# AI Tools and Client Financial Data: Rules for a Practice

- Author: AutoBankConvert (https://www.autobankconvert.com)
- Published: 2026-02-15
- Updated: 2026-02-15
- Category: Security and trust
- Canonical: https://www.autobankconvert.com/blog/ai-tools-and-client-financial-data

> Using AI tools with client financial data: which tools train on your uploads, what to strip first, and a practice policy for statements and client books.
AI tools can help with client financial data, but only under rules. Use business or API tiers that do not train on your inputs by default, strip names and account numbers before anything goes into a general chatbot, and keep whole statements in tools built for them. Write those rules down so every staff member follows the same ones.

## Which AI tools train on the client financial data you upload?

Whether an AI tool trains on your uploads depends on the vendor, the plan you are on and your account settings, not on the product name. The same chatbot can treat a free account's chats as training data and a business account's chats as off limits. The only reliable answer is the vendor's current data policy for your plan.

Two published examples show how much the plan matters. OpenAI's [API data controls guide](https://developers.openai.com/api/docs/guides/your-data) says data sent to its API is not used to train or improve its models unless the customer opts in. It also says abuse monitoring logs are kept for up to 30 days by default. Anthropic's [help article on the model improvement setting](https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-setting) covers its consumer plans: users choose whether chats are used for training. Turning the setting off excludes new chats, while data already used in training is not removed.

That last point is the one bookkeepers miss. A setting changed today does not reach back. If a staff member pasted a client's transactions into a consumer chatbot last spring with training on, switching it off now protects future chats only.

For every AI tool in use at the practice, find answers to four questions and keep them on file:

1. Are inputs used to train models by default on our plan?
2. Where is the setting, and who in the practice controls it?
3. How long are conversations and uploaded files kept?
4. Who at the vendor can read them, and when?

Vendors change these policies. Record the date you read each one, and check again at your annual security review.

## Consumer, business and API tiers of AI tools compared

AI products usually come in three tiers: consumer plans for individuals, business plans for organizations, and APIs for developers. Business and API tiers tend to exclude customer data from training by default and give an administrator control. Consumer plans leave the choice to each user's settings, which a practice cannot see.

The table below summarizes the general pattern. It is a guide to what to look for, not a statement about any particular vendor. Confirm each tool's terms yourself.

| Tier | Typical training default | Who controls the setting | Fit for client financial data |
| --- | --- | --- | --- |
| Consumer (free or personal paid) | Varies; may train unless the user opts out | Each user, privately | Poor: the practice cannot verify settings |
| Business or team plan | Commonly no training by default | An administrator for the whole practice | Acceptable after reading the terms |
| API, through a tool built on it | Commonly no training unless opted in | The developer of the tool you use | Depends on the tool's own terms |

The third row needs care. Many bookkeeping products use an AI model through an API. The model vendor's API terms matter, but so do the terms of the product in front of it, because that product stores your files too. Ask the product vendor directly whether uploaded files are used to train any model, theirs or someone else's.

A consumer account used for work is the weakest case, even with training switched off. The practice has no record of what was pasted, no way to remove a departed employee's chat history, and no control over the setting.

## What to strip from client financial data before an AI prompt

Before any client financial data goes into a general AI tool, remove what identifies the client or lets someone act on the account. That means names, full account and routing numbers, addresses, tax identification numbers and anything that names an individual. The AI rarely needs any of it to answer the question.

Most useful prompts need the shape of the data, not the identity behind it. Here is an example. A bookkeeper wants help understanding an odd pattern of card refunds. The raw line reads:

"2026-03-14 REFUND JANE DOE ACCT 0001234567 VISA 4417 -212.50"

The stripped version keeps everything the question needs:

"Date 2026-03-14, card refund, -212.50. Refunds of similar size appear on the 14th of each month."

Use this list as the strip-before-prompt check:

- Client business name and owner names: replace with "the client".
- Account, card and routing numbers: remove entirely, including last four digits if the pattern is unusual enough to identify.
- Payee and payer names that are people: replace with a role, such as "a contractor".
- Addresses, phone numbers and email addresses: remove.
- Tax IDs, Social Security numbers, dates of birth: remove, always.
- Whole statement PDFs and screenshots: do not upload. A page image carries every field at once.

If stripping the data makes the question impossible to ask, that is a sign the task belongs in a tool built for client data, not a general chatbot.

## A practice policy for AI tools and client statements

An AI policy for a bookkeeping practice is a short written rule set: which AI tools are approved, which data may go into each, what must be stripped, and who approves exceptions. One page is enough. It belongs inside the practice's wider security plan, next to the vendor list.

The wording below is a template. Edit the examples to match your own tools.

1. **Approved tools.** "Staff may use [example: our business AI plan] and the AI features inside [example: our accounting software]. No other AI tool may receive client information."
2. **Consumer accounts.** "Personal or free AI accounts must not be used for any client work."
3. **What may go in.** "Stripped transaction patterns, chart of accounts questions and drafting help that contains no client identifiers."
4. **What may not.** "Whole statements, screenshots of client books, tax documents, payroll files and anything with a name, account number or tax ID."
5. **Settings.** "The administrator confirms training is off for our plan and records the date checked."
6. **Exceptions.** "Any other use needs written approval from the practice owner first."
7. **Incidents.** "If client information goes into an unapproved tool, tell the owner the same day."

If your practice is covered by the FTC Safeguards Rule, AI vendors are service providers like any other. The [FTC's guide to the Safeguards Rule](https://www.ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-business-needs-know) says covered businesses must choose providers that can protect customer information and must spell out security expectations in contracts. The post on [whether the FTC Safeguards Rule applies to bookkeepers](/blog/ftc-safeguards-rule-for-bookkeepers) helps you work out if that duty is yours.

Retention belongs in the policy as well. An AI vendor's copy of a conversation is one more copy of client data outside your files. The guide to [how long to keep client bank statements](/blog/how-long-to-keep-client-bank-statements) sets periods for your own records; aim for vendor copies to be gone well before those.

## When should client financial data stay out of AI tools entirely?

Some client financial data should never go into a general AI tool, however good its terms. Full statements, tax returns, payroll files, anything a client marked confidential, and any client who said no to outside software all stay out. So do tasks where a confident but wrong answer would reach a client unchecked.

The off-limits list for most practices:

- **Whole bank and card statements.** Use a tool built for statement conversion, where each client's files stay in their own workspace.
- **Tax documents and identity documents.** The harm from exposure is too high for any convenience.
- **Payroll data.** It names individuals and their pay.
- **Clients who declined.** If an engagement letter says no outside AI, that is the rule for that client.
- **Final numbers.** Never send a total or a tax figure to a client because an AI produced it. Check it against the books.

Volume is a trap of its own. A practice facing a large backlog may be tempted to feed statements through a chatbot one at a time. For a project like that, the guide to [converting hundreds of bank statements in one project](/blog/converting-hundreds-of-bank-statements) shows how to run it as a batch in a dedicated tool. AutoBankConvert's categorization works from rules you write once plus each client's past coding, and each client keeps their own accounts, rules and history, so one client's data does not shape another's books.

## Telling clients how AI tools touch their financial data

Clients deserve a plain account of whether AI tools see their financial data, which ones, and under what limits. Put it in two places: the privacy policy the practice publishes, and the engagement letter each client signs. A client who asks should hear the same answer from every staff member.

A short paragraph is enough. Example wording: "We may use AI features in our approved accounting and statement tools to help categorize transactions. Those tools do not use your data to train models on our plan. We do not put your statements, tax documents or identifying details into general AI chatbots."

The guide to writing a [privacy policy for a bookkeeping business](/blog/privacy-policy-for-a-bookkeeping-business) covers where a statement like this fits among the other disclosures. Keep it accurate. If you change tools or a vendor changes its terms, update the wording before the next engagement letter goes out.

Being specific builds more trust than being vague. A client who reads "we use AI carefully" learns nothing. A client who reads which tools, which data and which limits can make a decision. If you want statement conversion and categorization kept inside a per-client workspace rather than a general chatbot, [request a spot with AutoBankConvert](/get-started) and add its stated safeguards to your policy.