Run your NDIS business on the software we run ours on.

AI for NDIS rostering and compliance work — what it should and shouldn't do

Most "AI for NDIS" is a chatbot bolted onto a form. Ours does the unglamorous jobs: it reads what your support workers actually type while they are still typing it, it checks its own output before you see it, and where it wants to change something it asks first.

Case notes are screened before they save. Two passes run on the worker's own text — a deterministic keyword check that understands negation, and a model that reads the note in context and catches the phrasings nobody thought to put on a list. If a note describes something that looks like an incident while the incident question says "none observed", the worker is asked about it there and then, not at the next file review.

It reads what the worker wrote, not a tidied-up summary of it. That ordering matters: a paraphrase drops exactly the detail a screen is looking for.

And it fails open, on purpose. Dead spot, flat battery, edge function down — the note still saves and the keyword pass still runs. Blocking a worker from recording what happened because a network call failed would be a worse outcome than a missed flag.

It checks its own drafts. The care plan generator is asked to quote a participant's documented words where the source allows — and then a deterministic check looks for every quoted span, verbatim, in the documents actually supplied. Anything it cannot find is demoted to plain text and marked for you to verify. Marking only: nothing is rewritten, nothing is deleted, nothing passes silently.

It hands back what it cannot do. A worker photographing a receipt gets the details read off it and filled in. When the image cannot be read, they are told so and asked to enter it manually — no silent partial fill, no confident number pulled out of a blurred total.

And it is switched off where consistency matters more than novelty. Worker training runs from fixed, authored curriculum, served the same way to every worker every time. Compliance training has to be the same thing twice; a module that regenerates is a module that can quietly change what your workers were taught.

What it does not do: act on its own. Where the system can propose a fix, it doesn't apply it — it queues it. The proposal sits as pending until a person approves or rejects it, and only an approved action can execute. The executor checks that approval before it runs anything, routes through safe handlers, and never writes to financial records directly.

That last constraint is the one worth knowing: the category of record where a wrong automated write costs the most is the category it is structurally not allowed to touch. In this sector the accountable decision belongs to a person, and software that quietly takes it isn't saving you work — it's moving your risk somewhere you can't see it.

What should AI actually do with an NDIS case note?