Should Professional AI Have to Pass an Exam?

This post is informational and reflects the state of the law as of August 2026. For your situation, consult a licensed attorney.

The FDA is floating the idea that a medical AI product should have to pass something like a board exam before it ships.

On August 18, 2026, the agency’s Center for Devices and Radiological Health published a discussion paper sketching a competency-based approach to premarket review of generative AI devices, modeled at a high level on how clinicians are evaluated and credentialed. Benchmarking on clinical knowledge, safety behavior, communication, and generalizability, then confirmation in real or clinically representative use.

The agency is careful about what this is. The paper is not guidance, does not change policy, does not set expectations for future submissions, and does not address whether any of it sits inside the FDA’s existing authority. It is a request for comment.

It is worth the attention of anyone building professional-services AI anyway, because the structure travels well past medicine.

Three ideas that generalize

The unit of review is the product, not the model. The paper evaluates the final user-facing system as configured for deployment, with its prompts, retrieval, guardrails, and orchestration logic. Most AI assurance work today sits a layer deeper, at the model. That is a meaningful relocation of the question, and it matches how harm actually reaches a user.

Risk runs on two axes. How far the output goes toward directing or taking an action, and how severe the harm is if someone relies on it and it is wrong. The paper walks a single clinical fact through four phrasings, from a general statement that dosages are sometimes increased to a specific instruction to increase a named drug from 10 mg to 20 mg daily, and treats the distance between those as a difference in kind.

A disclaimer may not change the risk tier. CDRH asks whether a patient-facing output “may not become any less directive because it includes a ‘talk to your doctor’ or an ‘I am not a medical professional’ statement.”

Most legal, tax, and financial AI products carry some version of that line. It is worth asking honestly whether yours is doing work or performing work.

Why this lands close to home

Law, accounting, engineering, and financial advice all regulate AI-assisted judgment already, through licensure, supervision rules, unauthorized-practice limits, and malpractice exposure. Every one of those gates sits on the person who signs the work. The tool itself stays outside the frame.

That allocation is coherent and it is also incomplete. A product that tells a small business owner whether to sign a lease is making a recommendation with real consequences, and the professional-responsibility apparatus reaches it only if a licensed person is in the loop and accountable. Increasingly, there is not one.

This is the first serious government attempt I have seen to run the assurance question at the product rather than at the professional, and the questions translate almost directly. If a competency exam is the right gate for a product that adjusts an insulin dose, what is the right gate for one that tells a business owner whether to sign?

What to do with this now

If you build or buy AI for professional workflows, the paper is a usable checklist even with no regulator behind it in your sector:

Classify your outputs by how directive they are, not by how the marketing describes them.
Test the deployed configuration, not the underlying model.
Treat disclaimers as a disclosure question, separate from the risk-tier question.
Write down what the product should refuse to do, and test that it refuses.

Comments on the discussion paper are open into the fall, and the docket is worth a look if your product sits anywhere near clinical, legal, or financial advice.

If you are building or deploying AI that gives professional-grade recommendations, talk to MB Legal about where the assurance and disclosure obligations land. See also before you prompt: AI, privilege, and confidentiality and the firm’s AI and emerging technology practice.

This post is informational and does not constitute legal advice. The FDA discussion paper described is a request for comment and does not change any legal requirement. For guidance on a specific situation, consult a licensed attorney.