For Washington healthcare and legal practices, the biggest AI risks in 2026 are not science-fiction failures – they’re regulated data leaking into consumer chatbots without a Business Associate Agreement, attorneys filing briefs with fabricated citations, and chatbots giving customers wrong information the business is then held to. The good news: every one of these risks has a known mitigation, and most of them are MSP-level operational hygiene.
Why Healthcare and Legal Have an Extra-Heavy AI Risk Profile
Both industries share three traits that turn ordinary AI risk into something more serious. They handle data protected by federal law (HIPAA for clinics, attorney-client privilege and bar ethics rules for firms). They have professional duties – competence, confidentiality, supervision – that already governed how technology gets used. And they sit in front of customers who trust them to know better. When AI fails in a marketing agency, you fix copy. When it fails in a clinic or a law firm, you may have a regulatory event, a malpractice question, or both.
The Western Washington dental, medical, and small-firm market we work with feels this acutely. Owners want AI for charting, transcription, and document review – exactly the use cases that touch the most sensitive data.
What HIPAA Says About AI in 2026
The starting point: consumer ChatGPT, Gemini, and Claude are not HIPAA-eligible. OpenAI, Google, and Anthropic don’t sign Business Associate Agreements for their consumer tiers, and inputs may be used for model training. HIPAA Journal’s running analysis summarizes this clearly. There is no “with a BAA, it’s fine” version of consumer ChatGPT – the consumer tier is simply off-limits for any data touching protected health information.
A BAA is available – but only on specific products. OpenAI offers a BAA for its Zero-Retention API endpoints and for ChatGPT Enterprise and Edu accounts (sales-managed), but not for ChatGPT Business – see OpenAI’s business associate agreement process. Google offers a BAA for Gemini for Workspace under specific configurations, and Microsoft offers a BAA covering Copilot for Microsoft 365 when used inside a properly licensed tenant. The product names matter. A clinic using “ChatGPT” through a browser tab is not the same as a developer using the OpenAI API with a signed BAA, even though they reach the same underlying model.
The proposed HIPAA Security Rule update. On December 27, 2024, HHS/OCR issued a Notice of Proposed Rulemaking to strengthen the HIPAA Security Rule (published in the Federal Register January 6, 2025; comments closed March 7, 2025). It would require AI tools that touch electronic PHI to be included in your formal risk analysis and risk management process, plus baseline cadences – vulnerability scanning at least every six months and penetration testing at least annually. As of May 2026, the rule is still proposed, not final. Separately, HHS’s AI strategy announcement signals HHS’s broader posture on AI in covered entities. Small practices that don’t have semi-annual scanning and annual pen testing in place should plan to, regardless of whether the NPRM is finalized in current form.
The minimum necessary standard still applies. Only PHI strictly necessary for the AI use case may be shared, even with a properly BAA-covered tool. A scribe app doesn’t need your full chart history to summarize today’s visit.
Knowing when AI becomes a Business Associate. A vendor that creates, receives, maintains, or transmits PHI on your behalf is a Business Associate and needs a BAA. Paubox’s overview of AI and BAA triggers covers the patterns – scheduling assistants, transcription tools, diagnostic helpers, patient communication bots. When you do sign a BAA, look for explicit language on no training of public models with your data, breach notification within 60 days, and encryption requirements. Our HIPAA IT requirements for Washington post covers the broader compliance picture, and the penalties post covers what’s at stake.
What the Bar Associations Say About AI for Lawyers
For law firms, the foundational document is ABA Formal Opinion 512, issued July 29, 2024 – the first formal ABA guidance on generative AI. It walks through the duties affected by AI use: competence, confidentiality, communication, fees, candor toward the tribunal, and supervisory responsibilities. The hook is Model Rule 1.1 Comment 8, which makes “keeping abreast of the benefits and risks of relevant technology” part of the duty of competence. (ABA opinions are guidance, not binding law – but state bars and courts treat them seriously.)
In Washington specifically, WSBA Advisory Opinion 2025-05 addresses lawyers’ ethics duties when using AI-enabled tools – competence, diligence, confidentiality, communication, candor, supervision, and fees – and is the central piece of guidance for WA attorneys evaluating AI workflows. Washington-licensed attorneys should read it in full at the WSBA Committee on Professional Ethics page. (Advisory opinion URLs occasionally change; verify the current link at the WSBA site before citing in a brief.)
The Liability Lessons Everyone Should Read Once
Four cases summarize the operational risk so far.
Mata v. Avianca (2023). Two New York attorneys filed a brief drafted with ChatGPT containing six fabricated case citations. The court sanctioned the two attorneys and their firm jointly $5,000, and the case became the canonical example of “do not file what you didn’t verify.” Original filing was 2022; sanctions came down June 2023.
Utah Court of Appeals (2025). Attorney Richard Bednar was sanctioned for filing a brief with fake ChatGPT citations, ordered to refund client fees and donate $1,000 to a legal nonprofit. The Mata pattern is not a one-off.
California (2025). Judge Wilner fined two firms a combined $31,000 for undisclosed AI use producing fabricated citations. The disclosure issue mattered as much as the citations themselves.
Air Canada (BC Civil Resolution Tribunal, February 14, 2024). Air Canada was held liable for incorrect bereavement-fare advice from its own customer-service chatbot – an award of approximately CAD $812 (damages plus interest and filing fees). Small money, large precedent: a business is responsible for what its AI tells customers. (Note: BC tribunal, not U.S. binding precedent, but cited widely as the first clear ruling.) Relevant for any clinic or firm running a chatbot on its website.
The common thread isn’t that AI is dangerous – it’s that you remain responsible for the output. AI is not a defense.
What FTC Enforcement Means Even If You’re Not in Healthcare or Law
Cross-cutting risk: the FTC launched Operation AI Comply in September 2024, an ongoing initiative that has produced more than a dozen enforcement actions for deceptive AI claims and “AI-washing” since. The DoNotPay “AI Lawyer” settlement is the most relevant for small professional-services firms – the FTC took action over claims that an AI tool could perform legal services without human supervision. If your marketing copy says your AI does something it can’t reliably do, the FTC has signaled it will treat that as a deceptive practice. Quietly clean up your website AI claims before someone else does it for you.
What an MSP Should Be Doing for You by Default
What we typically do for healthcare and legal clients across Western Washington – none of it exotic, all of it boring-but-effective:
- Block consumer-tier AI from devices that touch regulated data using DNS filtering and Microsoft Defender for Cloud Apps, and gate enterprise AI logins through Entra ID conditional access.
- Maintain an approved-tools list of AI products with active BAAs (or, for firms, with appropriate confidentiality terms) – and a documented “block by default” stance for everything else.
- Separate logins for consumer and work. No personal ChatGPT signed in on a work device. Period.
- AI note-taker disclosure. If a scribe or transcription tool is in use during a patient visit or client consultation, it gets disclosed and documented. This is both an ethics matter and increasingly an industry expectation.
- Document the BAA chain. Maintain a register of every AI vendor, the BAA on file, expiration date, and the data classification each tool is approved to handle.
- Tie everything to the AI Acceptable Use Policy and the broader shadow AI program so your written controls and your technical controls match.
For Pacific Northwest industry-specific managed IT, this is the difference between AI as a productivity tool and AI as a compliance event.
This article is not legal advice. Run your AI policies and tool choices past industry-specific counsel before finalizing – bar ethics rules and HIPAA risk analyses are not a DIY project.
ROI Technology Inc. supports HIPAA-covered clinics and small law firms across Western Washington with the practical controls behind AI policy. Contact us or call (888) 707-3652 for an AI risk and BAA review of your practice.