Using AI at work without leaking client data: a simple policy for small firms
If nobody at your company has set rules for AI, it’s almost certain someone has already pasted a client spreadsheet or a contract into a chatbot. Not out of bad intent: because it helps. The problem is that client data leaves your control, and your contracts and confidentiality duties still apply.
Banning it doesn’t work, because people keep using it quietly. What works is an approved tool and four clear rules. There’s a one-page policy template at the end that you can adapt.
Why pasting text into a chatbot is a data question
When someone pastes a client’s name, tax ID or account history into an AI tool, your company is sharing client information with a vendor. You answer for that even if an employee did it on their own.
What applies depends on your line of work:
- Client contracts and NDAs. Many restrict sharing confidential information with third parties. An AI vendor is one.
- Industry rules. The FTC Safeguards Rule covers financial businesses, including tax preparers; HIPAA covers health information; attorneys and CPAs have professional confidentiality duties.
- State privacy laws. California and a growing list of states regulate how personal information is shared and protected.
For a small firm, the most likely damage isn’t a fine. It’s losing a client who found out where their data ended up.
1. Use business accounts, not personal ones
The type of account matters more than anything else:
| Account type | Conversations used to train the model | Company control |
|---|---|---|
| Free or personal | Often allowed unless the user turns it off | None: the history lives in the person’s account |
| Business plan | Generally not, by contract | Central admin, retention and offboarding |
| AI on your own server | Doesn’t apply: data never leaves | Full |
Terms vary by vendor and by plan, so read the terms of the plan you’re about to buy, not what was true last year.
In practice: choose one approved tool, buy the business plan, turn on multi-factor authentication for the accounts and ask the team to use only that. If you already pay for Microsoft 365 or Google Workspace, you likely have an AI assistant under the same contract and the same protections as your email.
2. Decide what never goes in
A short rule works better than a long policy nobody reads:
- Never goes in: Social Security numbers, bank and card details, health information, passwords and documents of identified clients.
- Goes in without names: emails, contracts and spreadsheets, after swapping names and numbers for something neutral (“Client A,” “Vendor B”).
- Goes in freely: internal text with no personal data, templates, drafts and public information.
Swapping the names before pasting takes a minute and handles most cases. AI summarizes “Client A’s” contract just as well as a named client’s.
3. People approve what goes out
AI is confidently wrong: it invents clauses, swaps numbers and cites rules that don’t exist. Nothing that goes to a client, like a proposal, a reply or an invoice, should leave without a person reviewing it or without well-defined limits.
That goes double for automation. A workflow that drafts the reply and waits for someone’s approval is safe. One that answers on its own needs narrow rules: which topics, with what information and when to hand off to a human.
4. Know where your data flows
You should be able to answer three questions a client or an auditor might ask: what data, where and why. A simple table does it, one row per use, like this example:
| What we use it for | What data goes in | Tool and plan | Where it’s kept and for how long |
|---|---|---|---|
| Summarizing support emails | Client name and email | Assistant on the business plan | 30-day history, vendor’s US servers |
If you’re a tax preparer, this fits right into the written information security plan the FTC Safeguards Rule already expects you to have.
The blind spots: extensions, meeting recorders and personal apps
The chatbot is only the visible part. Three uses slip by:
- AI browser extensions that can read everything on screen, including your internal systems.
- Meeting recorders and transcription apps that store entire client conversations in services the company never chose.
- AI on personal phones, used with a photo of a document or a screenshot of a chat.
Cover all three in your rules. And give the team an easy way to request a new tool, or they’ll install it without asking.
A one-page policy template
Adapt it and agree on it with your team:
AI use at [company]
- Our approved tool is [name], on the company account. We don’t use personal accounts for work.
- We never put into AI: Social Security numbers, bank or card details, health information, passwords or documents of identified clients.
- Emails, contracts and spreadsheets go in only after names and numbers are swapped for neutral terms.
- Anything AI produces for a client is reviewed by a person before it goes out.
- Extensions, transcription apps and new tools need approval from [owner].
- If someone enters something they shouldn’t have, they tell [owner] the same day. Reporting it is never punished.
Item 6 matters more than it looks. Breach-notification deadlines in contracts and state laws are short, and you can only meet them if you find out quickly.
A good first step
Get the team together, ask without judgment which tools everyone already uses and for what, pick the approved one and agree on the template rules. It takes less than an hour and cuts the risk considerably.
AI rules work best on top of an organized base: multi-factor authentication and clean access in Microsoft 365. Accounting firms will find the same cautions built into the workflows in 5 automations every CPA firm should have.
If you want to go beyond the chatbot and use AI to automate processes with these safeguards built in, the free online assessment shows where to start. The technical side is part of our security essentials work.
This article is informational and isn’t legal advice.
Questions
Can we use the free version of ChatGPT at work?
For anything with no client data and no company secrets, yes. For anything involving client information, it's not a good idea: free and personal accounts often allow conversations to be used for model training, the company doesn't control the history and there's no contract protecting you.
Is there a law that covers this in the US?
There's no single federal privacy law, but several rules can apply depending on your work: the FTC Safeguards Rule for financial businesses, including tax preparers, HIPAA for health information, state privacy laws such as California's, and whatever your client contracts and NDAs say about sharing data with third parties. An AI vendor is a third party.
Do we need to tell clients we use AI?
Check your engagement letters and NDAs first: many restrict sharing client information with outside vendors, and some professions have their own guidance. A plain sentence in your terms or privacy notice saying you use technology vendors, including AI tools, under confidentiality terms covers most small firms.
Does running AI on our own server solve the problem?
It solves the sharing part: with the model on your own hardware, data doesn't leave for a vendor. Access control, human review and knowing what data is processed still apply. It makes sense when volume is high or the data is very sensitive.
What if someone already pasted client data into an AI tool?
Delete the conversation and turn off training on that account. Then work out what was exposed and whether a contract or law requires you to notify anyone; when in doubt, ask your attorney. Use the case to agree on rules with the team rather than to assign blame.