What should I ask an AI accounting vendor before buying?

At minimum: whether it holds a current SOC 2 Type II report and what it covers, whether your data is used to train any model, which model providers see your data and where, what the AI can do without approval, and what happens to your data when you leave. The tool above turns those into a list you can send.

Why ask about model training separately for each kind of model?

A vendor can truthfully say it doesn't let a third-party provider train on your data while still training its own models on your corrections, or on de-identified data shared across customers. Asking about each kind of model gets a complete answer.

Is this questionnaire legal advice?

No. It is a starting list of questions. The tax-data and Safeguards Rule questions reference official sources so you can read them, but how they apply to your firm is a question for your own adviser.

AI vendor due-diligence questionnaire

The questions to put to an AI accounting vendor before your data goes in. Tick what applies to your firm, then copy the list into an email or print it for the call.

Of the 72 tools in this directory, 34 don’t publish a clear answer on training with customer data and 26 don’t state a SOC 2 report. These questions get those answers in writing. Reviewed October 14, 2026.

12 questions.

Questions about AI and data handling for your product

Security and assurance

Do you have a current SOC 2 Type II report? What period does it cover, and are the AI features and your model providers within its scope or carved out?
Why it matters: A Type II report tests controls over a period. Scope matters: AI features or the model provider may sit outside it.
Which third-party AI model providers process our data, and are they listed in your subprocessor list? How will you notify us of changes?
Why it matters: Your data's real path often runs through a model provider you never contracted with.
Where is our data stored and processed, including by model providers? Can processing outside the United States be disabled?
Why it matters: Location affects which laws apply and, for tax data, what consent you need.

Data use and model training

Is any of our data, including prompts, uploaded documents, outputs and our corrections, used to train or improve models? Your own models, models shared across customers, or third-party models?
Why it matters: Many policies say 'not used to train' while still learning from corrections or using de-identified data. Ask about each kind of model separately.
If any training or improvement use is on by default, how do we opt out, who can do it, and does the opt-out cover data already collected?
Why it matters: Opt-outs are often per-workspace settings that a single admin can miss.
Do your model providers retain our prompts or outputs? For how long, and for what purpose?
Why it matters: Some providers keep prompts for a period for abuse monitoring unless zero-retention terms apply.
How is our data separated from other customers' data in search, retrieval, fine-tuning and caching?
Why it matters: Cross-customer leakage is the failure mode that matters most for client data.

Review and controls

What can the AI do without a person's approval? Who in our organization can change those rules, and is each change logged?
Why it matters: Autonomy settings are control settings. They should have an owner and an audit trail.
Does the audit trail show what the AI proposed, what a person changed, and who approved it? Can we export it?
Why it matters: Reviewers and auditors will ask who approved what. The log is your evidence.
How do we report AI errors, and what is your process when you find a systematic error that may have affected past output?
Why it matters: You want to hear about a pattern of errors from the vendor before you find it in a close.

Retention and exit

How long do you keep our data, and can we set retention to match our own record-retention policy?
Why it matters: Firms often have retention obligations that are longer, or shorter, than a vendor's default.
When we leave, what happens to our data, derived data and learned rules? Can we export our history and categorization rules first?
Why it matters: Rules a tool learned from your corrections are part of your process. Know whether you can take them with you.

Using the answers

Ask for each answer in writing with a link to the policy, trust-center page or contract clause that backs it up. Marketing pages change; a contract term or a dated policy is what you can rely on. Where a vendor’s answer contradicts its published policy, the published policy is usually what governs, so ask which one wins.

Before you send the questions, check the vendor’s directory listing. Each listing links to what the vendor already publishes about SOC 2, training and review, so you can skip what’s already answered and press on what isn’t.

Record the tools you approve in your firm’s AI policy; the AI acceptable use policy template has a register for them.

Questionnaire: questions

What should I ask an AI accounting vendor before buying?: At minimum: whether it holds a current SOC 2 Type II report and what it covers, whether your data is used to train any model, which model providers see your data and where, what the AI can do without approval, and what happens to your data when you leave. The tool above turns those into a list you can send.
Why ask about model training separately for each kind of model?: A vendor can truthfully say it doesn't let a third-party provider train on your data while still training its own models on your corrections, or on de-identified data shared across customers. Asking about each kind of model gets a complete answer.
Is this questionnaire legal advice?: No. It is a starting list of questions. The tax-data and Safeguards Rule questions reference official sources so you can read them, but how they apply to your firm is a question for your own adviser.

Sources

Official documents read on October 14, 2026. Vendor facts are sourced on each listing.