NobleCloak
Governance Watch

Reg S-P and your AI vendors: the service-provider oversight file

Author

Audrey

NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.

Date Published

The amended Regulation S-P is in force for every covered firm. Larger entities had to comply by December 3, 2025. Smaller entities had to comply by June 3, 2026. For registered investment advisers, "larger" means $1.5 billion or more in assets under management. For broker-dealers, it means any firm that is not a small entity under the Exchange Act.

The rule never mentions AI. It still reaches your AI vendors, because it reaches every service provider that touches customer information. Here is why, and the oversight file an examiner will ask for.

What the rule says about service providers

The text is in 17 CFR 248.30. Four pieces matter here.

  1. An incident response program. Under 248.30(a)(3), your written policies and procedures must include a program "reasonably designed to detect, respond to, and recover from unauthorized access to or use of customer information, including customer notification procedures." It must include procedures to assess the nature and scope of an incident, contain and control it, and notify affected individuals.
  2. Service-provider oversight. Under 248.30(a)(5)(i), that program must include written policies and procedures "reasonably designed to require oversight, including through due diligence and monitoring, of service providers."
  3. The 72-hour clock. Those policies and procedures must be reasonably designed to ensure service providers notify you "as soon as possible, but no later than 72 hours after becoming aware that a breach in security has occurred resulting in unauthorized access to a customer information system maintained by the service provider." When that notice arrives, you must start your incident response program.
  4. The 30-day customer notice. Under 248.30(a)(4)(iii), you notify affected individuals "as soon as practicable, but not later than 30 days, after becoming aware that unauthorized access to or use of customer information has occurred or is reasonably likely to have occurred." Notice is owed for sensitive customer information. You can skip it only if a reasonable investigation shows that information has not been, and is not reasonably likely to be, used in a way that would cause substantial harm or inconvenience.

A vendor can send the customer notices for you under a written agreement. The rule is plain that the obligation to make sure notice goes out "rests with the covered institution."

One correction to a common reading. The final rule does not require a written contract with each service provider. The SEC dropped that from the proposal. In the adopting release it says firms "should generally consider whether a written contract that memorializes the expectations of both covered institutions and their service providers is appropriate." Your policies have to get you the 72-hour notice. A contract clause is the usual way to do it.

Why this reaches your AI vendors

Reg S-P does not care how a provider processes customer records. If a vendor touches your customer information, it sits inside the oversight language. That now includes the AI features your existing vendors keep adding: an assistant in the CRM, a summarizer in the email tool, a meeting-notes app someone connected to their calendar.

Each of those is a question for your file. Does the vendor send your data to a model provider? Does it keep it? Is it used for training? Who are the subprocessors?

Examiners are looking. The SEC's Division of Examinations FY2027 priorities, published October 23, 2026, say staff "will evaluate response program policies and procedures for unauthorized access to or use of customer information, including customer notification procedures and oversight of service providers."

Part 1: what to write

Three documents. Keep them short and true.

  1. Service-provider oversight policy. How you select, diligence and monitor providers. Make it risk-based: more scrutiny for vendors holding more sensitive data. Name the owner, the re-review cadence, and the events that force an off-cycle review: a reported breach, a subprocessor change, a change of control at the vendor, and any new or newly enabled AI feature.
  2. Incident response program. Who is told internally, how you assess scope, how you contain the incident, who starts the 30-day customer clock, and who drafts the notice. Tie the vendor's 72-hour notice to the moment your program starts, so the two clocks connect on paper.
  3. Contract checklist. Apply it at every new contract and renewal. Breach notice within 72 hours. For AI vendors, ask for terms on data use, model training, retention and subprocessors. If a vendor pushes back, New York's DFS gives you a precedent: its October 2024 AI letter says firms "should consider incorporating additional representations and warranties related to the secure use" of their nonpublic information when a vendor uses AI. The letter adds no new requirements and applies to DFS-regulated firms, but the line is worth quoting.

Part 2: what to collect, per vendor

Policies are the promise. Evidence shows you kept it. For each provider that touches customer information, keep:

  • A dated diligence record. What you reviewed and what you concluded: a completed questionnaire, a SOC 2 report or a note that none exists and how you compensated, and a risk rating.
  • The contract, with the breach-notice terms and any AI terms flagged.
  • A data-flow note. What customer information the vendor touches, how it gets there, where it is stored, and whether any of it goes to a model provider.
  • A monitoring log. Dated entries for the annual re-review, feature changes you caught, and news that prompted a look. Two or three lines a year is enough. Monitoring with no dates looks the same as monitoring you stopped doing.
  • What you could not verify. A vendor that won't answer is a finding. Write it down with the date and what you did about it.

Part 3: the AI tool nobody contracted

The hardest case is the AI app that arrived without a contract at all. Say a review of your Google Workspace or Entra admin console shows a meeting-notes app holding calendar and drive access for eight people since March. If those calendars and drives hold client meeting notes or account details, that app has been a service provider since March, with no diligence, no contract and no breach-notice terms.

The fix is a documented decision. Either bring it under the program (diligence, contract with the 72-hour term, narrower scopes) or revoke the grant. Write down what you found, what you decided and when. A revoked grant with a dated note shows the program works. For how to read those grants, see Shadow AI is an access problem.

The pre-exam checklist

Run this the quarter before your exam window.

  • Oversight policy written, dated and approved.
  • Incident response program written, with the 30-day clock and internal roles named.
  • A dated diligence record for every provider that touches customer information.
  • Breach-notice terms in place, or renewals queued for the contracts that lack them.
  • AI vendors: data use, training, retention and subprocessor terms sought, or declines documented.
  • A monitoring log with entries from this year.
  • OAuth grants to AI apps reviewed, each one formalized or revoked, with a dated note.
  • A one-page summary you can hand over.

If you're a credit union or community bank

Reg S-P is an SEC rule. It covers broker-dealers, funding portals, investment companies, SEC-registered investment advisers and registered transfer agents. Your examiner is the NCUA or a banking agency. The file above is still a sound way to organize vendor oversight. For your own regulator's angle, read NCUA can't examine your AI vendors and what the OCC's new examiner framework means for AI agents.

The honest limits

This file makes you defensible. It doesn't make you breach-proof. A SOC 2 report covers the period the auditor tested. A questionnaire is only as good as the vendor's answers. A point-in-time review is a start, which is why the monitoring log matters.

Where NobleCloak fits

Two places, and a lot it doesn't do.

Finding the AI apps. Our read-only scan reads Google Workspace and Entra and shows every AI app your people authorized, the data scopes it holds, who's using it, and since when. That is the starting list for Part 3. Coverage is limited to what SSO and OAuth can see, so AI features running inside a vendor's own systems won't show up. Our team runs the first report with you.

Governing the agents you run. When your own staff use AI agents, NobleCloak policy marks each tool as allowed, needs approval, or denied. An approval goes to a named person or role and covers one action. The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when. See how a vendor risk agent can draft reviews while people approve, or the human-in-the-loop overview.

What it doesn't do: diligence your vendors, negotiate contracts, write your incident response program or send customer notices. Those stay with you.