Is ChatGPT safe for business?

This question usually arrives after somebody has already been using it. The honest answer is that the safety of ChatGPT for business work depends on three things: which account it’s being used on, what information is being put into it, and whether anybody in the business can find out afterwards.

The account type matters more than the tool

The same interface behaves quite differently depending on the subscription behind it. This is the single most misunderstood point, and it’s where most of the risk sits.

• Free and personal paid accounts: content may be used to improve the provider’s models unless the user has changed a setting. The business has no visibility over that setting, and no record of what was submitted.
• Business and enterprise accounts: providers generally commit that business data isn’t used for training, with administrative controls and retention settings. This is a real improvement, and it’s the minimum position for company work.
• Any account signed up personally by an employee: the agreement is between the employee and the provider, not the business, whatever the tier.

OpenAI’s own help centre confirms the split. On its Free, Plus and Pro plans it says “we may use your content to train our models” unless the user opts out, while ChatGPT Business, Enterprise and the API aren’t used for training by default.

An employee using a personal account for company work has, in effect, disclosed company information to a third party under terms nobody at the company has read. That’s the situation many businesses are actually in, and it’s different from the situation they think they’re debating.

Where the data goes

On a business account with training disabled, the practical risks are narrower but not zero. Data still leaves your environment and is processed on infrastructure you don’t control, most often outside Australia. It’s retained for some period. It’s reachable by the provider’s staff under their own controls. If your obligations to a customer, a regulator or an insurer include knowing where information is held, those are questions you now have to answer on somebody else’s behalf.

What it’s genuinely fine for

There’s a version of this conversation that overstates the risk, and it’s worth avoiding. Public AI tools are safe and useful for a large amount of ordinary work, on a properly licensed business account.

• Explaining a general concept, a standard or a piece of legislation
• Improving the wording of something that contains nothing confidential
• Drafting from a brief that carries no customer, pricing or people detail
• General research and background reading
• Code, formulas and technical help that contain no business data

Where it stops being appropriate

• Customer pricing, margin, cost models and rate cards
• Contracts, tenders and negotiation positions
• Board papers, strategy documents and anything market-sensitive
• Employee, remuneration and personal information
• Anything a customer contract says must stay in a particular jurisdiction

On personal information the regulator has already given its view. The OAIC recommends, as a matter of best practice, that organisations don’t enter personal information, and particularly sensitive information, into publicly available generative AI tools.

The difficulty with this list isn’t writing it. It’s that the work people most want AI help with, like the contract, the price increase, the board paper and the tender, sits almost entirely on the wrong side of it.

The work people most want help with is the work they’re least able to paste into a public tool.

The problem nobody mentions: it doesn’t know your business

Security dominates this conversation, but there’s a second issue that costs businesses more day to day. A public model has never seen your P&L, your contracts, your cost model or your commercial history. Ask it where you’re losing margin and it will give you a competent general answer about manufacturers, which isn’t an answer about you.

People compensate by pasting in more context, which is exactly the behaviour the security policy is trying to prevent. The tension is structural: a general model becomes useful on specific work only when you feed it specific information.

What a reasonable position looks like

• Provide a properly licensed business account, so nobody has a reason to use a personal one
• Write down what must never be submitted to a public tool, in language people will actually follow
• For work that involves company information, use a private AI environment where the data stays inside your control
• Check what your own customer contracts say about processing location before deciding anything
• Keep a person accountable for the output. Neither tool removes that.

The Australian Signals Directorate’s guidance on engaging with AI makes the same point about the second item: staff who use an AI system should be trained on what data can and can’t be put into it.

The realistic destination for most Australian businesses is both. A public tool for general work, and a private environment for anything that touches company information. What doesn’t work is a policy that forbids the second category without providing anywhere else to do it, because that policy has already been quietly overtaken in many businesses that have one.