By Ammar Ashour
Before an accounting firm shares live client information with an AI vendor, it should understand the task, the data flow, the safeguards, and the exit plan. A polished demonstration shows what a tool can do. It does not show what happens to the information entered into it or whether the vendor fits the firm’s obligations.
The practical approach is to review one use case at a time and keep a short written decision record. The ten questions below can help a small accounting or tax team organize that review. They are educational prompts, not a certification, legal opinion, or complete information security assessment.
Why AI vendor review belongs before the pilot
An AI tool may receive prompts, documents, accounting records, or feedback. Depending on the product and configuration, that information may move through a vendor and other service providers. The firm should establish the actual data path before deciding what information is appropriate to share.
NIST’s Generative AI Profile, published in July 2024, identifies data privacy and value chain integration as risk areas. It notes risks from unauthorized use or disclosure of sensitive data and from unclear or insufficiently vetted third party components. NIST describes the profile as voluntary risk management guidance, not a product approval.
For United States tax practices, the legal context also matters. IRS Publication 4557, revised May 2024, states that professional tax return preparers must create and enact security plans to protect client data. It also tells firms to select service providers capable of appropriate safeguards, require safeguards in contracts, and oversee their handling of customer information. The exact obligations for a particular firm depend on its activities and applicable rules.
Ten questions to ask an AI vendor
| Question | What a useful answer should clarify |
|---|---|
| 1. What exact task are we approving? | Name one use case, the intended users, and the output. “AI for accounting” is too broad to evaluate. |
| 2. What information will the tool receive? | List each data type, including uploaded files, prompts, connected records, usage logs, and user feedback. |
| 3. Where does the information go? | Identify the systems and service providers that store, process, or transmit it. Ask the vendor to document the data flow. |
| 4. Is our information used to train or improve models? | Clarify the default, available settings, contract terms, and whether different product tiers follow different rules. |
| 5. How is access controlled? | Ask about individual accounts, role based access, administrator privileges, multifactor authentication, and periodic access review. |
| 6. How is information protected? | Request current documentation about encryption in transit and at rest, isolation between customers, and security testing relevant to the service. |
| 7. What activity can we review? | Determine whether logs show who accessed data, what action occurred, when it happened, and which records were affected. |
| 8. What happens after a security incident? | Clarify notice timing, investigation responsibilities, available records, and the vendor contact responsible for escalation. |
| 9. How long is information retained and how is it deleted? | Ask about active systems, backups, derived data, account closure, and written confirmation of deletion when available. |
| 10. How do we leave or change vendors? | Identify export options, access revocation, deletion steps, contract termination terms, and the owner responsible for offboarding. |
Turn answers into evidence
A yes or no answer is rarely enough. Ask where the commitment appears and who is responsible for it. Evidence may include the contract, data processing terms, security documentation, administrative settings, a live demonstration, or an independent assessment. Match the evidence to the actual product and plan the firm will use.
FTC guidance on the Safeguards Rule says covered financial institutions should know what information they have and where it is stored, review access controls, evaluate third party apps, use appropriate safeguards, and oversee relevant service providers. The guidance also emphasizes that bringing in a provider does not transfer the firm’s responsibility for its security program.
Do not treat a general security page as proof that every requested control applies to your account. If the workflow needs a specific feature, such as an audit log or a setting that excludes data from model improvement, verify that feature inside the proposed plan and record the result.
Use a one page decision record
A small firm does not need a long report to make the decision visible. A one page record can contain:
- Owner: the person responsible for the use case and vendor relationship.
- Approved task: the narrow activity the tool may perform.
- Allowed information: the records that may enter the tool.
- Prohibited information: the records that must stay outside it.
- Evidence reviewed: the documents, settings, tests, and contract terms checked.
- Unresolved questions: the items that still prevent live use.
- Decision: approve a limited pilot, revise the design, or pause.
- Review date: when the decision will be revisited.
This record should connect to the firm’s broader security plan when one is required. It does not replace that plan. It helps prevent a limited demonstration from quietly becoming an open ended production workflow.
Keep the pilot fictional until material questions are answered
If the vendor cannot clearly explain data use, access, retention, incident handling, or deletion, keep the evaluation on fictional information. Do not use live client or taxpayer data merely to make the demonstration more realistic.
Hypothetical example: A firm wants an AI tool to draft missing document reminders from a tax organizer. The vendor can explain encryption and multifactor authentication but cannot confirm whether uploaded organizers are retained after account closure. The team records that gap and tests only with fictional organizers until the retention and deletion terms are resolved.
A pause is a valid decision. The goal is not to reject useful technology. It is to make the relationship between the task, the information, and the safeguards clear enough for a responsible choice.
Your next step
Choose one AI tool and one proposed use case. Complete the ten questions with the workflow owner and the person responsible for information security. If a material answer remains unclear, record it and keep live client data outside the pilot.
For the next layer of review, see AI Access to Accounting Software: A Practical Permissions Checklist. You can also browse the AI Insights hub.
Sources checked October 6, 2026. This article provides general educational guidance for United States accounting and tax workflow planning. It is not legal, cybersecurity, tax, or compliance advice. Cover image: AI generated editorial illustration.
