How do you show uncertainty in a voice interface with no screen?

Arctic Fields Refrigeration services industrial chillers for food-processing plants. Rui Tanaka is a field technician. Wrenchmate is the voice assistant built into his headset, reading live sensor data and answering spoken questions while both his hands stay on the equipment.

How to answer this, stage by stage

The interviewer already knows a screen can show a number quietly. What they want to know is whether you'll notice that a voice has to say its own doubt out loud, and that saying it too often is its own kind of failure.

Let's learn

Picture doing your job with both hands inside a running machine, and the only screen nearby is one you can't look at anyway.

Wrenchmate reads live sensor data off industrial chillers and answers a technician's spoken questions, like "what's the suction pressure on unit four." Before it existed, a technician either read a printed gauge by hand or radioed a base tech for a second opinion, taking about 5 minutes per check. With Wrenchmate, most answers land in about 30 seconds.

A screen can hedge by just looking a little uncertain. A voice has no "a little." It either says the doubt, or it doesn't.

Here's the turn: when Wrenchmate first launched, it read back every single measurement out loud for confirmation, a safety-first choice made when usage was rare. As technicians started using it dozens of times a shift for routine checks, the team cut that automatic read-back to speed things up, for every kind of query, not just the routine ones.

At its worst, an unqualified voice answer near a live safety action doesn't just waste a technician's time. It can send them to open a line believing it's been checked, when it hasn't.

What I would leave alone: routine, low-stakes queries, like checking a temperature reading nowhere near a live action, don't need any of this. Adding a confirm loop there would just slow down dozens of harmless checks a shift for no real safety gain.

The lesson: a voice interface doesn't get to stay quiet about its own doubt the way a screen can. It has to actually say something, and saying the same cautious thing every single time teaches people to stop hearing it at all.

Now here is the same thing as a story

The short version above is what you'd say out loud in an interview. Read this one for how close the actual near miss came.

His name is Rui Tanaka. He's serviced industrial refrigeration units for nine years, mostly by feel and by ear, long before Wrenchmate existed. He could tell a failing compressor by the pitch of its hum before any sensor confirmed it.

The good months with Wrenchmate were loud, busy ones. By mid-morning most days, Rui had cleared four or five chiller checks with a quick spoken question and a quick spoken answer, hands never leaving the machine, headset doing the reading for him.

On a Tuesday, three separate chillers had already needed attention by 10 a.m., and the compressor room was running at full volume. Rui asked, "What's the suction pressure on unit four?" reaching for the service valve as he spoke, expecting the usual quick, plain number.

Rui's shift ran on the left panel, dozens of times a day. The right panel only had to happen once.

Wrenchmate answered, flat and certain: "Reading normal, 42 psi." It had heard "unit fourteen," a chiller three rooms over that had already been serviced that morning, not the live unit Rui's hand was on.

His hand found the physical gauge out of ninety habit, not because Wrenchmate had asked him to check. The physical reading didn't match the number he'd just heard. He stopped, re-asked the question slower, and this time Wrenchmate correctly heard "unit four" and gave the real, still-pressurized reading. Nothing happened to him. But the near miss reached his supervisor's desk within the hour.

The middle steps used to be skipped for every query. Now they only fire when it actually matters.

The redesign didn't bring back the old always-confirm habit. It made the confirm step conditional: any query naming a unit right before a safety action, like opening a valve, now triggers an automatic read-back, whether or not Wrenchmate feels confident. Everything else stays as fast as it was.

One question decides whether the extra half-second happens at all.

PICK, spelled outNot a lecture on speech recognition. PICK is what decides how often the caution is worth the interruption.

None of these need a screen. All four fit inside a sentence someone can hear while their hands stay busy.

The recap, one line per letter: position is stakes-gated confirmation, not always-on or always-off, impact is seconds for Rui versus a real safety risk, cost asymmetry is 4 minutes against 180, and kill criteria is the moment a passive confirm stops meaning anything.

And if you want to be sure it really works, try it somewhere elseSame four letters, a grocery fulfillment warehouse instead of a refrigeration plant. This time the voice assistant is picking orders, not reading pressure gauges.

Northfield Grocery Fulfillment uses a voice-pick headset system so a picker named Grace Okonkwo can hear which item to grab next and confirm the count out loud, hands free to actually carry the bins. Mapped onto PICK: position is confirming out loud only when an item is one of a small set that's easy to misgrab, like two nearly identical spice jars sold in different sizes, and answering straight everywhere else. Impact is Grace losing a couple of seconds on a routine pick versus a wrong item reaching a customer's order and triggering a refund and a re-pick. Cost asymmetry: a routine confirm costs about 2 seconds; a wrong-item shipment costs roughly 25 minutes once you count the refund, the re-pick, and the apology email. Kill criteria: if pickers start confirming by habit without actually checking the label, the system needs to require reading the item's exact printed size back, not just saying yes.

Same four parts as Wrenchmate's spoken answer. A pressure reading and a spice jar size turn out to need the same shape of honesty.

The old decision at Northfield wasn't about read-backs at all, it was a merged-steps problem of its own kind: the system used to announce the item name and the expected count in one breath, with no pause for the picker to actually look at the shelf before confirming. That was fine when the catalog had few lookalike items. It stopped being fine once two nearly identical spice jars, in different sizes, sat on the same shelf.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "confirm only where a mistake would actually cost something, answer plainly everywhere else," and stop.
Cost: there's no budget to add confidence scoring to the voice model right now. Say so, and start with the cheap fix, a hard-coded list of high-stakes actions that always trigger a confirm, no machine confidence needed at all.
The model gets better, for real: if Wrenchmate's speech recognition improves and the mishear rate drops well under the 5% concern line even in loud rooms, the confirm rule can loosen, the same threshold running in reverse.

Where people run it wrong.
They add a confirm step to every single voice answer, and the technician stops actually listening to it.
They remove all confirmation to speed things up, without separating routine queries from safety-tied ones.
They use a spoken percentage as their hedge, which takes longer to process than the task in front of someone's hands.

How to use it live. When someone asks how you'd show uncertainty with no screen, ask back: which of these answers, if wrong, would actually cost someone something? Let that list, not a blanket rule, decide where the voice slows down.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0