AI Finance Security Questions: The 20 Your IT Team Will Ask, With Answers

Anirudh Badam
Co-Founder and CAIO, Adopt AI14 November 2026

You chose the tool. Then IT sent a spreadsheet with 260 rows.
That review is normal, and it is fair. An AI agent that reads your bank feeds and your ERP touches the most sensitive data in the company. This guide collects the AI finance security questions that come up most often, grouped into five areas, with a direct answer to each. Hand it to your security lead before the first call and the review gets shorter.
We make Adopt, so the answers use our own product as the reference. Every vendor should answer the same questions in writing. If one cannot, you have learned something.
Data: where it goes and how long it stays
1. What data does the tool read? Only what you connect. Bank feeds through read-only connections, ERP records, and the statements or invoices you upload. Ask for the field list per connection.
2. Where is the data stored, and in which region? Ask for the hosting region in writing. Adopt offers cloud, a customer VPC, full on-prem, and a single-tenant cloud in a chosen region when data residency is a requirement.
3. How long does the vendor keep it? Get a retention schedule and a deletion process, with the date a deletion request completes. Treat a vague "we delete on request" as an open item.
4. Is data encrypted in transit and at rest? This should be a yes with the standard named. It is a one-line answer, so ask for the line.
5. Does the vendor train models on our data? No. Adopt does not train on customer data, and its contracts with model providers carry zero-data-retention terms. Ask any vendor to show the clause.
Models: what the AI can and cannot do
6. Does the LLM touch the numbers? It should not. In Adopt, models write the format-conversion code, humans review that code, and deterministic code does the math. A model that sums a column will eventually sum it wrong.
7. What happens when the model is wrong? Ask for the failure path. In our own tests, a model that is right more often can still be worse, because it guesses where a more careful one stops. See the 97 percent model for the numbers. Look for abstention, escalation counts, and a reviewer queue.
8. Can we use our own LLM keys? Adopt supports bring-your-own keys. This matters to IT teams with an existing agreement and spend controls.
9. Which model providers see our data? Ask for the list of subprocessors and the terms with each. Vendor lists change, so ask for the date of the current one.
10. Can we see what the agent did on a given run? Ask for a per-run trace: inputs, steps, tokens, cost, and output. Adopt records a provenance file of source data, computations, and decisions for each agent output.
Access: who can see and do what
11. Does the tool respect our permissions? Adopt inherits role-based access from the source systems. A user who cannot see a ledger in the ERP cannot see it through the agent.
12. How are credentials stored? In a secret vault, never in prompts or logs. Ask who at the vendor can read them, and how access is logged.
13. Does it write to our systems? Integrations are read-only by default. Writing to an ERP is enabled with customer permission and needs human approval before anything posts.
14. Can the agent move money? No. Payment execution is deliberately left out of Adopt.
Controls: what stops a bad entry
15. Who approves what the agent prepares? A named reviewer. Nothing ships without their sign-off, and the approval is recorded. Internal controls for AI covers the review gates that satisfy an auditor.
16. Can we set different levels of autonomy by task? Ask whether the tool supports a dial per task, from reading only to suggesting to acting with approval. Start every task at the low end.
17. Can an auditor trace a number to its source? Every figure should link to the transaction or document behind it. Ask to see this on screen with your own data, not a slide.
Deployment: where it runs and how you prove it
18. Which certifications does the vendor hold? Adopt holds SOC 2 Type 2 and ISO 27001, and reports go out under NDA. It is compliant with HIPAA, GDPR, and CCPA. Say "compliant" for laws and "certified" for audited standards, and check that your vendor does the same. Adopt does not hold a SOC 1 Type 2.
19. Can it run inside our cloud? Yes, through a customer VPC deployment or full on-prem. If the agents call a hosted model API, the data crosses your boundary at that call, and the vendor should say so. Ask for the architecture diagram.
20. What does setup need from IT? For Adopt Cloud, roughly three to four short calls and three to five hours in total, mostly approving connections to your ERP and single sign-on. On-prem takes longer, around four to six weeks of sessions.
How to use this list
Send the 20 questions to every vendor on your shortlist before demos. Score the answers in writing. Then run a pilot on a month you have already closed, with nothing posting, so IT can inspect real behavior instead of promises. The plan in how to run an AI pilot in your finance team shows how to sequence that test, and evaluating AI finance tools covers the wider buying criteria. For the data question in detail, read AI accounting data privacy.
To start the review with answers on file, book a pilot and we will send the security packet to your IT lead the same day.
FAQs
Related Articles

The Accountant Shortage in Numbers: What It Means for Your 2027 Close

How to Automate Invoice Data Extraction and Categorization from PDFs

See it running on your workflows
Automate your accounting workflows

