What should AI actually do with an NDIS case note?

It should read what the worker wrote and ask them about it before the note is saved — not tidy the prose, not summarise the shift, and not decide whether anything happened. The useful job is narrow: catch the moment where a note describes something the same note says didn't happen, while the person who was there is still looking at the screen.

Everything else marketed as AI for case notes is either rewriting, which loses information, or reporting after the fact, which is too late to ask anyone anything.

The pain

A support worker finishes a shift, opens the note, and writes what happened. Then they hit the incident question and select "none observed", because nothing felt like an *incident* — it felt like a Tuesday.

Except the note says she was found on the floor. Or that he was more agitated than usual and it took two of them to settle him. Or that the medication was refused twice.

Nobody is hiding anything. The note and the checkbox are answered by two different parts of a tired brain, and the checkbox is answered by the part that is comparing today to every other day. The contradiction is visible to anyone reading the note as a whole — which happens weeks later, at a file review, when the person who could have explained it has worked eleven shifts since.

What ours does

Two layers run on the worker's own words.

Layer one is deterministic and offline. A keyword match that understands negation, so "no bruising observed" does not trip the same wire as "bruising observed". No model, no network, no latency — it runs regardless.

Layer two is a model reading the note in context, catching the phrasings nobody thought to put on a keyword list. Between them they cover the two failure modes: the wording you anticipated, and the wording you didn't.

If the note describes something that contradicts the "none observed" answer, the worker is asked about it there and then, on the form, before anything saves.

It screens the raw text, before the insert. That ordering is the whole feature, and we learned it the hard way — the screen used to run *after* the note row was written, on the AI-generated paraphrase rather than on what the worker actually typed. Two consequences, both real. It could not gate anything, because "none observed" was already stored by the time it answered. And it was reading laundered text: the generator rewrites a worker's words into progress-note register, and whatever the rewrite smoothed away, the screen could not see.

So it was moved. Same model, different position in the sequence, completely different feature.

What it deliberately doesn't do

It doesn't block the note. If the model is unreachable — dead spot, edge function down, flat battery on a long drive — the screen reports that it was unavailable and does not flag. Layer one still runs and still fires. A worker who cannot record what happened because a network call failed is a worse outcome than a missed flag, and that is a deliberate trade, written into the code as fail-open.

It doesn't decide whether something is an incident. It asks the worker. The answer is theirs, it goes in the record as theirs, and the classification decision — reportable, notifiable, none of the above — sits with the provider, where the legal responsibility already sits.

It doesn't rewrite the observation. The screen reads; it does not edit. The words in the record are the worker's.

And it doesn't catch everything. We have no detection rate to quote and we are not going to invent one. Two layers over the raw text catch more than one layer over a paraphrase, which is the honest claim and the reason the change was worth making. Your review still matters.

So is this "AI-powered compliance"?

No, and we would rather say so than sell you the phrase.

Compliance is a set of obligations that belong to your organisation. Software cannot hold them for you, and any vendor implying otherwise is selling you a feeling. What this does is narrower and more useful: it puts a question in front of the one person who knows the answer, at the only moment they can still answer it easily.

That is worth having. It is not the same as being compliant, and a tool that blurred the difference would be doing you harm at exactly the point where you were trusting it most.