Read this first
I haven’t worked in these four industries. This page is research, and each section says where it came from. Check it with your own compliance, regulatory or legal team before you act on it.
The industries share one problem: a wrong sentence costs more than a slow week. So the first job is a short list of things AI is allowed to say, and a person who signs off before anything goes live. Speed comes after.
Medical devices
Start with an approved-claims library. A table of every claim the company makes, with its evidence, the labelling or clearance it matches, and the date it was approved. AI drafts only from that table and flags any sentence that is not in it. Retired claims are marked, not deleted.
Who checks: regulatory affairs signs off each claim. Customer complaints go to quality, not to marketing.
Watch for: claims beyond the cleared use, comparisons without data, and complaints handled by the wrong team.
Who regulates the advertising? It depends on the device. One source says the FDA mainly covers prescription and restricted devices, with the FTC leading on the rest. Another describes the split differently. Ask regulatory affairs which applies to each product. I found no rule that treats AI-written claims differently, so the company owns what the tool writes.
Manufacturing
Start with product data turned into web content, then triage incoming requests for quote (RFQs). AI turns your own spec sheets into technical pages, and pulls material, quantity, tolerance and certification needs out of an RFQ. It compares them with a plain list of what you can make and drafts questions to ask back.
Who checks: an engineer checks every number and certification. The estimator decides what happens to each request. The tool never promises prices or lead times.
Watch for: tolerances, capacities, weights, certifications, prices and lead times written from the tool’s memory. A confident wrong spec costs the site an engineer’s trust.
On sales cycles: one agency guide puts them at six to eighteen months, with six to ten technical people involved. That is an agency’s figure, so check it against your own deals.
Law firms
Start with practice-area content that an attorney reviews. Then a limited intake chatbot, reporting, and client alerts built from legal updates. ABA Formal Opinion 512 (July 2024) puts the lawyer’s duties of competence, confidentiality, supervision and fees on every step.
Who checks: an attorney reviews the output. Client information stays out of tools unless the firm allows it, and a lawyer may need the client’s informed consent before it goes into a self-learning tool.
Intake chatbots: Florida Bar Opinion 24-1 says the bot must identify itself as a non-lawyer AI program, only gather facts, give no legal advice, and refer legal questions to a lawyer. The lawyer is responsible if it misleads a prospective client.
Fees: lawyers can bill for time actually spent, including checking AI output, but not for learning the tool. A fixed fee that stays the same after AI cuts the work may be unreasonable.
Advertising rules are set state by state, and I haven’t checked every one. Check yours.
Financial services
Start with a content pipeline that has compliance review and a saved approval record. AI drafts from approved language, helps with webinar follow-up, and can power a narrow website assistant. The firm owns whatever it publishes.
Who checks: compliance, on every piece. FINRA Regulatory Notice 24-09 says the communications rules apply whether a human or a technology tool produces the content, including AI features inside products a firm already uses.
Performance figures: for investment advisers, the SEC’s marketing FAQ says gross performance must be shown with equally prominent net performance over the same period. AI should never write a performance number.
Chatbots: FINRA guidance says chatbot content cannot be unwarranted, promissory or misleading. I’d limit it to approved answers and hand everything else to a person.