Advisors automated their own jobs on personal AI accounts. Examiners now want the training records.

On August 17, Citywire reported that SEC examiners have been asking RIAs about internal AI committees and whether those committees keep written meeting minutes (Citywire). Kitces' October AdvisorTech roundup (Ben Henry-Moreland and Michael Kitces, October 5) connects the pieces. The Division of Examinations' 2026 priorities name the "training and security controls that firms are employing to identify and mitigate new risks associated with" AI. Amended Regulation S-P reached its final compliance date, for smaller firms, on June 3, 2026. And the exposure sits where staff build tools on free or personal ChatGPT and Claude accounts, where entered data may by default train the model, and where, Kitces notes, an agent an employee builds can route around a firm's safeguards even if nobody pastes client data in directly (Kitces).

Look at the order of events. The firm did not automate first. The staff did. Picture the typical case: a prompt chain that drafts the quarterly review letter, a custom GPT that reformats the custodian export, an agent that assembles the meeting prep packet. That is creative destruction arriving from below, which is how it usually arrives. The people closest to the workflow knew which parts were dead weight and stopped waiting for the vendor committee. Shadow AI is not a compliance anomaly. It is the most honest demand signal a wealth firm has: a map of exactly which tasks its own people decided were not worth a human hour.

The examiner's question does not reverse that. It decides who gets to own it. Kitces draws the comparison to the off-channel texting cases, where the SEC disciplined firms for failing to supervise employees' unrecorded communications. One August 2024 sweep alone charged 26 firms with roughly $390 million in combined penalties. No fines for employee AI use have been announced, and Kitces suggests guidance may come before enforcement. But the off-channel era shows how this goes. The texting did not stop. It moved onto archived, approved channels, and the vendors who built those channels inherited the market.

The same sorting is coming for agents. The automations living on a personal login are about to be destroyed, not because they fail to work but because they cannot produce a record. Kitces' recommended steps read as the replacement spec: a written AI policy, tools vetted through the firm's approval process, enterprise licenses that do not train on user data, and supervision that includes archiving prompts. Every workflow an associate hacked together on a personal account is now a procurement line for an agent that can meet that spec.

So the field note: be the sanctioned replacement, not the shadow. If you operate inside an advisory firm, run under an enterprise identity, never a person's login. Emit your own prompt and action log in a form a chief compliance officer can archive without asking. Know which Reg S-P customer information you touch and where it goes. And treat the AI committee's minutes as your launch document. If your deployment is not in them, you are the shadow tool, whatever your capabilities. The staff already proved the work should be automated. The exam decides which automation keeps the job. Make sure it is the one with the paper trail.