Building an AI chatbot for Australian users? What the 10 December 2026 privacy changes mean for your code
If you build chatbots for Australian businesses, one date belongs in your sprint planning: 10 December 2026. That is when the automated decision-making transparency rules added to the Privacy Act 1988 by the Privacy and
If you build chatbots for Australian businesses, one date belongs in your sprint planning: 10 December 2026. That is when the automated decision-making transparency rules added to the Privacy Act 1988 by the Privacy and Other Legislation Amendment Act 2024 start to apply.
Most support bots won't be caught by them: answering "what are your opening hours?" is not a decision about anyone. But many bots are drifting from answering questions to doing things: approving a refund, deciding who gets a call-back, pre-screening a rental or loan enquiry, waiving a fee. Once your bot does that with personal information, the business has disclosure work to do, and the code has to support it.
This post covers what changed, which parts touch the code, and a small Python pattern for logging decisions so the privacy policy can describe them accurately. It's general information for developers, not legal advice: the client's lawyer or privacy officer decides how the law applies to them.
Who the Privacy Act covers
The Act's 13 Australian Privacy Principles (APPs) apply to "APP entities": Australian Government agencies and organisations with an annual turnover above A$3 million. Some smaller businesses are covered too, for example private health service providers and businesses that trade in personal information. So a clinic's chatbot is in scope even if the clinic is small.
The OAIC (the Office of the Australian Information Commissioner) has published guidance on privacy and commercially available AI products. Its headline advice is that organisations should not enter personal information, and especially sensitive information, into publicly available generative AI tools.
What changes on 10 December 2026
The new rules sit in APP 1 (APPs 1.7 to 1.9), and the OAIC released its final guidance on them on 30 September 2026. From 10 December, an APP entity's privacy policy has to say when a computer program uses personal information to make a decision, or to do something that substantially and directly relates to making a decision, that could reasonably be expected to significantly affect an individual's rights or interests.
The policy has to describe:
- the kinds of personal information used by those programs
- the kinds of decisions made solely by the program
- the kinds of decisions where the program does something substantially and directly related to the decision
They cover decisions made after that date, even if the personal information was collected earlier. Your client can only write that honestly if someone knows which flows in the bot make or shape decisions. That someone is usually the developer.
Step 1: inventory the decision points
Walk through every tool or function the bot can call and sort them into three buckets:
| Bucket | Example | What to record |
|---|---|---|
| Information only | FAQ answers, opening hours, order status lookup | nothing extra |
| Assists a human decision | bot scores a lead or drafts a refund recommendation that staff approve | inputs used, recommendation, who approved |
| Decides on its own | bot auto-approves refunds under $50, auto-rejects an application | inputs used, rule or model version, outcome |
Anything in the last two rows is what the privacy policy has to describe. Write the inventory in the repo (a DECISIONS.md next to the tool definitions works well) so it stays current as tools are added.
Step 2: log decisions in a structured way
When the policy says "we use your order history and the reason you give to decide refunds under $50", you want logs that prove it. A minimal decision record:
from dataclasses import dataclass, asdict, field
from datetime import datetime, timezone
import json, uuid
@dataclass
class DecisionRecord:
decision_type: str # e.g. "refund_auto_approval"
mode: str # "solely_automated" or "assists_human"
data_categories: list[str] # kinds of personal info used, not the values
outcome: str # "approved", "declined", "sent_to_staff"
model_or_rule: str # model name or rule version
reviewer: str | None = None # staff member when a human decided
id: str = field(default_factory=lambda: str(uuid.uuid4()))
at: str = field(default_factory=lambda: datetime.now(timezone.utc).isoformat())
def log_decision(rec: DecisionRecord, sink=print):
sink(json.dumps(asdict(rec)))
log_decision(DecisionRecord(
decision_type="refund_auto_approval",
mode="solely_automated",
data_categories=["order_history", "stated_reason"],
outcome="approved",
model_or_rule="refund-rule-v3 (<= AUD 50, first claim)",
))
Notice data_categories records the kinds of information, not the values. That keeps the log useful for the disclosure without copying more personal information into another system, which is an APP 11 issue of its own.
Step 3: keep personal information out of the prompt where you can
Every prompt you send to a model provider is a disclosure of whatever is in it. If the provider processes data overseas, APP 8 (cross-border disclosure) applies, and the Australian business generally stays accountable for what happens to it.
Simple wins:
- Strip emails, phone numbers and card-like numbers before the text reaches the model, and re-insert them in your own code if the reply needs them.
- Pass the model an internal customer ID instead of a name and address when the tool call can look the record up itself.
- Check whether the provider offers a region or data-residency option, and write down which one you used.
import re
PATTERNS = {
"EMAIL": r"[\w.+-]+@[\w-]+\.[\w.]+",
"PHONE": r"(?:\+?61|0)[2-478](?:[ -]?\d){8}",
"CARD": r"\b(?:\d[ -]?){13,16}\b",
}
def redact(text: str) -> str:
for label, pattern in PATTERNS.items():
text = re.sub(pattern, f"[{label}]", text)
return text
print(redact("Call me on 0412 345 678 or mail [email protected]"))
# Call me on [PHONE] or mail [EMAIL]
Regex redaction is a first layer, not a guarantee. Test it with real (anonymised) transcripts from the client.
Step 4: retention and deletion
APP 11.2 asks organisations to destroy or de-identify personal information they no longer need. For a chatbot that means a retention period for transcripts agreed with the client, and a scheduled job that enforces it. Keep the decision records (which hold categories, not values) longer if the client needs them for complaints.
Step 5: give people a way to reach a human
The law doesn't require a human handoff in every bot, but it is the simplest way to deal with a customer who disputes an automated decision, and it lowers complaint risk. Since June 2025 Australia also has a statutory tort for serious invasions of privacy, which is one more reason not to let a bot be the only route.
Checklist before 10 December 2026
- [ ] Decision inventory written and reviewed with the client
- [ ] Structured decision logs for every "assists" and "solely automated" flow
- [ ] Redaction before model calls; provider region recorded
- [ ] Transcript retention period agreed and enforced by a job
- [ ] Human handoff on every decision flow
- [ ] Privacy policy text drafted by the client from your inventory
If you'd rather have a team that has done this before scope it with you, we build bots like this as part of our AI chatbot development in Australia work.
FAQ
Does a basic FAQ chatbot need to be in the privacy policy?
The new automated decision-making disclosure is about decisions that could significantly affect someone's rights or interests. A bot that only answers questions usually isn't making such decisions, but the business's general privacy policy should still cover the chat transcripts it collects.
Our client turns over less than A$3 million. Does this apply?
Usually the APPs don't apply to small businesses, but there are exceptions, such as health service providers and businesses that trade in personal information. Ask the client to confirm their status.
Is sending prompts to a US model provider allowed?
It can be, but it is a cross-border disclosure under APP 8, so the business needs to take reasonable steps about how the provider handles the data. Record the provider, region and contract terms.
Where are the official sources?
The OAIC website (oaic.gov.au) covers the APPs, the 2024 amendments and its AI guidance; the amending Act is on legislation.gov.au.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.