Current as of November 20, 2026
When we talk with firms about AI, I always ask the same question: is it in your WISP? (Don’t know what a WISP is? Read this other post.) The most common answer is a pause, then “...should it be?” No judgment. AI tools showed up faster than anyone’s policies could evolve, including ours.
The good news is that you don’t need to be a security expert to evaluate an AI tool. You need five good questions, and the confidence to expect clear answers. Here they are, along with what a good answer sounds like.
1. Where does my client data go?
Why it matters: You can’t protect what you can’t trace. Tax professionals are required to oversee the service providers that handle client information, and that starts with knowing where it goes. The FTC expects you to choose providers that can keep client data safe, and to make sure your contracts require it.
A good answer sounds like: “Here are the systems and models that receive your data, including any third-party model providers and their retention terms, here’s where it’s stored, and here’s how it’s encrypted in transit and at rest.”
A red flag: Vague answers, or “it’s all in the cloud.”
2. Is it used to train anyone’s model?
Why it matters: If client information is used to train a model, you lose control over where it ends up.
A good answer sounds like: “No. Your data is never used to train our models or any third party’s,” in writing, for the plan you’re actually on.
A red flag: “We take privacy seriously,” with no yes or no. Also watch for different terms on consumer and business plans.
3. What happens to names and Social Security numbers?
Why it matters: The safest client data is the data the AI never sees.
A good answer sounds like: “Personal information is detected and removed or tokenized before anything is sent to a model, and you can see what was removed.”
A red flag: “Our model is secure,” with no mention of what the model actually receives.
4. Who can access it, and for how long?
Why it matters: Access and retention are where many data problems actually happen.
A good answer sounds like: “Access is controlled by person and by client. Here’s our retention policy, and here’s how you delete client data when an engagement ends.”
A red flag: No retention policy, or no way to delete data.
5. Does it fit my WISP?
Why it matters: Under the FTC Safeguards Rule, tax and accounting professionals are required to keep a written information security plan, and AI vendors that touch client data belong in it. And separately, if tax return information goes into the tool, Section 7216 consent may be required. That one’s on you, not the vendor.
A good answer sounds like: “Here’s our security documentation, here’s our audit status (for example, SOC 2), and here’s how we’ll notify you of an incident, plus the contract language that commits us to all of it.”
A red flag: No documentation, or no incident notification commitment.
An email you can send to any AI vendor
Subject: Security questions before we use [Tool] with client data
Hi [Name],
We’re evaluating [Tool] for use with client information in our tax practice. Before we do, could you answer the following in writing?
- Which systems and AI models receive the data we upload (including any third-party model providers), and where is it stored?
- Is our data used to train your models or any third party’s models?
- Is personal information (names, SSNs, addresses) removed or tokenized before it reaches a model?
- Who can access our data, what is your retention policy, and how do we delete it?
- Can you share your security documentation, audit status, and incident notification commitments so we can document you in our WISP?
Thank you,
[Your name]
How we answer these at TaxFigure
We built TaxFigure so the answers are simple. PrivacyVault detects and tokenizes client PII before anything reaches a model. Every client lives in its own workspace with access controls. Data is encrypted at rest, in transit, and in memory. SOC 2 Type II is in progress, targeted for early 2027. Dashboards that show exactly what was redacted and tokenized are coming soon, so you can check for yourself. Until then, ask us and we’ll show you.
Hold us to the same standard you hold everyone else.
Download the checklist
Get the one-page AI vendor security checklist to use with every tool your firm uses. Download the checklist
Want the bigger picture?
Read our plain-English guide, Is It Safe to Put Client Data Into AI?, for what happens to data in a typical AI tool and what you’re already required to do.
Sources
- IRS and Security Summit, IR-2026-92 (Aug. 18, 2026), reminder on written information security plans: https://www.irs.gov/newsroom/irs-security-summit-remind-tax-pros-they-need-a-written-information-security-plan-to-protect-client-data
- IRS Publications 5708 and 4557 (irs.gov/pub/irs-pdf/p5708.pdf and irs.gov/pub/irs-pdf/p4557.pdf)
- FTC Safeguards Rule, 16 CFR Part 314: https://www.ecfr.gov/current/title-16/chapter-I/subchapter-C/part-314
This post is for general educational purposes and isn’t legal advice.