The case for building this, including the parts that argue against us.
Two in five plan reassessments this year cut somebody's funding. The tribunal that hears the appeals has said in writing that it cannot keep up. The people assembling the evidence are frequently disabled themselves, and the research on scarcity says the load they are under measurably reduces their capacity to carry it.
What the system asks of people who have least to give
Before any argument about software, this is the situation the software would be entering.
Unpaid care is the largest workforce in the system, and it is not counted as work.
Around 2.8 million Australians provide informal care, including roughly 906,000 primary carers. Together they delivered close to 2.2 billion hours of care in a single year. Replacing it at formal-sector rates would cost just under $1.5 billion a week.Deloitte Access Economics, The value of informal care, for Carers Australia. 2020 data.
Almost half of primary carers are themselves disabled.
There are 1.2 million primary carers in Australia and 43.8 per cent of them have disability. The person assembling the evidence, chasing the reports and writing the reassessment request is very often managing their own condition while doing it.Australian Bureau of Statistics, Survey of Disability, Ageing and Carers, 2022.
Scarcity measurably reduces how well a person can think.
In a study published in Science, prompting people to think about their own money problems reduced performance on cognitive tests among poorer participants by around a standard deviation, while leaving wealthier participants unaffected. The same farmers scored lower before harvest, when poor, than after it. The authors are explicit that this is a load being carried rather than a permanent deficit: lift the load and the capacity returns.Mani A, Mullainathan S, Shafir E, Zhao J. Poverty impedes cognitive function. Science, 2013.
The burden has a name in the research literature, and three parts.
Learning costs, the effort of working out how the system operates. Compliance costs, the forms, the documents, the appointments. Psychological costs, the stress and stigma of the interaction itself. The framework's authors make a further point that matters here: the level of burden placed on a person is a political choice rather than an accident, and it falls hardest on those least equipped to absorb it.Moynihan D, Herd P, Harvey H. Administrative burden: learning, psychological, and compliance costs in citizen-state interactions. JPART, 2015.
What is happening to plans right now
These are the numbers behind the sentence people keep saying to us, which is that plans are coming back smaller.
Roughly two in five plan reassessments this year reduced someone's funding.
Around 40 per cent of plan reassessments in the first quarter of 2026, approximately 18,000 people, resulted in a funding reduction. This figure comes from press reporting rather than an NDIA publication, and we flag that because the distinction matters. It is the closest thing available to a measure of what participants are experiencing.Reported by the Australian Financial Review. Not published by the NDIA.
From October, community participation budgets reset by about half.
Budgets for social, civic and community participation are being reset from 1 October 2026, reduced by around 50 per cent and brought back toward 2023 levels, with capacity building daily activity allocations down around 10 per cent. It applies progressively as plans are reassessed or renewed, phasing in over roughly twelve months. Critical care and daily living supports are protected.Australian Government Department of Health, Disability and Ageing, NDIS legislation changes.
774,456 people have an approved plan, and the reassessment reaches all of them eventually.
As at 31 March 2026 the scheme had 774,456 participants with approved plans, growing by around 13,000 in the quarter. Every one of those plans comes up for reassessment, and from 1 October the reset applies at that point.NDIS Quarterly Report, Q3 2025 to 2026.
Eligibility tightens, and a large number of people are expected to leave the scheme.
Reporting on the reform legislation passed in August 2026 puts the number of participants expected to exit the scheme through tightened eligibility at around 160,000, alongside a $200 million Inclusive Communities Fund intended to partly replace what individual plans no longer cover. We include this figure with more caution than any other on this page. It comes from political reporting rather than an NDIA publication, the methodology behind it is not public, and it describes an expectation rather than a count. We have kept it off our participant-facing pages for that reason: it is alarming without telling any individual whether it applies to them.Press reporting on the NDIS reform legislation, August 2026. Not an NDIA figure, and not a count.
Requests to have a decision reviewed have been running near six per cent of participants.
Requests for review of a reviewable decision, as a share of active participants, rose from 4.8 per cent in the June 2023 quarter to 6.6 per cent in the December 2024 quarter, easing to 5.8 per cent by June 2025. On current participant numbers that is tens of thousands of people a year formally disputing a decision about their own support.NDIS quarterly reporting.
The tribunal that hears the appeals has said it cannot keep up.
The Administrative Review Tribunal received 7,375 NDIS applications between 14 October 2024 and 30 September 2025, and held 6,004 NDIS cases on hand at the end of that period. In its own words, the Tribunal does not have the resources to process this magnitude of lodgements, and the number of cases on hand has increased substantially.Administrative Review Tribunal, President's paper: the Administrative Review Tribunal, one year in.
The evidence that does not come in numbers
What a plan review actually costs a family.
We asked Wendy, a mother and carer, what preparing for a plan review was like. Her whole answer was: “Hell. One word. Hell.” It is on our homepage because it is the most accurate summary of the problem anyone has given us, and because a number cannot carry it.Bella Sláinte co-design programme.
The scale of what was already documented.
The Royal Commission into Violence, Abuse, Neglect and Exploitation of People with Disability held 32 public hearings and 1,785 private sessions, heard from close to 10,000 people, and reported across 12 volumes with 222 recommendations. The evidence about how this system treats people is not in dispute and has not been for years.Royal Commission final report, 2023.
Why we run co-design at all.
Peter's programme is sustained engagement with participants, support coordinators, allied health practitioners, planners, plan managers and small business operators. It exists because the failure mode of building software for a regulated care system is building something that satisfies the regulation and nobody else. Every design decision on this page traces back to something a person told us in that programme, or to a study we could check.Bella Sláinte co-design programme.
Whether a tool like ours actually helps
Drawn from a memo prepared for an allied health provider whose chief executive asked for the studies rather than a summary of them, because her therapists would want to check. The findings that cut against us are included, in full, for the same reason.
The gains go to the people still learning.
This is the most consistent result in the literature on AI assistance. Among 758 consultants in a randomised field experiment, the bottom half on a baseline assessment improved by 43 per cent and the top half by 17 per cent. Among 5,172 customer support agents across roughly three million conversations, the bottom skill quintile gained 36 per cent and the top quintile gained nothing statistically. The honest caveat: each group is measured against its own baseline, so part of the larger gain at the bottom is regression to the mean.Dell'Acqua et al., Organization Science 2026. Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 2025.
AI-drafted clinical documentation is worse than clinician-written documentation.
Eleven AI scribe products were compared against 18 human clinicians and scored by 30 blinded raters across ten domains. Human-written notes scored higher in every case, with AI worst on accuracy, thoroughness, usefulness, organisation and comprehensiveness. This is an argument for mandatory review rather than an argument against the tools.Reddy et al., Annals of Internal Medicine 2026.
When a tool produces the answer, the person's own skill degrades.
Shown in a randomised trial of students, and separately in experienced endoscopists, where adenoma detection in standard colonoscopy fell from 28.4 per cent to 22.4 per cent after clinicians became accustomed to AI assistance. The endoscopy study is observational and cannot establish cause. The mechanism matters more than either result: where the tool prompts or checks rather than answering, the harm disappears.Bastani et al., PNAS 2025. Budzyń et al., Lancet Gastroenterology and Hepatology 2025.
Training people to use AI critically does not protect them.
Forty-four physicians, every one of whom had completed twenty hours of AI literacy training, were randomised to error-free or deliberately flawed AI suggestions. Top-choice diagnostic accuracy fell by 18.3 percentage points when the suggestions were flawed. The defence has to sit in how the tool is built, not in an induction module.Qazi et al., NEJM AI 2026.
The claim that AI upskills practitioners is unproven.
No study has tested whether people perform better unassisted after using an AI tool. Where the question has been examined the finding is harm avoided rather than learning improved. This is why we removed the claim from this site that BellaAssist builds a durable skill in the person using it.Systematic review cited in our evidence memo, August 2026.
In a real workplace the gains are far smaller than in the trials.
Surveys of 25,000 workers across 7,000 workplaces linked to Danish administrative data found an average saving among adopters of 2.8 per cent of working hours, against the 15 to 50 per cent seen in controlled experiments. AI also created new work for 8.4 per cent of workers: checking outputs, review and compliance. In a clinical setting that new work lands on whoever signs a report they did not draft.Humlum and Vestergaard, NBER working paper 33777.
Six things the evidence decided for us
This is what working alongside the sector looks like in practice. Each finding above closed off a design choice we might otherwise have made.
Review is mandatory and cannot be turned off.
Because independently assessed AI documentation is worse than human documentation on every measured domain. BellaAssist submits nothing. A person approves everything that leaves it.
The tool checks and maps. It does not produce the answer.
Because deskilling appears where a tool answers for you and disappears where it prompts or checks. The Drafting Assistant flags language that will not land and offers a reframe. The practitioner writes the report.
Every suggestion carries the source it came from.
Because AI literacy training does not defend against automation bias, so the defence has to be in the design. A suggestion you can trace is a suggestion you can refuse.
Gaps are named in words, not buried in a score.
The readiness evaluation reports which of the 208 controls a submission satisfies and which it does not. A number on its own invites the reader to stop thinking, which is the behaviour the automation bias research warns about.
One handover pack, so the history is told once.
Because the load described at the top of this page is largely the load of repetition. A single version of the record, assembled automatically and owned by the participant, is the part of the burden software can actually remove.
BellaAssist does not touch behaviour support plans.
The NDIS Quality and Safeguards Commission stated in February 2026 that it does not endorse the use of AI tools in developing or reviewing behaviour support plans, and requires that no personal information of participants is disclosed to AI systems. We do not write, review or contribute to one. That boundary exists for this reason.
What nobody has told allied health practitioners
The most useful part of this page for anyone writing evidence for the NDIS, and the part no professional body has put in front of them.
Occupational Therapy Australia, the Australian Physiotherapy Association and the Australian Association of Social Workers have each published no AI position statement. There are CPD courses and magazine articles, but no policy.
The NDIA has published nothing on AI-assisted reports, assessments or evidence submitted to it. It neither accepts nor discourages. Practitioners generally assume a rule exists here. It does not.
No Australian allied health professional indemnity insurer has published a position on whether AI-assisted practice is covered. If you need to know whether you are insured, put the question to your insurer in writing and keep the answer, because there is no published one.
Australia has no binding high-risk AI classification. A mandatory guardrails framework that would probably have captured clinical AI was consulted on in 2024 and not enacted. The current national guidance is voluntary. Anyone who tells you health is classified high-risk in Australia is mistaken.
Speech Pathology Australia is the exception, with a genuine position statement published in May 2024. Ahpra and the National Boards have issued guidance interpreting the enforceable Codes of Conduct: practitioners must apply human judgment to any output of AI, and remain responsible for checking the accuracy of records created with it.
Check it rather than take our word for it
Every figure above is attributed to a named source with a date, and the academic findings carry the journal so they can be looked up. Where a study is weaker than its headline suggests, we have said so on the same line rather than in a footnote. Where a number comes from press reporting rather than the agency itself, we have said that too.
Three things on this page argue against using a tool like ours, and one of them made us delete a claim from this website. They are here because a sector being asked to change how it works is owed the whole picture, and because anything we left out, a good clinical governance lead would find anyway.
If you want to put a question to us about any figure on this page, or you think we have read a study wrongly, we would rather hear it: hello@bellaassist.au.