/

Guides

/

Guide · Compliance

Five security questions every RIA must ask an AI vendor

If you only ask five questions before an AI tool touches client data, ask these. Each comes with why it matters, the answer you need, the red flags and the evidence to request.

By the Verlo team

Rules and vendor pages checked: November 2026

Vendor note. Verlo makes AI software for financial advisors, so firms ask us these questions too. The list works for any vendor, and we point to what OpenAI, Anthropic and others publish so you can see what good evidence looks like. The rule references come from the eCFR, read on November 6, 2026. This is general information, not legal advice; have your compliance officer or counsel adapt it. If something here is out of date, tell us at support@verlo.finance. For how these questions fit into choosing a tool, see our guide to the best AI software for financial advisors.

The five questions. 1. Is our data used to train any model, yours or your providers’? 2. Where does our data go, and who else touches it? 3. What independent evidence do you have? 4. Will you tell us within 72 hours, in writing? 5. How do we see what the AI did, limit it, and get our data out?

A full due-diligence questionnaire has dozens of questions, and you should send one to every vendor that handles client information. But when you are sitting in a demo, or a team member wants to try a tool this week, you need a short list that catches the deal-breakers. These five do. Together they cover what Regulation S-P makes you responsible for: whether the vendor protects customer information, who else touches it, and whether you hear about a breach in time to act.

1. Is our data used to train any model, yours or your model providers’?

Why it matters. Client information used for training can leave your control and may not meet the contract condition for sharing it with a service provider under the privacy rule. Personal and free plans often differ from business plans here.

The answer you need. No, in writing, covering the vendor and every model provider behind it, with no exceptions for feedback or “improvement” features you cannot switch off.

Red flags. “We anonymize it.” Training defaults that depend on a setting each user must find. An answer about the vendor that skips its model providers.

Evidence to request. The contract clause, the model providers’ terms, and the data-use page for your plan.

Anthropic’s privacy center article on model training for commercial products, dated August 18, 2026, saying inputs and outputs are not used to train models by default
What a clear published answer looks like: Anthropic’s article for commercial products says it does not use inputs or outputs to train its models by default, with stated exceptions. Source: privacy.claude.com, captured November 6, 2026.

2. Where does our data go, and who else touches it?

Why it matters. Regulation S-P defines a service provider as a person or entity that receives, maintains, processes or is permitted access to customer information through services it provides directly to you. Your vendor’s hosts and model providers are likely to be part of the chain, and the 72-hour breach notice runs from a provider’s awareness of a breach of a system it maintains.

The answer you need. A named list of subprocessors and model providers with each one’s function and region, a data-flow diagram that matches what you will connect, and notice before anything changes.

Red flags. No subprocessor list. “We use several leading AI providers.” No notice before a model or host changes.

Evidence to request. The subprocessor list, the data-flow diagram and the data processing terms.

3. What independent evidence do you have?

Why it matters. Adjectives are free. An independent attestation is a third party testing the vendor’s controls, and it gives you something to put in your file.

The answer you need. A current SOC 2 Type 2 report (not just Type 1) covering the product you will use, with a recent period and a bridge letter if needed, plus a recent penetration-test summary you can read.

Red flags. A SOC 2 claim with no report you can see, under NDA or through a trust center. A report for a different product. A Type 1 report only, with no date for Type 2. Refusal to share anything under NDA.

Evidence to request. The report, bridge letter and penetration-test summary, or access through the vendor’s trust center. See what to put in your vendor file.

OpenAI’s Trust Portal compliance card with badges for SOC 2 Type 2, SOC 3, PCI DSS, several ISO certifications, CCPA, CSA STAR, GDPR, TX-RAMP and FedRAMP 20x
A badge shows which attestation exists, so ask which products it covers and for which period. Source: trust.openai.com, captured November 6, 2026.

4. Will you tell us within 72 hours, in writing?

Why it matters. Regulation S-P requires your oversight policies to be reasonably designed to ensure service providers notify you as soon as possible, and no later than 72 hours after becoming aware of a breach of a system holding your customer information. That is what starts your own incident response and your 30-day client-notice clock (17 CFR 248.30(a)(5)).

The answer you need. Yes, written into the contract, with a named contact, what the notice will contain, and help with your own client notifications.

Red flags. “We notify within a reasonable time.” Notice only to a generic support address. A refusal to put it in the agreement.

Evidence to request. The breach-notification clause and the incident-response summary.

5. How do we see what the AI did, limit it, and get our data out?

Why it matters. You must be able to supervise AI use, produce records for an exam, and leave a vendor without losing or stranding client data.

The answer you need. Exportable audit logs of who and what did what, admin controls to restrict features for the whole firm, retention you can set, an export in a usable format, and deletion on exit with written confirmation.

Red flags. No audit log. Settings that each user controls. Data that cannot be exported. Deletion only “on request, eventually.”

Evidence to request. A sample log export, the admin guide, and the exit and deletion clause.

A quick scorecard

Question

Pass

Conditional

Fail

1. Training

No training, in writing, including model providers

No training by default but a setting or exception remains

Training, or no answer

2. Data flow

Named subprocessors, data-flow diagram, change notice

List exists but incomplete

No list

3. Evidence

Current SOC 2 Type 2 for your product, plus pen-test summary

Type 1, or an old period, with a plan

No independent evidence

4. 72-hour notice

In the contract, with a named contact

Promised but not in writing

Refused

5. Visibility and exit

Logs, admin controls, export and deletion

Partial

None

A conditional answer can be acceptable with a contract term or a restriction on use, such as limiting the tool to data that is not client information. A fail on any of the five is a reason not to connect client data until it is resolved. Document the decision either way.

Frequently asked questions

What security questions should an RIA ask an AI vendor?

Five cover the essentials: whether your data trains any model, where it goes and who else touches it, what independent evidence the vendor has, whether it will notify you of a breach within 72 hours in writing, and how you see, limit and export what the AI did.

What is a good answer on training?

No training on your data by the vendor or its model providers, in writing, with no exceptions you cannot control. Check the plan you will actually use, because personal and business plans often differ.

Do I need a SOC 2 report from every AI vendor?

No attestation is named in Regulation S-P. A current SOC 2 Type 2 report is the most common evidence for vendors that handle client information, so most firms require it or a documented alternative. Tier your vendors by what they can reach.

What if the vendor will not answer?

Treat it as a red flag, document the request and the reply, and do not connect client data until you have what you need.

Go deeper

Updates to this guide

  • November 17, 2026: First published.

Sources

Checked November 5 and 6, 2026: 17 CFR 248.30, Anthropic, training on commercial products, OpenAI Trust Portal.