When Anthropic launched Claude for Small Business earlier this month, I did a word search of the announcement for HIPAA and BAA. Zero hits for either. That absence is the most important thing for any healthcare SMB to understand about this launch.
It does not mean the product is unsafe for healthcare. It does not mean Anthropic does not offer BAAs. Anthropic does offer BAAs on its Enterprise plan. What it means is that the SMB-tier offering as announced has not explicitly addressed HIPAA, and any healthcare practice considering it needs to ask specific questions before deploying. For the broader context on what Anthropic shipped, see our first blog on the topic. For the regulated financial services side of the same analysis, see our post on AI for financial services SMBs.
This post walks through what HIPAA actually requires from an AI vendor, what a BAA is and is not, and how the calculus differs across dental and DSO, behavioral health, ambulatory and specialty practices, and medspas. The fourth category is its own thing because most medspas sit outside the federal covered-entity definition for the majority of their workflows, which changes the analysis.
What HIPAA Actually Requires From an AI Vendor
At the risk of oversimplifying: if a third party touches your Protected Health Information (PHI), they need to be your Business Associate, and you need a Business Associate Agreement (BAA) with them. That is the structural rule.
An AI vendor that ingests PHI in the course of providing its service is a business associate. There is no exception for vendors that say they do not store the data. There is no exception for transient processing. OCR has been consistent on this since the HIPAA Omnibus Rule in 2013, which made business associates directly liable under HIPAA and pulled their subcontractors into scope. The introduction of AI services has not changed the underlying analysis.
So the question is binary: does your AI vendor offer a BAA on the specific plan you are using, and have you signed one?
For Anthropic specifically: BAAs are available on Enterprise plans through a click-to-accept HIPAA-ready configuration. The published policy is explicit that the BAA does not cover Cowork, among other surfaces. Claude for Small Business deploys inside Claude Cowork. Until Anthropic updates that coverage map, treat Claude for Small Business as a non-BAA service for healthcare purposes. The product may be fine for non-PHI workflows. It is not the right surface for PHI today.
The BAA Questions
Three things to ask any AI vendor you are considering for healthcare workflows.
Do you offer a BAA at this price tier, or only on enterprise plans?
Does the BAA cover all the features I plan to use, including connectors to third-party tools (which may bring additional sub-processors into scope)?
What sub-processors are covered under your BAA, and how am I notified if you add or remove one?
The third question matters more than people realize. If your AI vendor uses sub-processors for, say, document parsing or vector search, those sub-processors must also be appropriately covered. You are responsible for the chain. Our breakdown of the 18 technical safeguards in the HIPAA Security Rule walks through where vendor controls show up in the audit posture.
Dental Practices and DSOs
Dental is one of the highest-value AI use cases in healthcare SMB, partly because the workflows split cleanly between PHI-touching and non-PHI. Dental practices are also under more direct cyber pressure than most clinical segments, as we covered in Why Dental Practices are the #1 Target for Modern Cybercriminals.
Non-PHI Workflows: Lower Risk, Deploy First
- Accounting. AR chase against insurance and patient payment plans. Outstanding balance tracking is mostly identifiers and dollar amounts. Whether it counts as PHI depends on whether the data is tied to treatment information. Often it is not, in which case general invoice-chasing workflows apply.
- Practice marketing. New patient campaigns, recall outreach to patients who have not been seen in 18 months, social content. The patient list itself is PHI, but the marketing message generation is not, if you keep the data separated.
- Appointment reminders. PHI is in the patient identifier and appointment metadata. Requires BAA coverage.
- Internal team operations. Payroll, HR, vendor management, internal training. No PHI involved.
PHI-Touching Workflows: BAA Required, Deploy With Caution
- Treatment planning summaries, clinical chart notes, and insurance pre-authorization narratives. These require explicit BAA coverage and careful workflow design.
- Patient-facing communication that references specific treatments or diagnoses.
For DSOs Specifically
If you run a Dental Service Organization, the management services agreement structure matters here. PHI flows through the clinical practice entity, and the DSO management entity is typically a business associate of the practice. AI tools deployed at the management entity level need to fit inside that structure. Walk your existing BA chain. Confirm any new AI vendor extends through it.
Behavioral Health
Behavioral health has the highest PHI sensitivity in the entire healthcare SMB landscape. Two regulatory layers stack.
First, standard HIPAA. Same rules as any other covered entity.
Second, 42 CFR Part 2. If your practice qualifies as federally assisted and provides substance use disorder treatment, Part 2 imposes confidentiality protections substantially stricter than HIPAA. The federally assisted test is broad. It includes Medicare and Medicaid participation, tax-exempt status, and federal authorization, not just direct federal grants, which means most SUD treatment providers in the US are inside Part 2. Disclosure rules are different. Consent requirements are different. Penalties are different.
For Part 2-covered practices, the practical rule on AI is: no PHI touches an AI service unless the service is explicitly covered under both a HIPAA BAA and a Part 2 qualified service organization agreement. That is a narrow field today.
Where Behavioral Health Practices Can Use AI Now
- Practice operations: scheduling logistics, billing operations, payer interactions on benefit verification, payroll, vendor management.
- Marketing and content, with care around any patient testimonials or identifying language.
- Internal staff training scenarios using synthetic cases, not real patient data.
What I would not do yet in a behavioral health practice without explicit, signed, sub-processor-aware BAA coverage: any clinical note generation, treatment summarization, insurance authorization narratives, or patient communication drafting.
Ambulatory and Specialty Practices
Ambulatory care, specialty groups, and small physician practices sit somewhere between dental and behavioral health on the sensitivity spectrum. Same HIPAA framework, less acute than Part 2. For New York practices, the NY SHIELD Act adds reporting and safeguards obligations on top of HIPAA, so the state layer matters too.
Two specific considerations for this segment.
EHR Integration
Most ambulatory practices run an EHR (Epic, eClinicalWorks, Athena, NextGen, others). The Claude for Small Business package as announced integrates with QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365. Not your EHR. Which means anything that needs EHR data is going to require either copy-paste (workflow-fragile and PHI-risky), an additional integration layer, or a different tool entirely.
For administrative workflows that sit outside the EHR, the existing connectors are fine.
Device and Medical Billing Data
If your practice handles claims data through clearinghouses, billing data has its own data governance considerations. Most billing data is PHI. BAA coverage applies. Same playbook as dental.
Medspas: The Exception (With Caveats)
Here is where the analysis changes. Many medspas (medical aesthetic practices) are not covered entities under HIPAA for the bulk of their services, but the exception is narrower than is sometimes assumed.
HIPAA applies to covered entities, which include healthcare providers who transmit health information electronically in connection with certain HIPAA standard transactions, primarily insurance claims and related electronic exchanges. A medspa that does not bill insurance and operates purely on cash-pay aesthetic services is often outside the covered-entity definition under federal law. We have written more on the operational side of growing a medspa in MedSpa Growth: Scaling to Multiple Locations Without Losing Sight of Security.
That matters because it changes the AI compliance calculus. A medspa that sits outside the covered-entity test can deploy AI workflows for client communications, marketing, appointment scheduling, and even some clinical-adjacent content without the federal HIPAA overlay. State-level privacy laws still apply. Confidentiality obligations under state cosmetic medicine rules still apply. But the federal HIPAA framework largely does not.
Three cautions specific to medspas.
The Injectables and Electronic Prescription Nuance
If your practice prescribes Botox, fillers, GLP-1s, or any other prescription product, you are likely transmitting prescription data electronically in some form. Depending on how that data flows and which clearinghouses or e-prescribing systems are in the path, you can find yourself meeting the HIPAA standard-transaction test even on a cash-pay book of business. The covered-entity question is service-line specific, not entity-wide, and the safe assumption for any medspa offering injectables is that some workflows touch HIPAA. Talk to counsel before assuming you are exempt. Our blog post on securing peptides and regenerative medicine data covers the adjacent category where this is most acute.
The FDA Marketing Claims Trap
Medspa marketing copy is a legal minefield. The FDA aggressively enforces against unapproved drug claims, off-label promotion, and overstated efficacy claims. AI-generated marketing copy will pattern-match against successful marketing language found online, which is often non-compliant. If you let an AI draft promotional content for injectables, lasers, body contouring, or any device-based service, you need a robust review process to catch FDA-problematic claims before publication. The April 2026 warning letter to Pure Indulgence Aesthetics in Southlake, TX is the most recent reminder that the FDA is paying attention to this category.
Common AI failure modes: claims of permanent results, before-and-after framing without proper disclosures, comparison claims against competitor products, and weight-loss numbers that imply clinical guarantees. Any of these can trigger an FDA warning letter or a state board complaint.
When a Medspa is a Covered Entity For The Full Enterprise
If your medspa has any service line that involves insurance billing (some BHRT services, some weight-loss medications, some out-of-network plastic surgery), the practice as a whole may be a covered entity. Talk to counsel before assuming you are exempt. The exemption is service-line specific, not entity-wide.
Cross-Cutting: Safe Workflows for Healthcare SMBs
Across all four sub-verticals, here are the workflows where the risk-adjusted return is highest today.
AR chase against insurance and patient balances. Where dollar amounts are tied to identifiers but not clinical details, in many practices this is borderline PHI or non-PHI. Check with counsel. Often safe to pilot with BAA in place.
Practice marketing and patient recall communications, with PHI not leaving the patient identifier and contact info layer.
Internal operations: payroll, vendor management, staff training, and financial reporting. No PHI involved.
Contract review for vendor contracts and employment agreements.
Tax-season document organization for the practice's own books.
Cross-Cutting: Unsafe Workflows Without Specific Safeguards
Clinical note generation or summarization without BAA coverage and clinician-in-the-loop review.
Patient-facing clinical communication drafting without strict review gates.
Insurance pre-authorization or claims narrative generation without BAA coverage.
Anything involving Part 2-covered SUD treatment data outside a fully qualified service organization arrangement.
1AI-generated medspa marketing claims without legal or compliance review.
A Healthcare SMB Checklist Before Deploying Any AI Workflow
Have you confirmed BAA availability at the specific price tier and feature set you are deploying?
Have you mapped which workflows touch PHI and which do not?
If Part 2 applies, have you confirmed qualified service organization status?
Have you confirmed audit logging is sufficient for your OCR exam posture?
Have you documented the AI vendor in your security risk analysis?
Have you updated patient notices or practice consents if patient data flows through new vendors?
Have you trained your team on what they can and cannot put into the AI?
If you cannot answer all seven, pause and resolve them before going live.
If you want to talk through how a Managed AI program could be scoped for your practice with the right BAA chain, audit trail, and review gates, or to walk through our broader healthcare IT program, reach out to the Techvera Team for a consultation.
Disclosure: Techvera is an MSP serving small and medium businesses across North Texas (Denton) and New York, including healthcare practices in our customer base. Our internal operations are powered in part by Anthropic's Claude. Nothing in this post constitutes legal advice or HIPAA compliance guidance. Healthcare AI compliance requires consultation with qualified counsel and your privacy officer. The analysis above reflects publicly available information as of May 2026 and may not reflect later changes from OCR, FDA, or state regulators.
About the Author
Todd Mitchell
Chief Operating Officer
Todd Mitchell is the COO of Techvera, bringing operational expertise and strategic vision to help businesses transform their IT infrastructure.