AI for NDIS Providers: What Is Safe Under the Privacy Principles?

An NDIS provider can safely use AI for work that contains no participant personal information: policy drafts, easy-read pages, job adverts, rosters, templates and training material. For anything with participant details, the Commission expects information to be de-identified and none disclosed to AI systems, and a person remains accountable for every record.

The rule that follows from the Commission's February 2026 position is narrower than many providers fear. AI is not banned. The Commission says it does not endorse or approve AI tools for developing and reviewing behaviour support plans, and that this does not stop providers using AI as long as they meet their legal obligations. What breaks the law is not the tool; it is sending identifiable participant information to a third-party platform without safeguards. The sections below turn that into safe, careful and closed jobs for a provider, with wording you can adopt.

Follow me on Instagram@sagnikteaches

What the Commission actually said in February 2026

The Commission published a four-page position statement titled Use of artificial intelligence in the development of behaviour support plans. It is short enough to read in full, and you should: the position statement on the Commission's website. In summary:

Connect on LinkedInSagnik Bhattacharya

• The current legislative framework does not explicitly prohibit AI.
• It lists risks: disclosure of participants' personal information to third parties, processing or storage of information overseas, a lack of transparency about how data is stored, secured or used once entered, inaccurate or misleading content (including content not developed in consultation with the participant or carers) and automated decision-making without human oversight or clinical judgement.
• Unmitigated, these risks may lead to breaches of the NDIS Code of Conduct and other provider obligations. The statement's example is that disclosing sensitive personal information to a third-party AI platform such as ChatGPT without appropriate safeguards may breach the Code of Conduct requirement to respect the privacy of people with disability.
• The Commission expects that, if a provider uses AI, all information is appropriately de-identified and no personal information of participants is disclosed to AI systems, and that use is consistent with Australia's AI Ethics Principles.
• Behaviour support plans must still be person-centred, evidence-informed, reflective of the person's needs, upholding the person's dignity and right to privacy, and developed in consultation with the participant, family, carers, guardian, providers and support workers.

The statement is about behaviour support plans, but the reasoning applies across a provider's work, because the same privacy duties attach to every participant record. It is not a complete legal analysis, and it does not replace advice from your own privacy or legal adviser. It does give you the Commission's expectations in its own words. If you write or review plans, what the Commission says about AI and behaviour support plans turns the statement into a practitioner's workflow; the rest of this tutorial applies it to a provider's everyday privacy duties.

Subscribe on YouTube@sagnikteaches

The Privacy Principles in four plain-English questions

The Australian Privacy Principles (APPs) govern how organisations collect, use, store and disclose personal information. The Privacy Commissioner's office has published guidance on commercially available AI products, and four of its points matter most to a provider.

1. Is typing it into a tool a use or a disclosure? Entering personal information into an AI system can count as a use (if it stays under your control) or a disclosure (if it goes to a party outside your control). Both are limited to the purpose the information was collected for, plus reasonable related purposes the person would expect (APP 6). A participant who gave you details to arrange supports did not obviously expect them to go into a public chatbot.
2. Is it kept safe? You must take reasonable steps to protect personal information you hold (APP 11). With AI that means checking the vendor's security, where the system runs and whether data-sharing features can be switched off.
3. Does it leave the country? Cloud AI often processes data overseas, which engages the cross-border rules (APP 8). The Commission's statement names overseas processing as a risk too.
4. Is the information sensitive? Health and disability information is sensitive information, and collecting or generating it generally needs consent that is informed, current and specific (APP 3). A privacy notice alone does not equal consent.

The same guidance says the best-practice recommendation is not to enter personal information, especially sensitive information, into publicly available generative AI tools, and to carry out a privacy impact assessment when considering an AI product. Read the guidance on privacy and commercially available AI products before you choose a tool.

Sorting provider jobs by the participant data they touch

Turn the principles into something a support coordinator can use on a Tuesday afternoon.

If you cannot decide, treat the job as closed until your privacy lead decides otherwise. For choosing which jobs to start with, the automation assessment checklist for NDIS providers scores jobs by time saved and data risk.

Safe lane: jobs with no participant information

Policies and procedures

Providers carry a policy load: complaints, incidents, privacy, worker screening, restrictive practices, feedback. AI can adapt a generic template to your size and services, flag missing sections against a checklist you provide and write a one-page staff summary. Say a five-person provider adapts a complaints policy. The assistant produces a draft and a flowchart in words; the director reads every clause against her own process and her registration obligations, deletes any promise she does not keep (such as a response time she cannot meet) and approves it with a version date. Tailoring policy templates with AI covers the detail.

Easy-read and plain-language material

Participants have the right to understand what you offer. AI can turn a service description into short sentences, suggest headings and list where pictures would help. A person who uses easy-read, or an advocate, should review the result, because it is easy for a tool to produce text that is short but still confusing. Plain-English and easy-read pages with AI has a method.

Recruitment and training

Job adverts, interview question banks, induction checklists and quizzes carry no participant data. A provider hiring support workers can ask for an advert in plain language, a scoring sheet and a first-week checklist. Check each requirement for fairness and accuracy, especially screening and qualification wording, which is regulated.

Careful lane: de-identified content and assessed tools

De-identification sounds simple and is not. Removing a name is not enough if the text includes a rare diagnosis, a suburb, a school, a family circumstance or a date that together point to one person. A good practice is a two-person rule: the writer removes details, a colleague who does not know the participant reads the text and says whether they could guess who it is. If they can, it is not de-identified.

Service agreements and shift-note structure

A generic service agreement in plain English is a good AI task: terms, cancellation, how to give feedback, with placeholders for names, supports and prices. Fill the placeholders inside your own system, and have a second person compare the draft with your registration obligations and price arrangements before it is used. For shift notes, AI can help design the template (headings such as what happened, support provided, participant response and follow-up needed) without ever seeing a real note.

Rostering rules and invoice checks

Rostering constraints such as "worker A cannot work with participant 4" can be expressed with codes and tested on dummy data first. For invoice and claim checks, checking claims against shift notes explains a method that keeps real identifiers inside your system and uses AI only on structure.

When a tool is assessed and contracted

Some providers will want AI inside their own software. The privacy impact assessment the Privacy Commissioner recommends is the right step: list what data the tool sees, where it is processed, who can access it, whether it trains on your data, how long it is kept and how you delete it. Ask the vendor for written answers. If the tool processes information overseas, ask exactly where. A tool you have assessed and contracted may be acceptable for careful-lane jobs; the Commission still expects human oversight and clinical judgement, and its position on behaviour support plans stays as it is: it does not endorse or approve particular tools.

Closed lane: where AI must not draft the substance

Behaviour support plans, restrictive-practice documentation, incident detail, risk assessments for named people and health-related records are closed. The reasons are in the Commission's list of risks. Content may be inaccurate or misleading, may not reflect consultation with the participant and their carers, and may lead to decisions without clinical judgement. A practitioner who uses a tool to generate strategies is outsourcing the part of the job that the Rules require to be person-centred and evidence-informed.

For practitioners, the Commission's statement is more specific: if AI is used at all in behaviour support, information must be de-identified and no personal information disclosed. A practical reading is that a practitioner may use AI for generic structure, such as a layout for a plan, or for plain-language editing of text that has been properly de-identified and that they have written, but not to generate the plan's content from participant information. When in doubt, the registered practitioner and the provider's governance lead decide in writing, and the decision is recorded.

Two before-and-after examples of de-identifying

De-identification is a skill, so practise it. Both examples are invented for illustration.

Example one, a shift-note template. Before (do not paste): a real note naming the participant, the suburb, the day centre and the diagnosis. After: "Participant [P1] became distressed at [location type] when [trigger type]. Worker used agreed strategy [S2]. Participant settled within [minutes]. Follow-up: none." The structure survives, which is all a template needs, and no fact about a real person does.

Example two, a policy question. Before: "Our participant Sam, who uses a wheelchair and lives alone in [suburb], had a fall; how should our incident policy handle it?" After: "How should an incident policy handle a participant who has a fall at home while supported by one worker?" The second question gets the same useful answer with nothing that identifies anyone. Ask yourself each time: does the tool need this detail to do the job? Almost always, it does not.

A useful prompt for the second pair, once the text is clean:

Draft an incident-response procedure for a small disability support provider. Situation type: a fall at a participant's home with one worker present. Include steps for immediate safety, who to call, what to record (headings only), how to inform the participant's family or guardian, and what the coordinator reviews afterwards. Do not cite laws or reportable-incident time limits; I will add those from the regulator's current page. Mark anything you are unsure about with [CHECK].

The output will include a sensible skeleton and several [CHECK] flags. The director adds the reportable-incident timeframes from the Commission's current guidance, and checks that the procedure matches what workers actually do.

A worked example: one month's admin in a 13-person provider

Consider a provider with ten support workers, two coordinators and a director, supporting about 25 participants. Illustrative figures follow. The director sorts the month's admin into lanes.

• Safe: updating the privacy policy and writing a short "how we use AI" paragraph for the website. Time saved: about 2 hours.
• Safe: drafting three job adverts for a recruitment push and an induction quiz. Time saved: about 3 hours.
• Careful: rebuilding the service agreement template. A second person checks that no example from a real participant remains. Time saved: about 2 hours.
• Careful: designing the shift-note headings. Tested on made-up notes first. Time saved: 1 hour.
• Closed: progress notes, incident reports and plan summaries remain with workers and coordinators, using the provider's own systems.

Total return: roughly eight hours a month, nothing involving participant data in a public tool. The director records each AI use in a register with tool, purpose, data involved, approver and date. That register is also the best evidence if the Commission or an auditor asks how you manage AI risks. If audit preparation is the pressure, building the audit evidence folder with AI shows where a register fits.

A privacy impact check you can run in an afternoon

Before any new AI feature or tool is used even in the careful lane, run a short assessment and keep the answers. A small provider can do it on one page, with the director, a coordinator and whoever looks after IT or data.

1. Purpose. What job is this for, and what would we do without it? If the answer is "it is just quicker", note how much quicker and ask whether it justifies any risk.
2. Data. What exact information goes in? Tick: none, de-identified, identifiable, sensitive. Anything beyond de-identified sends the job back to the closed lane or to a formal assessment.
3. Vendor. Where is data processed and stored? Is it used to train models? How long is it kept? Can we delete it? Can we switch off data sharing? Get the answers in writing from the vendor's own documents, not from a sales call.
4. People. Who may use it, who approves outputs and who is told when something goes wrong?
5. Participants. Do our notice and agreement cover this use? Does anyone need to be asked first?
6. Review. When do we look at this again? Put a date in the register.

Say a provider is thinking about an AI add-on in its rostering software. The assessment shows that it reads worker names and participant initials, runs in the vendor's overseas cloud and keeps prompts for 30 days. The director decides against switching it on until the vendor confirms where processing happens and offers a setting that keeps participant identifiers out. The decision, the date and the open questions go in the register. Nothing was lost, and the provider has a record of acting on the Privacy Commissioner's advice to assess before adopting.

Wording for participants and staff

Participants should be told, in plain words, how AI is used. Illustrative notice text for a privacy statement:

We sometimes use AI tools to help write policies, plain-language information and staff training material. We do not enter your personal information, your support notes or your plans into public AI tools. If we ever want to use an AI tool with your information, we will tell you first, explain what it does and ask for your agreement. You can ask us about this at any time.

Only publish a sentence you can keep. If your provider decides to use an assessed AI tool with participant data later, update the notice and the agreement first.

For staff, a one-page rule might say: use only the approved tools; never type a participant's name, details, diagnosis or notes into any AI tool unless the director has approved that tool for that purpose; de-identify with a second-person check; read and own everything you submit; report mistakes at once without blame. A mistake reported quickly can often be contained, and your privacy adviser can tell you whether any breach-notification duty applies.

Mistakes that show up in practice

• The helpful worker. A support worker pastes a day's notes into a free chat tool to tidy the writing. The notes contain a name and a medical detail. The fix is a clear rule, an approved alternative (a template, or a tidy-up done by hand) and a no-blame reporting route.
• Features switched on by default. A software update adds AI summaries to your case-management system. Read release notes, ask the vendor, and keep a register so a new feature is assessed before staff use it.
• The confident wrong policy. An AI-drafted policy cites a requirement that does not exist or is out of date. Check every legal or regulatory reference against the current source, and never publish a policy nobody has read line by line.
• Surface de-identification. A name was removed but the suburb, school and rare condition remained. The two-person test catches this.

A quarterly check that staff follow the lanes

Every quarter, sample ten uses from the register, ask the staff member what went in, and check that it matches the lane. Look for any new tool staff have started using. Ask a participant or advocate whether the AI notice makes sense to them. Review the Commission's and the Privacy Commissioner's pages for updates, because this guidance has been moving quickly. If you would rather draft the rule with someone, that is what my AI implementation consultation covers; get in touch.

Questions NDIS providers ask about AI and privacy

Is it illegal for an NDIS provider to use ChatGPT?

No. The Commission's February 2026 statement says current rules do not explicitly prohibit AI and that providers may use it if their use complies with their legal obligations. The risk is in what goes in: sending participant personal information to a public tool without safeguards can breach privacy duties.

Does de-identifying a participant's information make it safe?

It reduces the risk a great deal, and the Commission expects it for AI use in behaviour support plans. It only works if the information is really de-identified: names, dates, locations, unusual circumstances and rare combinations can still point to one person. Have someone other than the writer check before pasting.

Can I use AI to write progress notes?

Not by pasting identifiable notes into a public tool. Keep personal information out, or use a system your provider has assessed, contracted and configured. Support workers must still check that every note reflects what actually happened. Notes are records, and the accuracy duty stays with the person who signs.

Do I need participant consent to use AI?

Where AI will handle personal or sensitive information, tell participants in your privacy notice and service agreement, and check whether consent is needed for that use. The Privacy Commissioner's guidance says consent for sensitive information must be informed, current and specific. Ask a privacy adviser to confirm your position.

Further reads

• How to Write NDIS Progress Notes: An AI-Assisted Audit Checklist — A checklist for progress notes where AI assists but people write.
• How to Get Clients as an NDIS Provider: AI for Enquiries, Referrers — Handle enquiries and referrers without participant data.
• NDIS Occupational Therapy Report Template: A First Draft With AI — A report template first draft, with the privacy limits.
• AI Accounting for NDIS Providers: Categorising Income and Claims — Categorising income and claims.
• NDIS Software for Small Providers: Per-User Costs and What You Get — What provider software costs per user.
• AI Chatbot for Healthcare: A Small Clinic's Data and Consent Checklist — Consent and data questions for any patient or participant chatbot.
• NDIS Service Agreement Template: Plain-English Drafting With AI — Plain-English service agreements drafted with placeholders.
• AI for Speech Pathology in Australia: NDIS Reports and Scheduling — NDIS report drafting with clinician review, scheduling and reminders, and consent for children's data under the Australian Privacy Principles.
• NDIS Plan Manager Invoices: Checking Claims Against Budgets With AI — Field checks, category matching and a flag-and-hold list for provider invoices, with a person approving every payment and participant data kept tight.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NDIS Quality and Safeguards Commission position statement on AI in the development of behaviour support plans (February 2026); Office of the Australian Information Commissioner guidance on privacy and the use of commercially available AI products; checked 4 November 2026.