NobleCloak
Playbooks

AI vendor risk and due diligence for a 1–2 person compliance program

Author

Audrey

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

Date Published

Most advice about AI vendor risk assumes a team you don't have. In Ncontracts' 2026 State of Third-Party Risk Management survey of 173 financial services professionals, "nearly two-thirds (63%) of TPRM programs operate with just one or two dedicated full-time employees." The same survey found institutions rank AI risk on par with cybersecurity as their top third-party concern, and "72% admit they are only partially aware of which vendors use AI."

This is the program for that one- or two-person team. The goal is a file that is proportional, dated and honest, and that you can finish this month.

Proportionality is already in the guidance

None of the frameworks ask a small institution to diligence a grammar checker the way it diligences its core processor.

Banks. The June 2023 interagency guidance on third-party relationships (Fed, FDIC, OCC) says it is each banking organization's responsibility "to calibrate its risk management processes, commensurate with the banking organization's size, complexity, and risk profile and with the nature of its third-party relationships." It is equally plain that outsourcing keeps the responsibility at home: "A banking organization's use of third parties does not diminish its responsibility to meet these requirements to the same extent as if its activities were performed by the banking organization in-house." It is guidance. In the agencies' words, "supervisory guidance does not have the force and effect of law."

That guidance is up for replacement. On November 13, 2026, the OCC, Fed, FDIC and NCUA extended the comment period on proposed third-party risk guidance to December 16. The agencies plan to rescind the existing guidance once the new version is final.

Credit unions. NCUA's Supervisory Letter 07-01, sent with Letter to Credit Unions 07-CU-13, says due diligence "should be tailored to the complexity of the third party relationship and may consist of reasonable alternative procedures to accomplish acceptable risk mitigation." The catch: a May 2025 GAO report found "NCUA lacks the authority to examine technology service providers." No examiner is checking your AI vendors for you. More in NCUA can't examine your AI vendors.

RIAs and broker-dealers. Amended Reg S-P, 17 CFR 248.30(a)(5), requires written policies "reasonably designed to require oversight, including through due diligence and monitoring, of service providers." Those policies must be designed so a service provider notifies you "as soon as possible, but no later than 72 hours" after it learns of a breach of a customer information system it maintains. Smaller entities had to comply by June 3, 2026. The file itself is in Reg S-P is live: the AI vendor file your examiner expects.

The common thread is effort scaled to risk, with dated evidence of monitoring. Equal, thin effort across every vendor reads as no risk assessment at all.

Your vendors are adding AI without telling you

The hardest part of the surface is the vendors you already approved. AI arrives as a product update: a summary button in loan origination, a notetaker that joins calls by default, a support portal that routes tickets through a language model. The contract stays the same while the product changes underneath it.

Your diligence covered the vendor as it was when you signed. That's why monitoring matters more than onboarding here, and why every renewal should reopen the file.

Step 1: run the data call in one week

You can't govern what you can't see. The output is one spreadsheet, one row per tool: name, vendor, users, data it touches, how it got in (approved, vendor feature, or shadow), what attestation you hold, and a risk flag.

Ask people. Send four plain questions to every department head. Keep it non-punitive or the honest answers stop.

  1. What software writes, summarizes, transcribes or "suggests" things for you?
  2. Have you pasted member, customer or account information into a chatbot or website?
  3. What have you connected to your work email, calendar or files?
  4. What AI feature did a vendor turn on this year that you didn't ask for?

Check the systems. Self-reports miss things.

  • OAuth grants. In the Google Workspace admin console, Security → Access and data control → API controls → Manage App Access lists each accessed app, the Google services it requested and how many users. In the Microsoft Entra admin center, Entra ID → Enterprise apps → All applications → Permissions shows admin consent and user consent per app. The full method is in Shadow AI is an access problem not a traffic problem.
  • Your vendor list. Read each vendor's release notes for "assistant," "summary" or "copilot."
  • Card and AP records. A small monthly charge is often a department's AI habit with a receipt.

A typical find: a meeting notetaker holds calendar and Drive read access for eight users, keeps recordings, and has no attestation on file. Someone clicked "allow." Log it, flag it high, move on.

Step 2: triage into three tiers

Sort by the data a tool touches and the access it holds.

  • Tier 1, real diligence. Write access, standing read access to member or client data, or a role in a core process. This list should be short.
  • Tier 2, light touch. Read-only, narrow scope, nothing sensitive. A one-line note and an annual review.
  • Tier 3, revoke. Abandoned trials and shadow apps with no owner. Remove the grant and log it. A revoked grant is a closed risk.

Most of your hours go to Tier 1. That concentration is your risk assessment.

Step 3: ask Tier 1 vendors eight questions

Real answers are specific and checkable: a scope, a number, a contract section. Marketing answers are adjectives like "secure" and "industry-leading."

  1. What data do you access, at what scope? Compare it with the grant you saw in Step 1.
  2. Do you train on our data or our customers' data? A clear yes or no, in the contract. "Aggregated data to improve our services" is a yes.
  3. Where is our data stored, for how long, and how is it deleted?
  4. Which sub-processors and model providers do you use? A current list, and notice when it changes.
  5. What attestation can you share? A current SOC 2 Type II or equivalent, with its audit period. "We are SOC 2 compliant" with nothing attached is a claim.
  6. What is your breach-notification commitment? A number of hours in the contract. RIAs should match it to their Reg S-P policies.
  7. Which AI features are on by default, and how will you tell us about new ones?
  8. What happens to our data when we leave? Export isn't deletion.

You won't get every answer. Write down the refusals. "Vendor declined to confirm training-data use in writing, [date]" is a dated finding, and it proves you asked.

Step 4: start the monitoring log today

Three short policies are enough: AI use, service-provider oversight and, for RIAs, Reg S-P incident response. Skip the GRC platform; a spreadsheet with the right columns is a defensible record at your size. Then start a log. A few dated lines per quarter: re-ran the grant list, re-checked Tier 1, reopened diligence at renewal. Oversight with no dated trail looks like oversight that stopped at onboarding. Point-in-time assessment vs continuous monitoring covers cadence.

When someone asks what you're doing about AI

An examiner, your board or your supervisory committee will ask. None of the guidance above names AI, and it doesn't need to: your AI vendors are third parties. A one-page summary answers three things.

  1. What AI is in use, ours and our vendors'? The dated inventory.
  2. What can each tool reach? The access from the grants.
  3. What did we decide about the risky ones? The questions, the answers and refusals, and what you revoked, restricted or accepted.

Name the file's limits. This program sees AI reachable through your identity provider and your known vendors. It won't see data pasted into a personal account on a phone, or AI a vendor runs on its own servers with no new grant. A file that admits its edges is easier to believe.

Where NobleCloak fits

Our read-only scan does the hardest part of Step 1. It reads Google Workspace and Entra and shows "every AI app your people authorized, the data scopes it holds, who's using it, since when." Coverage is limited to what SSO and OAuth can see, and our team runs the first report with you. See the AI Trust Report. Drafting policies from the scan is coming soon.

When agents act in your systems, policy marks each tool as allowed, needs approval, or denied. Approval goes to a named person or role, and one approval covers one action. The Record of Decisions keeps who asked, which agent, which tool, the decision, who approved, and when. See how human-in-the-loop works or The vendor risk agent reviews; people approve.

NobleCloak doesn't send vendor questionnaires, read contracts or see AI a vendor runs on its own servers. Steps 2 through 4 stay with you.