You can use a chat tool to draft customer replies without handing it the customer. Here is what the AI actually needs, what to strip out first, and a short rule your team can follow.
The short answer
Usually yes, once you take out the parts that identify the customer. A chat tool needs the substance of the question, your policy and the tone you want. It does not need the customer’s name, contact details, account numbers or anything about their health or finances. Swap those for placeholders, get the draft, then put the real details back yourself and read it before sending.
This is general information for Australian small businesses, not legal advice. If your business handles health, financial or other sensitive records, it is worth a conversation with your adviser about how AI tools fit your obligations.
What the AI actually needs to write a good reply
A chat tool writes a good reply from four things: what the customer is asking, what you sell or do, what your answer is (your policy, price, timeframe or decision) and the tone you want. None of those is the customer’s identity.
It is easy to assume the model needs the whole email to “understand” it. In practice, “a homeowner is asking whether our 90-day workmanship warranty covers a return visit” gives it everything it needs to draft the reply. A name, address and phone number add nothing to the writing and create a risk you did not need to take.
Two terms you will see in the guidance below. Personal information is, in plain terms, information about a person who can be identified from it, directly or by putting pieces together. The APPs are the Australian Privacy Principles, the rules in the Privacy Act that covered organisations must follow when they collect, use and store personal information.
The Office of the Australian Information Commissioner (OAIC) says privacy obligations apply to personal information you enter into an AI system, and to AI-generated output that contains personal information. As a matter of best practice, the OAIC recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. The Australian Cyber Security Centre (ACSC) gives small businesses the same practical advice: remove, anonymise or change personal details when using AI tools.
Keep, swap, never paste
The table below covers the common contents of a customer email. The rule of thumb: keep the question, swap the identifiers for placeholders, and leave out anything you would not want read aloud in a waiting room.
| In the customer’s message | What to do | Why |
|---|---|---|
| The substance of their question or complaint | Keep | This is what the draft is about |
| Your product, service, policy and the answer you want to give | Keep | The model cannot guess your warranty terms or your prices |
| First name or how they sign off | Swap for “the customer” or [NAME] | Tone does not depend on knowing who they are |
| Phone, email, street address | Remove | Never needed for a draft |
| Invoice, account, policy, job or client numbers | Swap for [INVOICE], [ACCOUNT] | You will type the real number back in afterwards |
| Exact dates that pin down a person or event | Loosen to “about three weeks ago” | Specific dates plus other clues can identify someone |
| Other people named in the email (a partner, a tenant, a child) | Remove | They did not write to you and have no reason to expect their details in a chat tool |
| Date of birth, Medicare, TFN, driver licence or passport numbers | Never paste | Identity details with no role in a reply |
| Health conditions, disability, mental health, medications | Never paste | Sensitive information, which the OAIC singles out |
| Financial hardship, income, Centrelink status, debts, legal disputes | Never paste | Reduce to the request itself, such as “asking to pay in two parts” |
| Attachments: ID scans, statements, medical letters, signed forms | Never upload | Describe what you need from them instead |
Removing names does not make a message anonymous. An unusual job, a small town, a specific date and a rare circumstance can identify someone between them. When the combination feels distinctive, loosen more than you think you need to.
A worked example: a leaking hot water system and an overdue invoice
This is a fictional example. The names, address, numbers and business are invented to show the method.
The email as it arrives
Hi, it’s Priya Natarajan from 14 Wattle Street, Ferntree Gully. Your team replaced the valve on our hot water system on 3 October (invoice 2047, $385). It’s leaking again from the same spot and there’s water across the laundry floor. I also got a reminder that the invoice is overdue. I’m on a disability pension and can only pay half this fortnight. Can someone come back this week? Call me on 0491 570 156.
What you paste instead
A customer of a plumbing business. We replaced a valve on their hot water system about three weeks ago under [INVOICE]. They say it is leaking again from the same spot and there is water on the floor. They have received an overdue reminder and are asking to pay in two instalments. They want a return visit this week.
Notice what went: the name, the address, the phone number, the invoice number and amount, the exact date, and the reason for the part payment. The reason is sensitive information and the draft does not need it; the request to pay in two parts is enough.
A prompt that gets a usable draft
Write a short, warm reply from a plumbing business. Apologise that the leak has returned. Explain that return visits on our own work are assessed at no charge within 90 days. Offer two time windows this week and ask them to confirm one. Agree to a payment in two parts, with the balance due after the return visit. Plain Australian English, no jargon, under 150 words. Leave [NAME], [INVOICE] and the time windows as placeholders.
The policy in the prompt is yours to supply. If your warranty is 60 days, say 60 days. The model will happily write a confident sentence about a policy you do not have.
When the draft comes back
- Paste it into your email client, not back into the chat tool.
- Replace [NAME], [INVOICE] and the time windows with the real details.
- Read the whole thing once as the customer would. Check the warranty wording matches what you actually do, the tone is right for someone with water on their laundry floor, and nothing has been invented.
- Send it yourself. The ACSC advises checking AI outputs for accuracy and keeping a person involved; this is the step that catches a wrong promise before it reaches the customer.
The whole exercise takes about a minute longer than pasting the raw email. For a fuller version of step 3, see our two-minute check for AI-drafted replies, with a worked example of what drafts typically get wrong.
Free accounts, business accounts and the settings to check
Whether your conversations are used to train the provider’s models depends on which product you are using and how it is set up. The ACSC notes that some providers use customer-submitted data to train their models, depending on configuration settings or the type of subscription, and recommends reviewing those settings and terms. Consumer accounts and business tiers are often configured differently, and the setting may have changed since you signed up.
As one example, OpenAI states that by default it does not train on business data from its ChatGPT Business, Enterprise and Edu offerings or its API (checked October 2026). We have not summarised other vendors here because the details differ and change; open your own tool’s data or privacy controls and read what it says today.
Three things to look for in any tool your team uses:
- Is there a setting that controls whether your conversations improve the model, and what is it set to right now?
- How long are conversations retained, and can you delete them?
- Who in your business is logged in, and are they using a personal account or one you control?
A business tier with training excluded by default is a sensible baseline. It is not a reason to start pasting full customer records. Even where a provider does not train on your data, the information still leaves your systems, sits on someone else’s servers and can be read by whoever shares that login.
Does the Privacy Act apply to my business?
According to the OAIC, most small businesses with an annual turnover of $3 million or less are not covered by the Privacy Act. There are exceptions. Health service providers must comply whatever their turnover, and the OAIC’s description is broad: it includes medical practitioners, pharmacists, allied health professionals, complementary therapists, child care centres and private schools. Businesses that trade in personal information, contracted service providers to the Commonwealth and some others must also comply. A small business that is not covered can also choose to opt in.
If you are under the threshold and none of the exceptions apply, the APPs may not bind you. Three reasons the habits in this article still matter:
- Customers expect care with their details whatever the law says, and trust is hard to win back in a local market.
- Other obligations may apply to you anyway: industry codes, professional standards, licensing conditions and the contracts you sign with larger customers often carry their own confidentiality rules.
- Your turnover and activities change. A habit built now will not need rebuilding if the threshold stops covering you.
If the Act does apply, the OAIC also expects transparency: update your privacy policy or notices to reflect how you use AI, and make sure any public-facing AI tool, such as a chatbot on your website, is clearly identified as one. The OAIC publishes a short checklist for choosing a commercial AI product that is worth working through before you settle on a tool for the whole team. From 10 December 2026, covered businesses that use software or AI to make significant decisions about people also need a new section in their privacy policy; our guide to automated decisions and your privacy policy explains what to include.
A one-page AI rule for your team
A written rule saves each person from guessing. The ACSC recommends an internal AI use policy that clearly defines what data can’t be uploaded. It does not need to be long. Here is a template you can copy, adapt and pin up:
- Approved tools. We use [tool name] on the business account only. No personal accounts for work.
- Never paste. Names with contact details, addresses, dates of birth, ID numbers, bank or card details, health or hardship information, or any attachment a customer sent us.
- Swap, don’t skip. Replace identifiers with [NAME], [INVOICE], [ACCOUNT] and loosen specific dates. Keep the question and our policy.
- Put the details back yourself. Add real names and numbers in the email client, never in the chat.
- A person reads every reply before it is sent and checks anything stated as fact against what we actually do.
- Not sure? Ask [owner or manager] before pasting. Nobody will be told off for asking.
- If something sensitive goes in by mistake, tell [owner or manager] the same day so they can decide what else needs to happen, including deleting the conversation and checking any obligations to the customer.
Walk the team through it once with a redacted example from your own inbox. Ten minutes of practice does more than a policy nobody opens.
This rule covers pasting customer messages. For a broader policy that also covers approved tools, checking output, account settings and how to introduce it, see our one-page AI policy template.
When a chat tool isn’t the right tool
Redaction works when someone pastes a customer email a few times a week. It stops working when the same staff member is pasting customer records all day and spending half their time swapping names in and out. At that point the chat window has become part of your workflow, and a general chat tool was never designed to be that.
The better fix is a properly configured business tool or an integration: something that connects to your existing inbox, job system or CRM with agreed access, retention and a review step, so the drafting happens inside systems you control rather than a browser tab. Before investing, map the workflow with our automation readiness checklist: how often it runs, who touches it, where it stalls and what a wrong reply would cost.
AppliedAgents builds this kind of workflow automation for Australian businesses, with the data-handling decisions made up front rather than left to whoever happens to be logged in. If you are at the “pasting all day” stage, that is the conversation worth having. If you are not, the habits above are enough.
Sources
- Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products
- Office of the Australian Information Commissioner, Checklist: privacy and the use of commercial AI products (PDF)
- Office of the Australian Information Commissioner, Small business
- Australian Signals Directorate’s Australian Cyber Security Centre, Artificial intelligence for small business (with NCSC-NZ and COSBOA)
- OpenAI, Enterprise privacy at OpenAI
This article is general information, not legal advice. Sources were checked on 26 October 2026. Product settings and provider policies change; check your own tool’s current data controls.
Practical thinking on AI, automation and the work behind a business.
