Guide

Australian Consumer Law and AI Replies in Your DMs

21 September 2026

Australian sellers ask a version of this question before automating anything: if the AI tells a customer something wrong, who wears it?

You do — and that's the whole answer. It's worth understanding properly before you switch automation on, because it changes which tool you should pick.

Under the Australian Consumer Law (Schedule 2 of the Competition and Consumer Act 2010), a representation made to a customer by your business is your representation. It doesn't matter whether a staff member typed it, whether it came from a template, or whether software generated it — there's no carve-out for automation, and none is coming. "The bot said it" has never been a defence and isn't one now.

That sounds alarming. It's actually quite manageable, because the risky answers in a DM thread are a short and predictable list.

This guide is general information about how the ACL applies to automated replies, not legal advice. For your own situation, check the ACCC and consumerlaw.gov.au, or talk to a lawyer.

Misleading conduct doesn't require intent

The provision that matters most for DM automation is section 18: a business must not engage in conduct that is misleading or deceptive, or likely to mislead or deceive.

The part sellers consistently get wrong is that intent is irrelevant. You don't need to have meant to mislead anyone. You don't need anyone to have actually been misled. If the conduct was likely to mislead, that's enough — an honest mistake, an out-of-date spreadsheet, and a confidently wrong AI answer all land in the same place.

Section 29 then sets out specific false or misleading representations, including claims about a product's standard, quality or grade, its price, and whether it's available at all.

Put those together and the picture for automated DMs is clear: an AI that guesses isn't a productivity feature — it's a compliance exposure that runs twenty-four hours a day and never gets tired of guessing.

The four DM answers that carry real risk

Almost every ACL problem in Instagram DMs comes from one of four questions — which happen to be the four questions customers ask most.

"Is this still available?" — Telling a customer something's in stock when it isn't is a representation about availability. If your AI answers from a product list that was accurate last Tuesday, it will eventually promise someone a sold-out item. This is the most common automated DM answer and the easiest one to get wrong.
"When will it arrive?" — Delivery estimates are representations about the future, and the ACL treats a claim about a future matter as misleading unless you had reasonable grounds for it. "Should arrive before Christmas" is a real promise, not a pleasantry — and the reasonable grounds need to be yours, not the AI's optimism. Seasonal deadlines are where this bites hardest.
"How much is it, is the sale still on?" — Quoting a price that's changed, or a discount code that expired last week, is a price representation. Was/now pricing draws particular ACCC attention, because a "was" price has to be one the item was genuinely offered at for a reasonable period — an AI reciting an old promotion can walk you straight into that.
"Will this fix my skin?" — Efficacy claims are their own category, and for skincare, supplements and anything health-adjacent you're potentially in Therapeutic Goods Administration territory as well as the ACL. Claims your marketing copy is careful about can be reinvented casually by a chatbot trying to be helpful. If you sell in these categories, this is the answer to constrain hardest.

Consumer guarantees you can't write your way out of

Separately from what your AI says, the consumer guarantees apply automatically to most retail sales — goods must be of acceptable quality, match their description, and be fit for any purpose you said they were fit for. For most consumer goods these guarantees can't be excluded, restricted or modified, whatever your terms page says.

Where a failure is major, the customer chooses the remedy, including a refund. Where it's minor, you can choose to repair, replace or refund. A "no refunds" or "no refunds on sale items" policy isn't enforceable against these rights, and stating one has itself drawn ACCC enforcement.

This matters for automation because refund questions arrive in DMs constantly, and an AI fed a strict store policy will repeat that policy confidently. A bot telling an Australian customer "sorry, all sale items are final" isn't just unhelpful — it's stating something the law doesn't permit you to state.

The practical fix isn't a better-worded policy. It's not having the AI adjudicate refunds at all: they're low-volume, high-stakes, and frequently involve facts the AI has no access to.

What compliant automation actually looks like

The pattern that keeps you safe isn't a more cautious chatbot — it's one that knows the difference between what it can verify and what it's guessing, and behaves differently in each case.

Answer from live data, not a snapshot. Stock and price answers should come from the store's current state at the moment of the reply, not a catalogue copy that syncs occasionally.
Re-check before committing. The gap between deciding on an answer and sending it is where prices change and the last unit sells — verifying just before sending is what stops a stale quote reaching a customer.
Escalate uncertainty instead of resolving it. When the AI can't ground an answer, handing the conversation to a human is the correct outcome. A confident guess is the failure mode that creates liability.
Keep some questions off-limits. Refunds, returns disputes, health and efficacy claims, and anything involving a specific delivery guarantee belong with a person.
Keep a record. Being able to show what was said, on what basis, and when is what turns a complaint into a manageable conversation.
Be honest that it's automated. A customer should be able to reach a human — which is also a Meta platform expectation, covered in our guide on what Meta allows for Instagram DM automation.

How Sauzy is built around this

We designed the reply pipeline around one assumption: a wrong answer costs more than a slow one.

Product answers are grounded in your live Shopify catalogue rather than generated from memory, so stock and pricing come from your actual store. Before a reply goes out, the assistant re-checks the product data it was about to quote against what's live in Shopify — if the price or stock has moved in the meantime, it stops and escalates rather than sending the stale figure. Discount answers are filtered by the promotion's real start and end dates, so an expired sale isn't quoted as current. Order status answers come from a genuine lookup against the order, not an estimate.

When the assistant isn't confident enough in an answer, it doesn't send a hedged version — it hands the conversation to you and emails you a link to it. You can take over any conversation from the inbox, and hand it back when you're done.

That behaviour exists because it's the commercially correct thing to do. It also happens to be close to what the ACL expects — not a coincidence, since the law is mostly asking you not to tell customers things that aren't true.

The short version

Automating your DMs doesn't change who's responsible for what your business says. It changes how fast your business can say it, which cuts both ways.

Pick a tool that answers from live store data, verifies before it sends, and escalates when it's unsure. Keep refunds, returns and health claims with a human. Those two decisions cover most of the exposure — and neither of them requires you to answer DMs at midnight yourself.

For the current detail on consumer guarantees, misleading conduct and penalties, the ACCC is the authoritative source, and worth checking directly rather than trusting any guide — including this one — to be current.