The NDIA has quietly deployed machine-learning tools to assist with drafting support plans for NDIS participants — a significant step in using algorithms within one of Australia’s major disability support programs. Documents recently obtained by media outlets show that the agency has used machine learning to generate draft budget recommendations, especially for first-time participants, based on key information in their profiles.
According to briefing papers from the agency, the machine-learning systems are used to produce what the documents call “Typical Support Packages” (TSPs) for initial NDIS participants. The algorithm analyses participant profile data — things like disability type, age, support needs categories and historical patterns — and then generates a recommended budget or support framework. However, the agency emphasises that final decisions remain firmly in the hands of human delegates: the machine-learning output is only advisory, and staff must review and approve any plan. The agency’s internal policy states that AI tools must not access participant records unless explicitly authorised.
Parallel to this, the NDIA also trialled Microsoft Copilot — a generative AI assistant — with roughly 300 staff over six months. That trial focused on staff productivity: drafting emails, summarising meetings, generating internal documents. Staff productivity reportedly improved (with a cited 20 % reduction in task time) and staff satisfaction was high (about 90 %). Notably, though, the trial did not involve Copilot being used to generate participant plans or budgets — that role remained restricted to the earlier machine-learning tool. The agency draws distinction between generative AI assistants and algorithmic budgeting tools.
The ambition behind using ML is clear: to speed up the plan-and-budget creation process for participants, streamline turnaround times, reduce administrative bottlenecks, and provide consistency for straightforward cases. For many participants, lengthy delays in plan development can lead to delayed supports or services. The algorithmic assistance is positioned as a way to alleviate that.
But disability advocates and AI ethics experts urge caution. They warn that machine-learning tools, while efficient, struggle with complexity, nuance and the unique individual circumstances that characterise many NDIS participants. One academic pointed out that “if a person doesn’t fit neatly into a box … a machine-learning approach can’t predict the supports they’ll need.” There is also concern about automation bias — the phenomenon where human decision-makers defer too much to algorithmic suggestions, rather than critically evaluating them. In a system where budgets can significantly impact a person’s supports for daily living, transport, therapy or equipment, the stakes are high.
Trusted safeguards become critical in this context. The NDIA asserts that no automation is used to determine eligibility or final funding decisions, and it stresses that staff remain responsible for each plan. Yet the documents show that the algorithmic outputs are still shaping the initial drafts. Experts stress that transparency, auditability and participant awareness of any algorithmic influence are essential. When systems assist with supports, participants have a right to understand how recommendations were generated and ensure their individual needs are not down-graded to historical averages or data patterns.
The use coincides with a federal push to adopt AI across government services, with a new whole-of-government AI plan supporting training, tool access and governance. However, some disability advocates say that systems imposed too quickly or without rigorous oversight risk repeating past failures of automated decision-making in welfare sectors.
For now, participants of the NDIS and their supporters may wish to ask whether an algorithm provided a draft for their plan, and ensure that human review was substantive. Organisations working with NDIS participants might also want to monitor how this machine-assisted drafting evolves and whether safeguards hold firm.
In conclusion: The NDIA’s use of machine‐learning to assist with initial drafts of participant support plans marks a tricky balance between efficiency and risk. While the potential to reduce delays and enable faster access to supports is promising, the reliance on algorithmic recommendation in a high‐stakes human services context demands careful oversight, transparency and constant attention to individualised care rather than one-size-fits-all models.
