Is it safe to put company data into ChatGPT?

When I open up the Slido questions at a workshop, there's nearly always some version of 'what am I allowed to put in this thing?', often several times over. (They're the same question wearing different hats, and the question is 'am I about to get myself in trouble?')

So here's my short answer. Company data is reasonably safe in ChatGPT, Claude or Copilot when you're on a business plan that doesn't train on your chats, and you leave out any personal information the AI doesn't need. A client's details pasted into a free personal account is where people come unstuck, and upgrading to a paid personal plan changes less than most people expect.

What is the South Road billboard test?

It's the one question I get people to ask themselves before they paste anything into an AI tool. South Road is the long, busy, very unglamorous road that runs north to south through Adelaide (if you've sat in traffic on it at 5pm, you know), and I picture a billboard on it.

'If you put something into a free AI tool, ask yourself... would I be comfortable seeing this on a billboard on South Road tomorrow? If the answer is no, don't put it into a consumer/free tool without checking the data settings and your organisation's policy.'

Most of what people want help with passes easily. A draft newsletter, a page of meeting notes with the names taken out, a job ad, the 14th version of your About page... all fine on a billboard. A client's medical history, a staff member's performance review or your customer list with phone numbers attached doesn't pass, and none of it belongs in a personal account.

The other way I explain it is email.

'Think about AI like email. Email isn't inherently secure or insecure. It depends on the account, permissions, encryption, who has access and what you're sending. AI is exactly the same. What are you putting in? Who can access it? What can the AI access? And what happens to the data?'

Nobody bans email because someone once sent the payroll file to the wrong Dave. You set the accounts up properly and you teach people what not to send, and AI works the same way.

If you wouldn't want it on a billboard on South Road, it doesn't go in a personal AI account.

Does paying for ChatGPT or Claude make it safe?

Paying on its own doesn't fix it, because the setting that matters follows the type of plan. Personal ChatGPT and Claude plans, free or paid, can use your chats to train the provider's models. Business plans sit under commercial terms where training is off by default.

Here's where the big three sit, as checked on 17 November 2026. These pages change often, so click through before you rely on this.

Tool and planTrains on your chats?Where to check
ChatGPT Free, Plus and Pro (personal)Yes by default. Turn off 'Improve the model for everyone' in Settings, then Data controls.OpenAI, data controls
ChatGPT Business, Enterprise and Edu, and the OpenAI APINo, not by defaultOpenAI, data controls
Claude Free, Pro and Max (personal)Yes, if the model improvement setting is on. With it on, Anthropic keeps the data for up to five years, and 30 days with it off.Anthropic privacy centre, consumer terms update
Claude Team, Enterprise and the Claude API (commercial terms)No, not by defaultAnthropic, commercial products
Microsoft Copilot app with a personal Microsoft accountYes, unless you opt out under Privacy in Copilot's settings. Microsoft's FAQ describes the older app, and an updated app arrived on 18 August 2026, so check the settings in the version you use.Microsoft Copilot privacy FAQ
Microsoft 365 Copilot and Copilot Chat, signed in with a work accountNo. Prompts and responses aren't used to train foundation models.Microsoft, enterprise data protection

Two catches sit in the fine print. On personal ChatGPT and Claude accounts, a thumbs up or thumbs down can send that conversation into training even with the setting off. And ChatGPT's temporary chat keeps a conversation out of your history and out of training, while OpenAI can still hold it for up to 30 days for safety checks.

If client data is in play, the plan you want is the one with a business agreement behind it: ChatGPT Business, Claude Team or Enterprise, or Copilot through your work Microsoft 365 account.

Can I put customer data into Claude or ChatGPT?

You can, on the right plan, and usually with less of it than you'd think. Here's my answer when someone asks it in a workshop.

'I'd treat customer data the same way you would anywhere else. Ask: do I have permission to use it here, does our organisation allow it, and do I actually need the identifying information?'

The regulator ends up in a similar place. The OAIC's guidance on privacy and commercially available AI products says privacy law covers personal information you put into an AI tool and anything it produces about a person, and that putting customer information into an AI system has to fit the purpose you collected it for (Australian Privacy Principle 6). As a matter of best practice, it recommends organisations don't enter personal information, and particularly sensitive information, into publicly available AI chatbots.

That's where my golden rule comes in.

'If the AI doesn't need the personal information to do the job, don't give it the personal information.'

In practice that means stripping out names, emails, phone numbers and account numbers before you paste, then putting them back in yourself at the end. Ask Claude to draft a reply to 'a customer whose order arrived damaged' and you'll get the same email you'd get after pasting in Margaret's full address and order history. Health details, criminal records, religion and the other sensitive categories under the Privacy Act stay out on every plan, unless your organisation has a specific, approved setup for them.

Privacy policies are about to pick up a new job too, for any business using software to make decisions about people. I've written up what's due by 10 December 2026 separately.

What are the 4 P's of AI security?

It's the checklist I give people who want something shorter than a policy document, and the simplest way I know to remember AI security. I teach it in the data section of our Untangle workshops.

  1. Product. What tool/account are you using?
  2. Permissions. What can it access?
  3. Privacy. What happens to the data?
  4. Purpose. Why are you giving it the data in the first place?

'If you can answer those four things, you're already having a much better security conversation.'

Permissions is the one people skip. When you connect an AI tool to your email or shared drive, it generally works with whatever your account can see, and Microsoft says Copilot respects your existing permissions and sensitivity labels. So if your shared drive has a folder called 'Payroll FINAL do not open' that everyone can open (hate to say it, but most businesses have one), the AI can open it too. Fix the folder first.

A full room of people working on laptops during an AI workshop Jess Tresidder ran at Stone & Chalk in Adelaide
A full room, laptops open. People build during the session rather than watch slides.

What if the AI provider gets hacked?

If there's an unofficial office IT person in the room, this tends to be their question, and it's a good one.

'We're all doomed. Kidding. That's a legitimate risk. But it's not unique to AI. For instance we already trust banks, CRMs, email providers, cloud storage, HR systems and accounting software with sensitive information. The question isn't whether a system is capable of being breached. Almost every system is. The question is what controls exist to reduce the likelihood and impact (and what can we do to reduce sensitive information being leaked).'

The bracketed bit at the end is the part you control. Data you never pasted can't leak from anyone's servers, which is the golden rule doing its job again. The OAIC also suggests checking a product's security record before you adopt it, including a search for past incidents, so do that once for whichever tool your team settles on.

What should my team do this week?

  1. Find out which plan everyone is on. People using personal accounts for work is one of the most common gaps I see, and usually the easiest one to close (hate the word easiest, but it is).
  2. Switch off model training on any personal accounts that stay in use.
  3. Write a one-page 'never paste' list: client identifiers, health information, staff records, passwords and bank details.
  4. Run the 4 P's over every tool you connect to email, files or your CRM.
  5. Teach the South Road test at your next team meeting. It takes five minutes!

And if you're ever squinting at a paste and wondering whether it passes the billboard test, take the names out and ask again. South Road has enough going on already.

Questions people ask

Does ChatGPT's temporary chat make it safe to paste client data?

It helps for a one-off chat. OpenAI says temporary chats don't appear in your history, don't update memory and aren't used to train its models, but they may be kept for up to 30 days for safety purposes. The data still leaves your computer and reaches OpenAI's servers, so the billboard test still applies.

Is Microsoft Copilot safer than ChatGPT for work?

It depends which Copilot you're signed into. Copilot used with a work or school account comes under Microsoft's enterprise data protection, and Microsoft says those prompts and responses aren't used to train foundation models. The free Copilot app with a personal Microsoft account can use your conversations for training unless you opt out. Check the account before you compare the tools.

Do I need to tell customers we use AI?

The OAIC recommends updating your privacy policy and collection notices with clear information about how you use AI, and making sure any public-facing AI tool, such as a website chatbot, is clearly identified as AI. From 10 December 2026 there's also a legal requirement to disclose certain automated decisions that significantly affect people in your privacy policy.

Does turning off model training delete my old chats?

No. On ChatGPT, switching off 'Improve the model for everyone' stops future chats being used for training, but your saved chats stay in your history until you delete them yourself or delete the account. If you've pasted something you regret into a personal account, delete that conversation separately as well as changing the setting.

Sources

  1. OAIC, Guidance on privacy and the use of commercially available AI products
  2. OpenAI Help Center, Data controls in ChatGPT
  3. Anthropic Privacy Center, Is my data used for model training? (consumer)
  4. Anthropic Privacy Center, Is my data used for model training? (commercial products)
  5. Anthropic, Updates to consumer terms and privacy policy
  6. Microsoft Support, Privacy FAQ for Microsoft Copilot
  7. Microsoft Learn, Enterprise data protection in Microsoft Copilot and Copilot Chat