Asking With Gloves On
The device assistant we built for the plant hall was used far less than expected, and the reason had nothing to do with its answers. People were putting a tool down, removing a glove, and deciding the question was not worth it.
Why It Was Not Being Used
We had measured answer quality, latency and coverage, and all three were acceptable. Usage was about a third of what the wired terminals had suggested it would be, and nobody had complained, which is the worst kind of feedback.
Watching a shift explained it in twenty minutes. Asking a question meant setting down a tool, removing a glove, waking a device and typing a part number with cold fingers. The threshold at which a question is worth asking is much higher than any of us had assumed from a desk.
What Voice Changes and What It Breaks
It removes the physical cost almost entirely, and usage roughly tripled within a month of the voice interface arriving. That part was straightforward and is not why this article exists.
It introduces one failure the typed version could not have. A misheard part number is a valid part number for a different component, and the assistant then answers correctly about the wrong thing, fluently, with a source. Typing produces a not found; speaking produces a confident answer.
The Research on Speech and Tools
Pahwa and colleagues published work in 2026 benchmarking speech-driven tool use, examining how spoken queries turn into tool calls across domains and acoustic conditions rather than treating transcription and action as separate problems.
That coupling is the point for us. Our failure is not a transcription failure in isolation, it is that the transcription feeds an identifier lookup with no tolerance for being slightly wrong, and the two have to be evaluated together.
| Spoken input | What we do |
|---|---|
| A part or order identifier | Read it back, wait for confirmation |
| A general question | Answer directly. No confirmation |
| Anything triggering an action | Confirm, always, on screen and aloud |
| Unrecognised in noise | Say so and offer the keypad |
The Read-Back
Identifiers are read back before anything is looked up: you said part four seven one one, is that right. It costs three seconds, it is what aviation and logistics have done for decades, and it removed the class almost entirely.
General questions are not read back, because confirming every utterance destroys the benefit that made voice worth building. The distinction is whether a mishearing produces a wrong answer or no answer, and only identifiers have the first property.
Spoken and Shown
Every answer is both spoken and displayed. The spoken version is short, the displayed version carries the source and the full value, and a technician who wants to check reads the screen when they have a hand free.
That combination came from watching people rather than from a design decision. Nobody trusted a spoken specification value enough to act on it without looking, and an interface that only speaks would have been used for orientation and not for the work.
The Acoustics Nobody Budgets For
A plant hall at shift change is loud in a way that changes recognition accuracy substantially, and our first evaluation had been done in an office. Recognition of ordinary words held up; recognition of alphanumeric identifiers did not, which is the exact reverse of what would have been convenient.
We spent a week collecting audio in the hall at three times of day and used it to choose settings and to set the confidence threshold below which the system asks for the keypad instead. That week was worth more than any model choice we made.
What We Did Not Build
Voice control of anything with an effect. The assistant answers questions and does not update records, book parts or close work orders by voice, even though the same customer asked for all three.
The reasoning is the read-back argument taken to its conclusion. A confirmation is a reasonable cost for a lookup and a poor guarantee for an action taken in a noisy hall by someone holding a tool, and we would rather that step happen at a terminal.
What We Do Not Claim
We do not claim voice is right for every assistant. It is right where hands are busy and eyes are on something else, and in an office it is usually worse than typing for everyone except people who cannot type comfortably.
We also do not claim the read-back removes mishearing. It removes acting on a mishearing that the user notices. A user who confirms without listening is back where they started, and we do not know how often that happens.