This week I sat with a question from small accounting and bookkeeping offices: is it safe to put client data and AI in the same room? A tax preparer wants to check a finished return for mistakes. A bookkeeper wants monthly reports built from bank statements. An office manager chases payroll hours by email. All of these involve someone else's private information.
The honest answer is that the bigger risk isn't the AI. It's what your staff are already doing with it.
Is your office already putting client data into AI?
Probably, and nobody wrote down any rules for it. A 2026 survey of accounting firms, published by CPA Practice Advisor, found:
- 56% of firms use generative AI
- 46% report staff putting confidential client data into public AI tools
- about 90% have no written AI policy
So the choice isn't really "AI or no AI." The choice is between AI with rules and AI without them.
Which AI account your office uses matters more than which tool
Most people don't know that the type of plan decides what happens to what they type in.
Consumer plans, the personal accounts people sign up for on their own, often use chats to improve the model unless someone finds the setting and turns it off. Business plans usually don't train on your data by default. A staff member on a personal account and a staff member on the company's business account can type the same thing and take on very different risks.
My recommendation to offices is simple:
- Give staff business seats, paid by the firm
- Ban personal accounts for client work
- Switch off the optional feedback features at the account level
This step is boring and cheap, and it covers most of the risk.
The rules already apply to you
Tax preparers are not working in a grey area. In June 2026 the IRS told practitioners to handle client data only through secure, firm-approved AI. It also said firms need written AI procedures, staff training and vetting of outside tools.
Two more rules sit on top of that:
- IRC 7216 requires written client consent before tax-return information is used or disclosed outside narrow exceptions. The IRS has not said whether pasting a return into an AI tool counts as disclosure.
- The FTC Safeguards Rule treats tax preparers as financial institutions, so they need a written security plan that covers their service providers.
I'm not a lawyer, and I never tell a client "you're compliant." What I do tell them is that a short AI policy, a paragraph in the security plan and a consent clause reviewed by their own counsel are now normal office paperwork.
Can a masking tool hide client data before AI sees it?
Partly. I tested a free tool that runs locally on the office computer, finds personal details and swaps them for placeholders before anything is sent out.
On a practice tax return, the tool out of the box caught about 63% of the account numbers and IDs and about 53% of the names. Every name it missed was a non-Western name. After tuning, it caught 95% of the structured details and 93% of the names. It still missed one nickname.
That result taught me two things:
- Masking lowers the risk. It does not remove it. Names written in free text are the weak spot, and that is exactly where immigrant-owned businesses and their clients are most exposed.
- A person reads every masked file before it is used. That step is not optional.
I'd never sell masking as "nothing real ever leaves your office." It's a seatbelt, not a guarantee.
The habit that protects you most
Give AI only what the task needs. If you're checking a return for math errors, the AI doesn't need the client's name or Social Security number. If you're building a report from bank statements, it doesn't need the account number. Get your staff to ask "does the AI need this?" before every paste, and you'll prevent more problems than any tool will.
The takeaway: settle the account type, write the one-page policy and teach the minimum-data habit, and AI becomes something your office can use openly instead of something it hopes nobody asks about.
If your office is somewhere in that 46%, this is a good week to have that conversation.
