Author
Jeffrey Hill
Founder, ValaisOS LLC
Jeffrey Hill is the founder of ValaisOS (va-LAY O-S) and the author of Advisor Insights. He has spent 24 years building and operating technology platforms for wealth management, private markets, and financial data, most recently as a principal AI platform architect and operator, with earlier roles at Communify, S&P Global, IHS Markit (as a managing director), and Wall Street On Demand. His work has focused on advisor technology and the operating practices that let advisory firms preserve client context, exercise judgment, and stay within their regulatory obligations while adopting new tools.
Advisor Insights is written for compliance officers at established firms and for independent advisers who want to be in compliance without a compliance department. Every article is checked against primary sources. The research, drafting, editing, and production are done in collaboration with AI agents that Hill operates himself, which is part of why the publication takes the governance of AI-assisted work seriously: it is the publication's own working method.
Based in Colorado. LinkedIn · hello@valaisos.com
Subject areas
- Advisor operations
- Responsible AI
- Knowledge continuity
- Technology governance
Follow ValaisOS on LinkedIn →
Regulation S-P makes the advisory firm responsible for overseeing its service providers through due diligence and monitoring, but a firm cannot diligence a party it does not know exists. Of nineteen AI vendors serving financial advisors and RIAs whose public documents we reviewed, eight publish a subprocessor list we could read. Six name a model provider and two name a transcription provider. None names a vector store, the system holding a derived copy of client documents. The one vector database we found named is on Anthropic's own subprocessor list, filed under web search, and three of the nineteen vendors now appear as connectors in Claude for Financial Advisors.
Claude for Financial Advisors, released by Anthropic on September 14, 2026, is an openly licensed plugin of eight skills and a twelve-system connector list. The skills are Markdown instructions, the plugin holds no client data, and the repository's own README states that most of its guardrails, including advisor approval before any write to a client system, are enforced by the model following them rather than by the runtime. That makes the plugin a well-written first draft of a firm's AI operating procedures. The controls that hold on a bad day, the record of what happened, and where client documents go when a connector is missing are the firm's to write.
Every AI product sold to advisory firms promises that a person approves the work before it counts. Claude for Financial Advisors gates its CRM writes behind a single batch confirmation, and its repository's design rules require a pause before anything leaves the session. The approval happens. The question a firm has to answer is where it was written down. Anthropic's Enterprise audit log records conversations, projects, and file uploads by identifier, and excludes chat content. The closest thing to a recorded approval event is Cowork telemetry an administrator has to configure. Until the firm makes approval an artifact in its own system, the evidence of review is a chat turn.
Rule 204-2 (17 CFR 275.204-2) does not mention AI. It lists categories of records an SEC-registered adviser must keep, and an AI-generated document falls into a category, or does not, on the same terms as a human-written one. A sent client email relating to advice is a required record whoever drafted it. An unsent draft usually is not. What AI changes is how much of the work now exists only as unsent drafts and vendor logs, and whether the firm can reconstruct what was sent from what it kept.
Under Regulation S-P, a service provider is any entity that "receives, maintains, processes, or otherwise is permitted access to customer information through its provision of services directly to a covered institution" (17 CFR 248.30(d)(10)). The test is whether records about clients flow to the vendor or are accessible by it, not what the vendor is called, how big it is, or whether it uses AI. Since June 3, 2026, every SEC-registered adviser owes each service provider due diligence, monitoring, and written policies reasonably designed to secure a 72-hour breach notification, and the firm keeps the duty to notify clients itself.
Yes, it applies to you. Since June 3, 2026, amended Regulation S-P (17 CFR 248.30) has applied to every SEC-registered investment adviser regardless of size. A firm must maintain written safeguards for customer information, run a written incident response program, notify affected individuals within 30 days of becoming aware of a breach of sensitive customer information, oversee service providers with due diligence and monitoring, and keep records of all of it under Rule 204-2(a)(25). This article walks the rule paragraph by paragraph and ends with what a small firm should do this quarter and what the file should contain.
Advisory work is starting to be done with a system in the middle of it. The finished note, email, or plan looks the same either way, and that is the problem. Institutional memory used to live in people who could be asked. When part of the work comes from a system, memory depends on three things captured at the time: what the system saw, what it produced, and what the person decided. A firm that keeps those can hand work forward. A firm that keeps only the finished artifact has a record it cannot retrace.
An advisor's handwritten meeting note was never a recording. It was a judgment about what mattered, and that judgment was the firm's memory. Automated notes invert this: the record becomes complete and unweighted, and the note the advisor used to write is the note that quietly stops being written. The transcript is a gain. The substitution is a loss. Firms that keep a short judgment layer beside the automated record get both.
Between August 4 and August 21, 2026, an attacker used a defect in Lenovo's ID registration to sign into roughly 5,000 Dropbox accounts with no password and no access to the victim's inbox. Dropbox says every affected account had two-factor authentication off. For advisory firms that had it on, there was no letter, no investigation, and no Regulation S-P clock. The article explains why a password was irrelevant and a second factor was not, what that one setting spared firms under the amended rule, which second factor to use, and how to require it across every cloud tool that touches client files.
The first wave was clients arriving with AI-generated answers. The new wave is clients arriving with AI-generated question lists for their advisor. The list is usually reasonable. The prompt that produced it tells you what the client is actually concerned about, and that is where the real meeting starts.
A surprise tax bill is almost never a knowledge failure, because every advisor involved knows what brackets do. It is a timing failure. The information existed in March, sat unnoticed in a payroll system or a brokerage feed through the summer, and arrived at the advisor's desk as an April emergency. The gap between where the data lives and where the judgment lives is what manufactures the surprise, and closing that gap is the least glamorous and most valuable work in an advisory practice.
Nobody calls their advisor in tears about tax-loss harvesting. The moments that make an advisory practice are human, being the first call after the promotion, the sale, the birth, the loss, and no machine can replace them. The honest question about AI in an advisory practice is not whether it can replace those moments. It is what it clears out of the way so there is room for them. The firms getting this right are deliberately deciding which work deserves a human and which work deserves an audit trail.
Most vendor security questionnaires are long, generic, and produce answers nobody reads. These nine questions are shorter and harder, because each one has an answer that would change whether you sign. They cover the breach-notification clock, where your client data physically goes, whether it trains a model, what you can export, whether the vendor's records satisfy your recordkeeping rule, and who carries the loss. Each question is paired with what a usable answer looks like.
The duty to notify clients after a vendor breach sits with the advisory firm, not the vendor. Of nineteen AI vendors serving wealth management whose public documents we reviewed, ten publish a notification deadline, and eight of those ten put it in a data processing agreement, the terms of service, or a subscription agreement rather than the privacy policy. Two name Regulation S-P. Three of the ten deadlines carry qualifiers that change what they mean. The article closes with nine questions a firm can send any vendor.
A production chain is only as strong as its weakest link, and automating most of a chain concentrates value and risk in whatever remains. In advisory work, what remains is judgment on one end and execution on the other. The overlooked part is the handoff between them, where a decision leaves the person who made it and enters the system that acts on it. Professions that automated safely, like structural engineering, survived because they had an artifact that made judgment attributable. Advisory work has no stamp, and that seam is where accountability is established or lost.
An AI policy states what the firm intends. Exams, arbitrations, and client disputes are settled by evidence of what actually happened. AI work leaves no record unless the workflow is designed to produce one, so the gap between a firm's policy and its receipts is wider than most firms realize.
Personalized product outreach used to be expensive, and the expense was an accidental control. AI removes the cost but not the obligation. A generated reason a product fits a client is a hypothesis, not evidence of fit, and the firm now owns every tailored message it sends.
When a client calls about a market headline, they are rarely asking what the market will do. They are asking whether the news applies to them. Headlines describe aggregates; clients live in particulars, and the advisor's work is translating one into the other with specifics: what you hold, what the plan assumed, what would change it. Morningstar and Vanguard research put a measurable cost on acting without that translation, and the firm's systems should already know which households a headline touches.
A client with strong income, maxed retirement accounts, and no debt had $140,000 sitting in a checking account for years, visible on every statement and seen by no one. Idle cash is a background condition of most books: as of July 2026 the FDIC's national average interest checking rate was 0.07%. Software is well suited to noticing across every account at once; deciding what the money is for remains human work, and the system that notices must be governed.
When a client asks what you think about AI, a credible answer does three things: acknowledges what the technology does well, names precisely what it cannot know, and connects that limit to the judgment and accountability an advisor provides. Dismissal and breathless enthusiasm both close the conversation. The article gives a version an advisor can say out loud, the research behind the limit, and the prerequisite: firsthand evaluation inside firm-approved systems.
When a client brings an AI-generated financial conclusion into a meeting, the useful first move is to ask to see the prompt, because the answer depends on what the system was told and what it was never told. A 2026 working paper by Choukhmane, de Silva, Lin, and Akuzawa found AI financial guidance was often directionally sound yet leaned on round-number heuristics and varied with the inputs. The article gives five questions an advisor can ask and explains how to preserve the reasoning that follows.
Advisory firms rarely lack information. Account records, plans, CRM notes, and emails record what happened. What they often fail to preserve is why: what prompted a decision, what the client cared about, which alternatives were considered, and what remains open. Client context is that reasoning, and it is what lets another authorized professional continue the work without guessing. The article gives five warning signs, a continuity test, and four practical improvements.
