounds like a human wrote it. None of those need a name, a date of birth, or a chart number to work — and the moment they don't, the entire HIPAA question stops applying to that workflow.

The micro-action, and you can do it today: open whatever AI tool your practice is already using, look at the last ten prompts your team ran, and mark each one needed PHI or didn't. In every practice I've done this with, the second pile is much bigger than anyone expected. That exercise takes fifteen minutes and it tells you whether you have a contract problem or a habit problem — and they have very different price tags.

If the workflow genuinely does need PHI, price the API route with a BAA next to the Enterprise quote before you sign anything. For a small practice running modest volume, metered API usage under a BAA is frequently far cheaper than a per-seat enterprise agreement sized for a hospital — and it's the option nobody's sales team is incentivised to mention. For how the consumer tiers price out by comparison, Claude's pricing breakdown has the full ladder.

What I tell owners

Almost nobody who asks me this question has a HIPAA problem. They have an unwritten-rules problem that a HIPAA question finally surfaced.

Staff adopted a tool because it made a genuinely tedious job faster, nobody said not to, and eighteen months later there's no policy, no approved-tools list, no idea which subscriptions are personal versus practice-owned, and no one who can say with confidence what has been pasted where. The compliance exposure is real, but it's a symptom. The disease is that AI arrived in the practice the way a stray cat arrives at a back door — incrementally, affectionately, and with nobody making a decision.

The fix is unglamorous and it is mostly not a purchase: decide which two or three workflows AI is actually for, write down which tool is approved for each and what may never be pasted into any of them, and de-identify by default so the expensive contract question only applies to the small number of workflows that truly need it.

If you'd rather not work out that order yourself, the AI Readiness Assessment ($2,500 flat — two weeks, a written roadmap, fee credited against a build within 90 days) is exactly this exercise done properly: which processes, in what order, on which tools, with the governance written down instead of assumed. The do-it-yourself version is the CFSB Playbook ($397). And if you want to talk through a specific practice, that's what strategy calls are for. Practice-specific workflow examples live in AI for dental practices.

One last thing, and I'd rather say it plainly than hedge it: this is a technology explainer written by a fractional CTO who implements these systems, not legal advice. HIPAA enforcement turns on specifics — your data, your workflows, your risk analysis. Use this to ask your compliance counsel a much sharper question than "is ChatGPT okay?", which is a question nobody can answer well. "Which of our AI tools are covered by a signed BAA, and who accepted it?" is the one worth asking.