Records, Communications and Supervision
An AI-Drafted Message Is Still a Business Communication
Record-keeping and supervision regimes are indifferent to how a communication was produced. A client message, a piece of research, a marketing post or an internal instruction sits inside the same retention, retrieval and review obligations whether a person wrote it from scratch or edited a model's draft. The US position splits by firm type: broker-dealer supervision runs through FINRA Rule 3110 and communications with the public through FINRA Rule 2210, while books-and-records requirements for investment advisers sit with the SEC, which has also pursued firms over business conducted on unrecorded channels. In the EU, MiFID II imposes recording and retention duties on communications relating to transactions. The failure mode is constant: work done where the firm cannot retain or produce it.
- Retention, retrieval and supervision duties attach to the communication, not to how it was drafted
- In the US, broker-dealer supervision is FINRA Rule 3110 and public communications FINRA Rule 2210; adviser records sit with the SEC
- In the EU, MiFID II imposes recording and retention obligations on transaction-related communications
- An AI tool outside the retained estate is simply another unrecorded channel, however useful it is
What Has to Be Retained and Reviewable
If AI is used in client-facing work, decide in advance what the record consists of. At minimum, the communication as sent, held in the existing retained channel — that part is not new. Beyond that, institutions increasingly need to reconstruct how something was produced: which tool, which version, what was asked, what the model returned, what a human changed, and who approved it. That is a logging decision taken before deployment, because it cannot be recovered afterwards. Supervision has to adapt as well, since review programmes calibrated on human-written communications will meet a far larger volume of fluent, near-uniform text — exactly the material that sampling-based review triages worst.
- The communication as sent belongs in the existing retained channel; that obligation has not changed
- Reconstructing production — tool, version, prompt, output, edit, approver — is a logging decision made up front
- Volume and uniformity break review programmes calibrated on human-written communications
- Set retention periods from your firm's records policy, never from a tool vendor's default
Marketing, Recommendations and Claims About AI Itself
Two further exposures sit close together. AI-generated marketing and research meet the same fair, clear and not misleading standards as anything else, and models are fluent producers of unbalanced enthusiasm, missing risk warnings and implied performance. Which rule applies depends on the firm: in the US, broker-dealer communications with the public fall under FINRA Rule 2210, while the SEC's marketing rule applies to investment advisers. Where a communication is personalised at scale it can also cross from marketing into a recommendation, engaging MiFID II suitability in the EU and Regulation Best Interest in the US; in the UK the FCA's Consumer Duty adds outcomes-based expectations on top. Overstating AI in your own product — "AI washing" — is itself a regulated-communications problem.
- AI-generated marketing and research meet the same fair, clear and not misleading standards as any other
- In the US the applicable rule turns on firm type: FINRA Rule 2210 for broker-dealers, the SEC marketing rule for advisers
- Personalisation at scale can turn a communication into a recommendation — MiFID II suitability in the EU, Reg BI in the US
- Overstated claims about your own use of AI are a regulated-communications problem in their own right
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.