What makes a request “unclear” — and why a person should decide
How AskLight separates clear requests, unclear speech and things that aren't requests at all — and why unclear speech goes to a supervisor.
AskLight team · · 2 min read
When a resident speaks to the AskLight device, the AI has one job: work out whether they’re asking for something, and if so, what. There are three possible answers — and the middle one is the most important.
Clear requests
“Could someone bring me water?” is clear. So is “I need the bathroom, and then help getting back to bed” — that’s two requests, and AskLight creates two tasks so each can be claimed.
Indirect requests can be clear too. “It’s really cold in here. Could I get another blanket?” is a comfort request even though it starts with a remark about the temperature.
Not requests
Plenty of speech in a room isn’t a request: the television, a visitor chatting, a resident saying “I’m fine, thank you,” or describing something that happened yesterday. AskLight creates nothing for these, so staff aren’t interrupted.
Negation matters. “No water, but please help me stand” contains one request — help standing — not two.
Unclear
Then there’s everything in between. “My pillow fell on the floor.” Is that a request to pick it up, or just a remark? “Help.” Very short, little context.
A system built to look impressive would guess. AskLight doesn’t. Unclear speech goes to the supervisor’s review list, where a person can approve it as a task, edit it, or dismiss it — and that decision is recorded.
Why not just guess?
Because the cost of a wrong guess is real either way. A guess that creates a task from TV audio interrupts staff and wears down trust in the alerts. A guess that ignores a real but awkwardly worded request leaves a resident waiting.
Sending uncertainty to a person keeps the alerts staff receive meaningful, and keeps residents’ awkward or quiet requests from disappearing.
Measured honestly
It would be easy to make the numbers look good by sending everything to review. So we count “went to review” separately from “became a task”, and measure misses and false alerts on their own. We’d rather show you real results on your unit in a pilot than quote a single accuracy figure.