Colour theme
Book a consultation

Keeping private data out of public AI tools

Set clear rules for AI prompts, uploads and connected apps. Verify data handling, test protection and give staff a practical way to report mistakes.

A laptop keyboard with a blue illuminated AI key.

An employee pastes a customer email into an AI chat to improve the wording. The task sounds ordinary, but the email may contain personal information, pricing or details covered by a client agreement.

Before staff use AI tools with work information, make the rules specific enough that they can decide what to enter. Finding AI tools staff use without the business knowing is covered in the shadow-AI guide. This guide focuses on the information leaving your organization.

Define what staff can submit

Separate public material, internal information and information that needs special handling. Give examples from your own work: customer records, payroll, health information, credentials, unpublished bids and proprietary code.

In a public tool the business hasn’t signed up for, use only public or made-up material. A business AI service still needs rules about what work information can go into it. Agreeing that a tool can draft public copy doesn’t mean it can process a client file.

The Privacy Commissioner’s generative AI principles address legal authority, appropriate purposes, safeguards and transparency. Review the law, contracts and sector requirements that apply to your organization before introducing personal information into a new use.

Verify the exact service and account

Data handling varies by service, plan, account and settings. OpenAI’s data-use page, for example, says its Free, Plus and Pro plans may use conversations for training unless the user switches that off, while its Business and Enterprise plans don’t by default. Don’t assume a paid account provides every protection your work requires.

Check training use, retention, deletion, human access, processing locations and connected services. Record the terms and settings reviewed, and make sure staff sign in with the business account, not a personal one.

For example, Microsoft’s enterprise data protection guidance says covered Copilot prompts, responses and Graph data aren’t used to train foundation models. It also explains that web queries have separate handling and that agents can have their own terms. That distinction needs review when enabling web search or third-party connections.

A no-training commitment doesn’t mean the service stores nothing, and it doesn’t settle whether you’re allowed to use personal information that way.

Reduce the input before sending it

Ask whether the tool needs the complete record. A fabricated example can help draft a template without uploading a live customer conversation. Remove credentials and unnecessary attachments, and check document comments, hidden sheets and metadata.

Replacing a name with an initial doesn’t guarantee anonymity. A rare job title, exact date or other details may still identify someone. If meaningful information remains private, use a tool your business has cleared for that kind of information, or do the task without uploading it.

Keep generated output within the same handling rules. An AI summary can repeat details from its inputs.

Test technical protection with synthetic data

Data loss prevention can help, but coverage depends on licences, managed devices, browsers, destinations and policy conditions. Microsoft’s AI data-leak deployment guidance describes controls for uploads, clipboard activity and browser prompts.

Pilot the policy with synthetic sensitive content. Test typed prompts, pasted text and file uploads separately. Include the browsers and devices staff use, and check whether the policy audits, blocks or permits an override.

Microsoft’s paste-control troubleshooting guide documents specific destination-group and classification requirements. A policy that blocks one upload path may miss a different input method. Record those gaps and give staff a workable alternative they’re allowed to use.

Give staff a clear reporting route

Practise with examples that don’t contain live personal information. Show staff how to sign in with the business account and where to ask about an uncertain input.

If private information is submitted to the wrong service, stop further sharing and report it promptly through the internal incident process. Record the service, account, time and information involved without spreading another copy. Ask the responsible team to assess available deletion, containment and any reporting obligations.

Keep monitoring proportionate and tell staff what’s recorded. For underlying Microsoft 365 access, the Copilot permissions guide addresses a separate part of the rollout.

Talk to us if you need help turning an AI-use policy into clear choices and tested controls.


Featured Image Credit