Staff using ChatGPT at work: what to do when you find out

What you have probably just found

On a Tuesday afternoon an owner glances at a screen over a team member's shoulder and sees a chat window with a client's name in it. Nobody was hiding anything. The tool was just open next to the quote.

If you have found one person, assume there are more. The typical pattern is a free or personal account, no announcement, and nobody ever having said yes or no. Common uses are rewriting emails, summarising long documents, drafting job ads, sorting out a spreadsheet formula and turning rough notes into minutes. Most of that is harmless. The risk sits in two places: which account it is, and what information went in.

You do not need every answer before you act. You need three: which tools, what kind of information, and roughly how often.

Think about what this looks like in your kind of business. In a property management office it is tenancy notices and reply emails. In a recruitment agency it is job ads and candidate summaries. In a bookkeeping practice it is client letters explaining why a figure moved. Each involves another person's information, which is why the habit deserves a conversation rather than a shrug.

Staff using ChatGPT at work: the first conversation

Have it privately, soon, and curious rather than accusing. Something like: 'I noticed you're using ChatGPT for the quotes. I'm not annoyed. I want to understand what it's helping with so we can set it up properly.' Then listen.

Ask what they use it for, which account they log in with, what they typically paste in, and what they wish it could do. Thank them for telling you. The person who answers honestly today is the person who will flag a problem next month, so protect that.

If sensitive material has gone into a personal account, make a note of what and roughly when before anyone deletes conversation history. You will want that record to judge the risk. Deleting first and thinking later removes the facts you need.

Resist solving everything in the first conversation. The aim is to learn, to show you are not hunting, and to agree on one small thing, such as 'until we have a rule, please keep client names out'. Follow up in writing, kindly, the same day.

Why a blanket ban usually backfires

The instinct to ban is understandable. It feels decisive and it is easy to write. It also tends to move the habit onto phones and home laptops, where you cannot see it and where the account is definitely personal. You lose visibility without removing the behaviour.

A ban also ignores why people reached for the tool. If the reason is that writing client letters from scratch takes an hour and the tool takes five minutes, the need does not go away because you disapprove. A narrow pause is different. Saying 'until we sort this out, nothing with client names or financials goes into any AI tool' is clear, temporary and easy to follow.

There is a fairness point as well. If nobody told staff what was allowed, punishing the first people caught tells everyone else that the rules are a trap. Make the rule first, then expect people to follow it.

Work out what has actually been going in

Keep a note as you learn. A simple list of tool, what it is used for and what goes in becomes the start of your tool register.

A single conversation with one person tells you about one person. To see the wider pattern, use a few methods together, because none catches everything.

Personal accounts on personal devices will never appear in your systems, so you have to ask. Make it safe to answer, which usually means anonymously and in groups, with a clear statement of what the results will be used for. Calibre Forum's survey is built around this: each person answers on their phone, and the owner sees results only in groups of three or more. A short form of your own can do a similar job if you are careful about anonymity.

• Ask everyone, in a way that does not feel like a trap.
• Check the third-party app permissions in your Microsoft 365 or Google Workspace admin console, or ask your IT provider to.
• Look at browser extensions installed on company machines, if you manage them.
• Scan card statements and expense claims for AI subscriptions.
• Ask your IT provider and your software vendors which AI features have been switched on in tools you already pay for.

Sort tasks into fine, needs a rule, check first and stop

Rather than judging tools, judge tasks. Four buckets cover most cases.

Test the buckets with a few staff before you publish them. They will spot tasks you did not think of, like running a lease clause through a chatbot to explain it, and they will tell you which rules sound fine on paper but are unworkable on a Friday afternoon.

• Fine: brainstorming, rewording your own text, explaining a concept, drafting a generic template with no identifying detail.
• Needs a rule: anything involving client names, internal financials, staff matters or unreleased commercial plans. These belong in an approved tool on a business account only.
• Check first: anything your client contracts, engagement letters or professional obligations speak to. Ask the named person before using AI on it.
• Stop: health information, identity documents, payment details, passwords and anything under a confidentiality clause or legal privilege.

Give people a sanctioned way to do the same thing

People use personal accounts because it is the easiest option on offer. If you want them to stop, make the better option easier. That might mean a business subscription whose terms you have read, with the admin settings checked, one named person to ask, and a one-page list of what is approved.

Run a short session using your own tasks. Take a real supplier email, a real file note with the details removed, a real job ad. Show what the approved tool does well and where it makes things up. People who have seen it confidently invent a clause number are more careful than people who have been given a rule.

Then write the rule down. A two-page policy with a list of approved tools and four or five plain clauses is enough to start with.

Decide who owns it. If the answer is 'IT', check that IT has the time. If the answer is 'whoever is free', nobody does.

If client information has already gone in

Stay calm and be methodical. Work out what the information was, whose it was, which account it went into and what that vendor's terms say about retention and training. Many cases turn out to be a client's name and a generic request. A few will be more serious.

If personal information is involved, Australian privacy law may be relevant, including the Notifiable Data Breaches scheme. Not every paste is a notifiable breach; the scheme is concerned with eligible data breaches likely to cause serious harm. Whether yours qualifies depends on the facts, so take advice if the information was sensitive or the volume was large. Check your client contracts too, because some require you to tell the client regardless.

Whatever you conclude, keep a short written record: what happened, what you assessed, what you decided and what you changed. If a client or insurer asks later, that record is worth more than a recollection.

Finally, tell your leadership team what you found and what you decided, so managers are not improvising different answers to the same question in different corners of the business.

Common questions

This guide is general information, not legal, insurance or financial advice. Check your own obligations with a qualified adviser.