In April 2023, engineers at Samsung’s semiconductor division pasted internal source code and meeting notes into ChatGPT to get help with their work. Within weeks the company had banned generative AI tools on its devices and networks. Nobody involved meant to leak anything. They were trying to work faster.

Small firms face the same risk with client data, at a smaller scale and with a sharper professional duty attached. A paralegal summarizing a deposition, an accountant asking why a client’s return doesn’t reconcile, or a clinic manager rewording a patient complaint letter can each send identifying details to an outside company in a few seconds.

What happens when you press enter#

When someone types into an AI assistant in a browser or app, the text travels to the provider’s servers to be processed. What happens next depends on the product and the account type.

The provider may keep the conversation for a period, for abuse monitoring, for your own chat history or for longer. Some consumer plans use conversations to improve future models unless the user opts out. Business and API agreements often say they don’t train on customer data, and they set shorter retention periods. Any of these terms can change, and they apply to the specific plan the person used, which may not be the one the firm approved.

For a professional firm, the practical point is simple: once client information is in the prompt, it has left your systems and is governed by someone else’s terms.

The seven ways client data gets into prompts#

1. Pasting whole emails and documents#

The most common path. Someone copies an email thread to ask for a summary or a draft reply, and the thread carries names, signatures, phone numbers, matter numbers and every earlier message quoted below. The person was thinking about the last paragraph. The prompt contains all of it.

2. Uploading files#

Attaching a PDF or Word file sends more than the visible text. Files carry metadata such as author names, company names, file paths that often include client or matter names, tracked changes, comments and sometimes earlier versions. A contract that looks clean on screen can carry an opposing party’s comment history.

3. Context that identifies without a name#

Removing names helps, but facts can be just as identifying. “Our client, a dentist in Silver Spring who was sued by a former hygienist in March” points to very few people. Dates, locations, job titles, amounts and unusual events combine quickly.

4. Account and record numbers#

Bank account numbers, policy numbers, tax IDs, medical record numbers and docket numbers often appear in pasted text without anyone noticing. They are direct identifiers, and some of them are regulated data on their own.

5. Personal accounts on personal devices#

When a firm has no approved AI tool, staff use their own. Verizon’s 2025 Data Breach Investigations Report found that 14 percent of employees routinely accessed generative AI systems on corporate devices, a figure that leaves out use on personal phones entirely. Personal accounts run on consumer terms, with whatever training and retention settings the individual left in place.

Many assistants let users share a conversation by link. A link sent to a colleague can be forwarded, and anyone who has it can read the conversation, including the client details in the first message. Chat history also stays on the account, so a later breach of that account exposes every past prompt.

7. Connectors and browser extensions#

AI features that read your mailbox, calendar or document store, and browser extensions that “help” with every page, can send content to a provider without anyone pasting anything. Check what each connected tool can read before approving it.

What the professional rules say#

For lawyers, ABA Formal Opinion 512, issued July 29, 2024, is the main national guidance. It applies existing duties to generative AI. On confidentiality, it says Model Rule 1.6 covers all information relating to a representation, and it warns that self-learning AI tools may disclose that information, even inside a firm’s own closed system. For those tools, the opinion concludes that lawyers should get a client’s informed consent before entering information relating to the representation, and that boilerplate in an engagement letter doesn’t count as informed consent.

State bars have added their own opinions. Florida Bar Ethics Opinion 24-1, for example, recommends getting the affected client’s informed consent before using a third-party generative AI program if doing so would disclose confidential information.

Accountants, clinics and financial advisers have parallel duties under their own professional rules, under privacy laws such as HIPAA and state breach statutes, and under client contracts. Many outside counsel guidelines and vendor agreements now include clauses on AI use.

Controls that work for a small firm#

Write a short AI use policy#

One page is enough to start. Say which tools are approved, which accounts staff must use, what may never go into a prompt (client names, identifiers, privileged content, health information) unless the approved tool removes it, and who to ask when unsure. Have every staff member read and sign it.

Give people an approved tool#

Most leaks come from people trying to do their jobs. If the firm provides a tool that handles client data safely, staff stop reaching for personal accounts. Choose a business plan with written terms on training and retention, and turn on MFA and single sign-on.

Remove identifiers before text leaves the firm#

The strongest technical control is to strip names, email addresses, phone numbers and account numbers out of each prompt automatically before it reaches the AI provider. The model still gets enough context to help, with placeholders where the identifiers were. This catches the details people miss when they paste in a hurry.

A simple example of what the provider should see:

What staff type What the AI provider receives
Draft a reply to Adaeze Okafor (a.okafor@example.com) about the delay on account 4417-2210. Draft a reply to [PERSON_1] ([EMAIL_1]) about the delay on account [ACCOUNT_1].

Clean files before anyone uploads them#

Strip document properties, comments, tracked changes and hidden content from files before they go anywhere outside the firm, including AI tools. The same habit protects you when sending files to clients and opposing counsel.

Log and review#

Keep a record of which tools are used and by whom. A quarterly look at AI usage tells you whether the policy matches what people do day to day.

Train with real examples#

Show staff a pasted email thread and point out every identifier in it, including the ones in the signature block and the quoted messages further down. One concrete example does more than a list of rules.

Removing names isn't the whole answer

Automatic redaction catches names, emails and account numbers reliably. It can’t judge whether a combination of facts identifies a client, and it can’t decide whether a matter is too sensitive for AI at all. Keep a person’s judgment in the loop for those calls.

What to do when it has already happened#

If someone realizes they pasted client information into an unapproved tool, act quickly and calmly. Delete the conversation from the account’s history, check the provider’s settings for training and retention on that account, and write down what was shared, when and with which tool. Then tell the person responsible for confidentiality at the firm. Depending on what was disclosed and your professional rules, the firm may need to assess whether the client must be told. Treat it as a reason to improve the controls, so people report mistakes instead of hiding them.

A quick self-check for your firm#

Ask these questions this week:

  • Do we have an approved AI tool, and do staff know which one it is?
  • Is it a business plan with written terms on training and retention?
  • Does anything remove client identifiers before prompts leave the firm?
  • Do we have a written AI use policy that every person has signed?
  • Do our engagement letters or client communications address AI use where our rules require it?
  • Do we know what our connectors and browser extensions can read?

If you answered no to more than two, start with the approved tool and the policy. They take the least time and change behavior the fastest.

Questions people ask

If I remove the client's name, is the prompt safe?

Not always. Account numbers, email addresses, addresses, dates and unusual facts can identify a client as clearly as a name. Strip every direct identifier, and think about whether the remaining facts point to one person.

Do AI companies train their models on what I type?

It depends on the product and the account. Some consumer plans use conversations to improve models unless you opt out, while many business and API agreements say they don't. Read the current terms for the exact plan your staff use, and confirm retention periods as well as training.

Does ABA Opinion 512 ban lawyers from using AI?

No. Formal Opinion 512, issued July 29, 2024, says lawyers can use generative AI if they meet their existing duties of competence, confidentiality, communication, supervision and reasonable fees. For self-learning tools, it says lawyers should generally get the client's informed consent before entering information relating to the representation.

Is blocking AI tools the safest option?

Blocking tends to push staff toward personal accounts on personal phones, where the firm has no visibility at all. Most firms do better with an approved tool, a written policy and a technical control that removes identifiers before text leaves the firm.

Sources

  1. ABA issues formal opinion on use of generative AI tools (Robinson+Cole via JD Supra)
  2. ABA Formal Opinion 512 and confidentiality (ABA Litigation Section)
  3. Samsung bans staff's AI use after spotting ChatGPT data leak (Bloomberg via Business Standard, May 2023)
  4. Reflections on the 2025 Verizon DBIR (DomainTools)
  5. Florida Bar Ethics Opinion 24-1

Keep reading

This guide is general information for small firms. It is not legal, insurance or compliance advice for your situation.