The AI Vendor Questions Every Financial Firm Should Ask Before Sharing Client Data

Every AI vendor in financial services is selling the same promise: faster work, fewer errors, happier clients.

But most of them need one thing from you to deliver it. Your client data.

That is where the conversation should slow down.

I just came back from ADVISE AI 2026 in Austin, hosted by Financial Planning and its publisher, Arizent. Credit to the Financial Planning team: they put more than 400 advisors, compliance officers, operations leads and technology vendors together. The advisors there hoped to find solutions. They seemed to want to be AI-forward. They are trying to navigate the how.

It was not a CPA crowd. But I work with CPA and tax strategy firms, and the concerns I heard in Austin echo conversations I hear from their leadership. The people in the room at ADVISE AI are caught between a rock and a hard place: they don’t want to be left behind, but they are held back from moving forward. When they sign vendor contracts, they also own the downside. Business owners, CFOs, COOs, compliance leads – this piece is for you.

I run a marketing firm, not a compliance practice. I handle the marketing piece, and that means I build: the systems that attract prospects, house them and nurture them until they are ready to talk. Those systems are not built to be compliance tools. When a prospect becomes a lead and shares private information, it is handed off to the firm. That does not make me the authority on your regulatory obligations; your compliance team and counsel are. But a lot of what I saw in the vendor marketplace last week looks very close to what I build, and that gives me a useful view of what is under the hood. What follows is what I hear, what I learn from training organizations and from AI companies themselves, and the questions I would want answered if I were in your seat.

In this piece:

• Why client data is where AI stalls
• The vendor questions you must get answered in writing, and the red flags my spidey senses watch for
• How to tell a product from a prompt 🤨
• The risk few firms are talking about: staff using their own AI
• A 90-day plan to get started without touching client data
• Where AI can go safely, and why I think marketing is the place to start
• How to measure whether AI is actually paying off

Client data is where AI stalls

The AI use cases that save the most time are the ones that need personally identifiable information (PII).

Reconciling statements. Drafting planning summaries. Reviewing returns. Answering client questions with account context. All of these functions runs on names, Social Security numbers, account numbers, balances and tax returns. In my own work, this is exactly what we try to keep out of the marketing layer.

Everyone is struggling with this. And a lot of people are pitching “we will roll out/implement AI for you.”

Before a financial firm hands that data to any outside tool, it needs to know where the data goes, who can see it and how long it is kept. For many firms, this is a legal requirement, not just good practice. SEC-registered advisers and broker-dealers must protect client information under Regulation S-P, which the SEC amended in 2024 to include oversight of the vendors that handle it. Tax preparers, CPA firms included, are limited in how they can use or share tax return information under Section 7216 of the tax code, and the FTC Safeguards Rule requires them to keep a written plan for securing client data. Above all, every firm owes that care to the clients who trusted it with the most sensitive numbers in their lives. 

I cannot help but notice many of the vendors I talk to build on the same underlying models, most often Anthropic’s Claude. The demos are impressive, especially if you don’t know how to build.  But when the question turns to compliance (where client data goes, who can see it, how it is archived, who is liable), I have not yet heard an answer I would want a financial firm to rely on.

So the firm hesitates, and the vendor builds around the gap. Neither gets the speed or accuracy the technology could deliver. I don’t see that hesitation as a failure of nerve. To me, it is a reasonable response to a question most vendors have not answered yet.

“We use Anthropic” is not a data policy

Many AI vendors in this space build on Anthropic’s Claude models, in part because Anthropic’s commercial terms say it does not train its models on customer content from its commercial services.

Do I take it on faith? No. I take it as the starting point for better questions.

Two distinctions matter:

• Training is not retention. “We don’t train on your data” does not mean “we don’t store your data.” Logging, retention periods and who can access stored prompts are separate questions.
• The model provider is not your vendor. The company you sign with sits between your data and the model. It has its own servers, its own employees, its own sub-processors and its own policies. Its terms are the ones that bind your data.

One attorney I work with suggested another route: running AI on Amazon Web Services (AWS), which, in that attorney’s experience, is what many lawyers use because it is highly secure and meets IRS and legal requirements. I am not endorsing any one setup. But it is a good reminder that how a vendor connects to the model matters as much as which model it uses. A lot of the vendors I encountered last week build their tools and platforms on Anthropic, and it is still the Wild West.

The vendor questions to get answered in writing

These are some of the questions you will want answered, in writing, before any client data touches an AI tool. Your compliance team and counsel will likely have more. Feel free to copy this list into your next vendor conversation.

Data handling

• Which model, and which plan? Consumer, team, enterprise and API accounts carry different data terms. Ask which one your data runs through. Even within one provider, plans differ. With Anthropic, for example, Team and Enterprise plans get the same protection against training, but only Enterprise lets a firm set how long chats are kept. By default, those chats are kept indefinitely unless an administrator sets a retention period.
• Training: Are our inputs or outputs used to train any model, by the model provider or by you?
• Retention: How long are prompts, outputs and files stored? Where? Is zero data retention available?
• Access: Which of your employees can see our data, under what circumstances, and is that access logged?
• Sub-processors: Who else touches the data? Get the current list and how you will be notified of changes.
• Connection: Do you connect to the model directly, or through a cloud service such as AWS? Whose account does our data sit in?
• Location: Where is the data processed and stored?

Security and evidence

• Independent assurance: A current SOC 2 Type II report or equivalent, not a marketing page about security.
• Is our data encrypted while it’s being sent and while it’s stored
• Breach notification: How fast will you tell us, and what will you tell us?

Compliance and recordkeeping

• Archiving: Can we save and retrieve AI conversations the way regulators require us to keep other client communications?
• Audit rights: Can we, or our auditors, verify what you say?
• Model changes: Will you notify us before switching the underlying model or provider?

Contract and exit

• Data processing agreement: Will these answers be in the contract, or only in the sales deck?
• Liability: Who carries the risk if client data is exposed through your system?
• Exit: If we leave, how is our data deleted, and how do we confirm it?
• Term: Can we start with a short pilot instead of a multi-year commitment?

If a vendor struggles to answer these, or will only answer them in conversation, I would take that as a signal to slow down.

Red flags I would watch for

• The vendor cannot tell you which AI model or plan your data runs through.
• Security answers live only in a slide deck, with no report or contract language behind them.
• The trial asks you to upload real client files before any agreement is signed.
• Claims that a tool is “FINRA approved” or “SEC compliant.” As far as I know, regulators do not certify AI tools, and your firm stays responsible for how it uses them.
• Pressure for a multi-year contract before you have run a pilot.
• No clear way to export or delete your data when you leave.

Know whether you are buying a product or a prompt

A lot of what is being sold as an “AI solution” is a well-written prompt, a document library and a connector to a model your firm can already access.

Because I build marketing systems with the same AI tools, the seams in a vendor demo show quickly: the underlying model, a system prompt, a document library, a connector, a form. Much of it is very close to what I build for marketing, and what I build is not designed to hold client data. When you have only seen one demo, it is easy to be impressed by the wrapper, and some firms end up paying monthly for something their team could set up fairly quickly.

That does not mean every vendor is overpriced. Some solve hard problems that are worth paying for: secure client data handling, compliance archiving, audit trails, and integrations with custodians and tax software. The point is to know which one you are buying before you sign.

A simple test: ask the vendor what their product does that your firm could not do with a business account on the same underlying model and a well-organized set of your own documents. A good vendor will have a clear answer. That answer is what you are paying for.

Caveat emptor. It is the Wild West right now, and the people approving purchases are usually busy running a firm, not testing how the product works.

Your team may already be using AI. Just not the firm’s.

Two things kept coming up in my conversations at ADVISE AI, and there was no working session on this.

The first: many people I talked to “know someone in the industry” who doesn’t worry about any of this and puts client information into AI anyway.

The second came up even more often, and it worries me more. Staff are using AI at work, but their own. Personal accounts, on personal plans, sometimes on personal phones.

I understand why. The tools are useful, and people want to work faster. But when an employee pastes a client’s statement or tax return into a personal AI account, the firm has no contract with that provider, no say in its data terms, no record of what was shared and no way to get it back. Consumer plans often carry different terms from business plans, and some use conversations to train models unless the user turns that setting off.

It isn’t just advisors. In Liberty Mutual’s 2026 study of independent insurance agencies, the share of agents using AI at work jumped from 37% to 65% in a single year, while only 18% said their agency had a well-defined AI policy. This is an issue across the board for business owners. 

I don’t think banning AI is the answer. In my experience, people reach for their own tools when the firm hasn’t offered a better one. Giving them one – an approved business account with clear data terms, a short policy on what never goes in, and one person who owns the decision – seems like a far safer path. Or the safest path right now. 

Three questions to ask your team this week

Ask them anonymously, and make it clear nobody is in trouble. You want honest answers.

1. Which AI tools have you used for work in the last month, and on which account?
2. What kinds of information have you put into them?
3. What would make your job easier if the firm gave you an approved AI tool?

The answers tell you where your exposure is today, and what your team actually needs.

A 90-day plan to get started without touching client data

Most firm leaders do not have the bandwidth to build AI tools or to track a technology that changes every week.

They are running businesses. AI moves faster than anyone with a full client load can follow. A tool that was best-in-class in the spring may be obsolete by fall, and a vendor that looked stable may be acquired or gone.

From what I have seen, that can push firms toward one of two places: paralysis (“we’ll wait until it settles”) or overcommitment (a long contract with the first vendor who made it sound simple). Neither really protects the business. A small, deliberate start can help. Here is one way to sequence it.

Days 1 to 30: get visibility

• Name an owner. One person accountable for AI decisions, even if it is a part of their role.
• Take inventory. Use the three questions above. Based on my conversations in Austin, the answer may be “more than leadership thinks.”
• Write a one-page acceptable-use policy (see below), and have compliance review it.

Days 31 to 60: give people a safe option

• Choose one approved business AI account for the firm, using the vendor questions above.
• Show your team how to use it, and what never goes into it.
• Start one marketing pilot with no client data. For example, turn one recorded webinar into a blog post, an email and a set of social posts, with your usual compliance review.
• Write down your starting numbers before the pilot begins (see the scorecard below).

Days 61 to 90: measure and decide

• Compare results to your starting numbers.
• Review what worked, and what didn’t, with compliance.
• Decide the next step: expand what is working in marketing, or move carefully to an internal use case with guardrails.

What a one-page AI policy can cover

• Approved tools, and on which accounts.
• What never goes in: Social Security numbers, account numbers, tax returns, statements, passwords, and client names alongside financial details.
• What is fine to use AI for: drafting from your own expertise, summarizing public documents, research, and marketing content.
• Review before release: anything going to a client or the public gets a human review, and compliance review where required.
• Recordkeeping: how AI-assisted communications are captured.
• Who to ask when someone is not sure.

Where AI can go safely: a map for financial firms

Here is how I think about sorting AI use cases with financial clients. It is a starting point, not a compliance framework, and the dividing line is simple: does the task require client data?

Start now (no client data)

• Educational content: blogs, newsletters, LinkedIn posts and video scripts built from your firm’s expertise, reviewed by your experts and compliance.
• Search and AI visibility: structuring your website and content so Google, ChatGPT and other AI tools can find, understand and cite you.
• Event and webinar production: outlines, slides, promotion, follow-up sequences and recap content.
• Repurposing: turning one recorded webinar or podcast into articles, clips, emails and FAQs.
• Market and competitor research from public sources.
• Internal briefings from public documents: summarizing tax code changes, regulatory notices and industry research for your team.

Next, with guardrails (internal or de-identified data)

• Meeting and call summaries, only with an approved tool, client consent where required and a review step.
• Internal process documentation: SOPs, checklists and training materials with no client details.
• De-identified analysis, with your compliance team defining what “de-identified” means.
• Sales and service insights from internal call recordings, using an enterprise account with clear data terms.

Wait, or proceed only with a vetted vendor and written approval

• Anything involving client PII, tax returns, account numbers or statements.
• Client-facing AI that gives answers or recommendations without human review.
• Automated decisions about suitability, eligibility or advice.

Why marketing is the safe place to start

The marketing layer is the most flexible and least restrictive part of a financial services business, and it is where AI can do real work today without touching client data.

Marketing content is built from your firm’s expertise, not your clients’ files. It is meant to be public. It already goes through a review process. That makes it the lowest-risk place to put AI to work while the harder data questions get answered.

It is also where a firm can find out whether AI actually pays off. Return on investment is one of the concerns I hear most, and the industry surveys reflect it: in Inside Public Accounting’s 2026 IT Survey, nearly 43% of accounting firms said they had not yet seen a measurable impact from their AI tools, and in a small Wolf & Company survey of community bank leaders, 30% cited unclear ROI as a challenge. Marketing is a lower-risk place to answer that question for your own firm, because the results are measurable.

And the work is needed, because search has changed. Google now leads with AI Overviews and AI Mode. Prospective clients ask ChatGPT, Gemini and Claude for recommendations before they ever visit a firm’s website.

At MAICON in Cleveland the week before ADVISE AI, Wil Reynolds of Seer Interactive cited a study published in Kevin Indig’s Growth Memo: in Google’s AI Mode, 74% of buyers’ shortlists came straight from the AI’s answer, and about 26% overrode the AI’s ranking, usually for a brand they already recognized. AI models look for “receipts”: consistent evidence across your website, your content, the press and the people who vouch for you.

For a financial firm, I believe that means the ones publishing clear, specific, expert education are the most likely to be recommended. Firms that stay quiet risk being left out of the answer.

Least restrictive does not mean unrestricted. RIAs still answer to the SEC Marketing Rule. Broker-dealers still answer to FINRA Rule 2210. CPA firms still answer to AICPA and state board standards. AI does not change those rules. It changes how fast your team can produce work that meets them, and your compliance review stays in the loop.

How to measure whether it is working

Write down where you are before you start, then check the same numbers each month for three months. Some of these are easier to track than others; start with the ones you can.

Visibility: are more of the right people finding you?

• Searches for your firm’s name and your people’s names
• Website visits from search, and from AI tools such as ChatGPT and Perplexity (these often show up as referral traffic in your analytics)
• A monthly spot check: ask a few AI tools the questions your ideal clients ask, and note whether your firm is mentioned

Leads: are more of them reaching out?

• Inquiries, consultation requests and webinar or event registrations
• How many of those are a good fit for your firm

Prospect insight: do you know more before the first conversation?

• Where each lead came from
• What they read or watched before reaching out
• The questions they ask, which tell you what to publish next

Effort: is it saving time?

• Hours spent producing each piece of content
• How long compliance review takes

Those are numbers a firm can track within a few months, and they give leadership real evidence before anyone makes harder decisions about client data.

What I am telling my own clients

I feel the same tension I heard from so many people in Austin. I want my clients to get the full benefit of AI. I also never want to compromise their businesses, their licenses or their clients’ trust in any way.

So here is where I have landed, for now. Make vendors earn your client data, in writing. Give AI an owner and a policy. And start in the marketing layer, where your firm’s expertise is the raw material and nothing sensitive is exposed.

Authority is fragile. It takes years to build and one careless decision to damage. AI can help a firm build it faster, as long as leadership decides where AI is allowed to go.

Frequently asked questions

What should a financial firm ask an AI vendor before sharing client data?

Ask which model and plan your data runs through, whether it is used to train any model, how long it is retained and where, who can access it, who the sub-processors are, and what independent security evidence exists, such as a SOC 2 Type II report. Also ask about breach notification, recordkeeping, audit rights, notice of model changes, contractual liability and how data is deleted when you leave. Get the answers in the contract, not just the sales conversation.

Does Anthropic (Claude) train on customer data?

Anthropic’s commercial terms say it may not train models on customer content from its commercial services, which include its API and business plans. Its privacy center lists two exceptions: conversations a user rates with a thumbs up or down, which may be kept and used for training, and data a customer chooses to share. That covers training only. Data retention, logging and your vendor’s own storage and sub-processors are separate questions, and the vendor you sign with sets the terms that bind your data.

Is “we don’t train on your data” enough assurance?

No. Training and retention are different. A vendor can avoid training on your data and still store prompts, outputs and files for a period of time. Ask how long data is kept, where, who can access it and whether zero data retention is available.

Who should own AI decisions in a financial firm?

One named person, even if it is part of a broader role such as COO, CFO or chief compliance officer. That person keeps an inventory of AI tools in use, maintains a short acceptable-use policy and requires vendor due diligence before any tool touches client data.

Should a firm build AI tools or buy them?

Build when the task is simple and uses no sensitive data, such as content drafting or research. Many “AI solutions” are a prompt and a document library your team can set up. Buy when the vendor solves a hard problem you cannot: secure client data handling, compliance archiving, audit trails, or integrations with custodians and tax software.

Where can a financial firm use AI safely right now?

In marketing and in work built from public information. Educational content, search visibility, webinars, repurposing and research all deliver value without client data. Internal and de-identified work comes next, with guardrails. Anything involving client PII should wait for a vetted vendor and written approval.

Does compliance still apply to AI-generated marketing?

Yes. The SEC Marketing Rule, FINRA Rule 2210, and AICPA and state board standards apply to your communications no matter how they were drafted. AI changes how fast you produce content, not the rules it has to meet.

LeadGrow’s client results are documented by the people who hired us: 530% email list growth and 590% LinkedIn audience growth over a four-year engagement with SuGo Capital (Sarah Sullivan, CEO), and 60% subscriber growth plus a new buyer channel on YouTube for American Gold Exchange (Dana Samuelson, President). We have also conducted a comprehensive marketing audit for uDirect IRA Services, a self-directed IRA firm.

Work with LeadGrow: Every engagement starts with the Authority Blueprint™, a diagnostic of where your firm’s trust, visibility and content stand today. Book a strategy call