Vendor note. Verlo makes AI software for financial advisors, so some firms will run this checklist on us. We wrote it to work on any AI vendor. You can request our security documents through the Verlo Trust Center. Our guide to AI software for financial advisors compares the tools themselves. This is general information, not legal advice. Have your compliance officer or counsel review it before you adopt it.
If your firm uses an AI tool on client information, the company behind that tool is likely to be a service provider under Regulation S-P’s definition, and the rule expects you to oversee it through written policies. That means due diligence before you connect it and monitoring after.
This guide gives you a process you can repeat for every AI vendor, a 52-question due-diligence questionnaire (DDQ) you can send, and a list of the evidence to ask for. Download the DDQ as a spreadsheet.
What Regulation S-P expects of you
The amended rule is at 17 CFR 248.30. For an SEC-registered investment adviser, the parts that matter for AI vendors are these:
You are a covered institution. The definition includes investment advisers registered with the SEC.
You need a written incident response program for detecting, responding to and recovering from unauthorized access to customer information.
You need written policies for overseeing service providers, including due diligence and monitoring.
A service provider must notify you as soon as possible, and no later than 72 hours, after it becomes aware of a breach of a system that holds your customer information.
You must notify affected individuals as soon as practicable, and no later than 30 days after you become aware of unauthorized access that occurred or is reasonably likely to have occurred, unless an exception applies.
A service provider is any person or entity that receives, maintains, processes or otherwise is permitted access to customer information through its provision of services directly to you. An AI vendor connected to your CRM, custodian data, email or documents is likely to fit.
The compliance dates have passed. Larger firms (advisers with $1.5 billion or more in assets under management) had to comply by December 3, 2025, and other firms by June 3, 2026, per FINRA’s reminder. The $1.5 billion threshold is summarized by Goodwin.
The rule is principles-based. It does not list required attestations, questionnaires or contract terms, and it does not set how often to re-review a vendor. What it does require is that your policies are written, that you follow them, and that you can show you did. The process below is one way to do that.
Step 1: Decide how much diligence the tool needs
Tier the tool by what it can reach, not by what you intend to use it for. A tool connected to your CRM can see client records even if you only ask it about marketing.
Tier | What the tool can touch | Examples | Diligence |
|---|---|---|---|
1 | Customer information: names with account details, balances, holdings, tax or estate documents, meeting recordings or transcripts, client emails | An AI notetaker, an assistant connected to your CRM or custodian, a document-analysis tool, a forms filler | The full DDQ, evidence review, contract terms, and a re-review at least once a year |
2 | Firm information that is not customer information: internal procedures, templates, marketing drafts with no client details | A drafting assistant used only on internal or public text | The short form (questions A1, A4, C2, C3, D1, F2, I2 and I3) and a written use restriction |
3 | Public information only | A search or research tool | Record the decision and the restriction, and revisit it if the use changes |
Step 2: Send the DDQ
Send these questions to the vendor’s security or compliance contact. A good answer is specific, written and backed by a document. Wherever the answer is “yes”, ask for the evidence in the last column. The spreadsheet adds columns for the vendor’s response, what you received, your notes and a pass, conditional or fail rating.
A. Company and viability
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
A1 | Who owns the company, and who funds it? | Clear ownership and funding, with no undisclosed related parties | Corporate overview, funding or financial statements |
A2 | How long has the product served advisory firms, and how many RIAs use it? | Specific numbers and references you can call | Customer references |
A3 | Does the product depend on a single AI model provider or cloud? | A named list of dependencies and a fallback plan | Architecture overview |
A4 | What happens to our data and service if you are acquired or shut down? | Written transition and data-return commitments | Contract clause, continuity summary |
A5 | What insurance do you carry? | Cyber and professional liability cover with stated limits | Certificate of insurance |
B. Data: what, where and who sees it
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
B1 | What customer information will the tool access, and from which systems? | A written list by data type and source | Data-flow diagram |
B2 | Can we limit access to specific data, clients or users? | Role-based, least-privilege controls that follow our existing permissions | Admin guide or demo |
B3 | Where is data stored and processed, by region and cloud provider? | A named cloud and region, US-only where required | Hosting statement |
B4 | Is data encrypted in transit and at rest, and who holds the keys? | Named encryption standards and a described key-management approach | Security whitepaper |
B5 | Do your employees access our data, and under what controls? | Access is limited, logged, reviewed and time-bound for support | Access policy |
B6 | Is our data logically separated from other customers’ data? | Tenant isolation that is described and tested | Architecture or penetration-test summary |
C. AI models and training
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
C1 | Which AI models and model providers process our data? | A named list, with notice before it changes | Subprocessor list |
C2 | Is our data used to train or improve any model, yours or a provider’s? | No, in writing | Contract clause, provider terms |
C3 | Do the model providers retain prompts, files or outputs? | Zero-data-retention terms with each provider, or a defined short period | Provider agreements or written attestation |
C4 | Can the AI take actions (send, write back, delete), or only draft? | Actions need approval, and admins set the scope | Product documentation |
C5 | How do you protect against prompt injection from untrusted email or documents? | Described controls and testing | Security testing summary |
C6 | How do you test accuracy, and what happens when a model or prompt changes? | A defined evaluation process and change notices | AI governance policy |
C7 | Can answers be traced to their sources? | Citations to the underlying records | Product demo |
D. Security program and attestations
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
D1 | Do you have a current SOC 2 Type 2 report? | Yes, for a recent period, shared under NDA | The report and a bridge letter |
D2 | When was your last independent penetration test, and were the findings fixed? | Within the last year, with a summary you can read | Penetration-test summary |
D3 | Do you enforce multi-factor authentication and support single sign-on? | MFA is enforced and SSO or SAML is supported | Admin settings |
D4 | How do you manage vulnerabilities and patching? | Defined timelines by severity | Vulnerability management policy |
D5 | How are your employees screened and trained? | Background checks and recurring security training | HR security policy |
D6 | Do you follow a recognized framework such as ISO 27001 or NIST? | Certified or mapped to the framework | Certificate or mapping |
E. Access controls and logging
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
E1 | Is every user and AI action logged with who, what and when? | Tamper-resistant, exportable audit logs | A sample export |
E2 | Can we export logs and records for an SEC exam? | Yes, on request within a defined time | Process description |
E3 | Can we turn features off for the whole firm, such as web search? | Yes, through admin controls, with a record of changes | Admin guide |
E4 | How is access removed when one of our employees leaves? | Immediate deprovisioning through SSO or an admin action | Process description |
F. Incident response and notification
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
F1 | Do you have a written incident response plan, and when was it last tested? | Yes, tested at least annually | Plan summary and test date |
F2 | Will you notify us of a breach affecting our data within 72 hours of becoming aware of it? | Yes, written into the contract | Contract clause |
F3 | What does a breach notice contain? | What happened, which data, what you are doing, and a contact | Notice template |
F4 | Who is our contact during an incident, and how fast do you respond? | A named escalation path with response times | Contact sheet |
F5 | Can you support our own notifications to clients, which have a 30-day deadline? | Yes, with the facts and cooperation we need | Contract clause |
F6 | Have you had a breach affecting customers in the last three years? | Disclosed with the outcome and the fix | Written statement |
G. Subprocessors
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
G1 | Who are your subprocessors, and what does each one do? | A published list with function and location | The list |
G2 | Do you flow security, training and breach-notice terms down to them? | Yes | Contract excerpt |
G3 | Will you tell us before you add or change a subprocessor or model provider? | Advance notice and a right to object | Contract clause |
G4 | How do you assess your own vendors? | A documented diligence process | Policy |
H. Records, retention and supervision
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
H1 | What do you store (recordings, transcripts, summaries, chat logs), and for how long by default? | Clear defaults that we can change | Retention schedule |
H2 | Can we set retention to match our books-and-records policy? | Yes, through admin settings | Admin guide |
H3 | Can we retrieve records quickly for an exam or a client request? | Yes, with search and export | Demo |
H4 | Do you keep AI drafts separate from the final, approved version? | Versioning of drafts and approvals | Product documentation |
H5 | Can records flow to our existing archive, such as email archiving? | An integration or an export is available | Integration list |
I. Contract and exit
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
I1 | Who owns our data and the outputs? | We do | Contract clause |
I2 | Can we export all of our data in a usable format? | Yes, in a defined format | Export documentation |
I3 | How soon is our data deleted after we leave, and can you certify it? | A defined period, such as 30 days, with certification | Contract clause |
I4 | Are price increases and usage limits fixed in the contract? | Capped and stated in the order form | Order form |
I5 | What service levels and support hours do you commit to? | A written SLA | SLA |
J. AI governance
# | Question | A good answer looks like | Evidence to request |
|---|---|---|---|
J1 | Do you have an AI governance policy, and who is accountable for it? | A written policy with a named owner | Policy |
J2 | How do you keep a person in the loop for client-facing output? | Review steps before anything reaches a client | Product demo |
J3 | How do you handle model errors that customers report? | A defined process and response times | Support policy |
J4 | Do you make marketing claims about accuracy or compliance, and can you support them? | Specific, substantiated claims | Documentation |
Step 3: Ask for evidence, not just answers
Most vendors will say yes to most questions. The evidence is what separates a yes from a promise. Ask for:
A current SOC 2 Type 2 report and, if its period ended some months ago, a bridge letter.
A summary of the latest independent penetration test and how the findings were fixed.
The subprocessor list, including model providers, with each one’s function and location.
The data processing agreement and master services agreement.
A security overview with a data-flow diagram for your data.
The incident response summary and the breach-notification clause.
A business continuity and disaster recovery summary.
The AI governance policy, including how the vendor tests accuracy and handles model changes.
A certificate of insurance showing cyber and professional liability cover.
Written confirmation of the model providers’ terms on training and retention.
Many vendors publish part of this in a trust center and keep the sensitive documents behind a request. Anthropic’s, below, lists its attestations by product, its subprocessors and a FAQ, and makes sensitive documents available on request.


Step 4: Score the answers and decide
Rate each area pass, conditional or fail. Pass means the answer came with evidence. Conditional means the answer is acceptable but a gap needs a contract term or a compensating control, such as a use restriction. Fail means the gap is unacceptable for the data involved.
These are red flags worth stopping on:
No written commitment to tell you within 72 hours of becoming aware of a breach.
Your data may be used to train models, or the vendor cannot show that its model providers do not retain it.
No subprocessor list, or no notice before subprocessors or models change.
No independent attestation, and no willingness to share a penetration-test summary.
No audit log, or no way to export records for an exam.
You cannot export or delete your data when you leave.
The AI can send, write or delete without an approval step you control.
A compliance label such as “SEC compliant” or “Reg S-P compliant” is a prompt for a question, not a conclusion. Ask what the label covers and what documents support it. Software can support your compliance, and your firm remains responsible.
Record the decision so you can show it later:
Field | What to record |
|---|---|
Vendor and product | Name, version or plan, and the date |
Tier and why | The tier, and what the tool can reach |
Systems connected | Each system and the data it exposes |
DDQ | The date sent, the date returned and who reviewed it |
Evidence | Each document received, with its date |
Gaps | Each gap, and the contract term or control that covers it |
Approval | Who approved, and when |
Next review | The date, and the events that would trigger an earlier one |
Step 5: Secure the contract terms
Put the important answers in the contract, not only in the DDQ:
Breach notice within 72 hours of the vendor becoming aware, and cooperation with your own notifications.
No training on your data by the vendor or its model providers.
Zero data retention, or a stated short period, with each model provider.
Advance notice of new subprocessors and model providers, with a right to object.
Cooperation with SEC examinations and access to your records.
Export of your data in a usable format, and deletion on a stated timeline with certification.
Capped price increases and stated usage limits.
Insurance, and a limit of liability that fits the data at stake.
Step 6: Keep monitoring
Re-review each Tier 1 vendor at least once a year, and sooner when something changes. Triggers include a new subprocessor or model provider, a breach, an acquisition, a new feature that widens what the AI can read or do, and a new system you connect. Keep a calendar of review dates and a folder of the evidence.
What an exam-ready vendor file contains
Your written policy for overseeing service providers.
The tier and decision record for each vendor.
The completed DDQ.
The evidence you received, with dates.
The signed contract and any amendments.
A monitoring log: reviews, notices received and actions taken.
Your incident contacts for each vendor.
Frequently asked questions
What is a DDQ?
A due-diligence questionnaire is a standard set of questions you send a vendor before you connect it to your data. For an RIA it covers security, data handling, incident response, subprocessors, records and exit. The answers, and the evidence behind them, form the core of your vendor file.
Does Regulation S-P require a SOC 2 report?
No specific attestation is named in the rule. It requires written policies for overseeing service providers, including due diligence and monitoring. A current SOC 2 Type 2 report is a common piece of evidence because it is an independent review of a vendor’s controls, but check with your compliance officer or counsel what your policy should require.
How often should we re-review an AI vendor?
The rule does not set a frequency. At least once a year for any vendor that touches client information is a common practice, with an earlier review when a vendor changes subprocessors or models, has an incident, is acquired, or adds features that widen what it can access.
Do we need to do this for ChatGPT or Claude?
Yes, if client information goes into them. They are likely to be service providers under the rule’s definition, like any other AI vendor, and the plan you use matters. See our guide on using ChatGPT or Claude with client data.
What if a vendor will not answer the questions?
Treat it as a red flag and document it. A vendor that will not say how it handles your clients’ information, or will not put its breach-notice commitment in writing, has given you the answer you need.
Go deeper
Can financial advisors use ChatGPT or Claude with client data?
Regulation S-P amendments: what RIAs must do about AI and other service providers
Updates to this guide
November 17, 2026: First published.
Sources
Checked November 5, 2026. 17 CFR 248.30 (Regulation S-P), FINRA’s Regulation S-P compliance-date reminder, Goodwin’s alert on the $1.5 billion threshold, Anthropic Trust Center.