The input you do not ask for
IFA 2026 was full of appliances that work out what you want instead of asking. The design problem that creates is not accuracy, it is what happens when the guess is wrong.
IFA finished in Berlin this week, and the appliance halls had a theme running through them that was easy to miss under the robot demos. Miele showed the G8, its first dishwasher with a camera inside the door. Shut it and a light flicks on, the camera takes one frame, the light goes out, and the machine decides what it is looking at. A bottle gets a harder spray. A pot gets more intensity. Plastics get more heat on the dry. Miele put its own framing on the stand: fewer decisions, less to think about. It reaches Germany, Austria and Switzerland in April 2027.
A few halls over, Anker showed the SleepLab Pro, a bedside speaker at $269.99 with 60GHz radar in the top of it. It reads the micro-movements of your chest to get heart rate and breathing, and stages your sleep from that, with nothing on your wrist and no subscription attached.
Two very different products, the same move. Both of them removed a question they used to ask.
The question is the expensive part
We build a lot of things that ask people questions. Check-in flows, booking forms, intake, kiosks at a door with a queue behind them. The pattern is always the same: every question you ask costs you people. Some of them stop, some of them guess, some of them put in something wrong because the honest answer did not fit the options.
So a sensor that answers the question for you looks like a straight win. The dishwasher already knows what is in it. The radar already knows you are asleep. Nobody had to be asked, nobody had to be right, and the flow got shorter by one step. That is the same instinct behind the on-device OCR in Scan My Receipt: the amount is on the receipt already, so do not make somebody type it in.
Where it actually gets hard
The failure mode changes, and this is the bit that gets underestimated in the specification. When you ask a question and the answer is wrong, it is wrong in a way both sides can see. There is a field, it has a value in it, and the value can be corrected. When you infer the answer, the mistake is silent. The machine has already acted on it. The dishwasher has already picked its cycle by the time you would have noticed it read a pot as a plate.
That gives you three things to design that a form never needed:
- 01Visibility. The person has to be able to see what was inferred, in plain language, before it matters. A guess that is never surfaced cannot be corrected.
- 02A correction path that is cheaper than the original question. If overriding the guess takes longer than answering would have, you have added work and called it automation.
- 03A confidence floor. Below some threshold the honest move is to fall back and ask. Silent low-confidence guessing is how you lose trust in the whole system, not just the one field.
The confidence floor is the one that gets cut for time, and it is the one that matters most in a room with people waiting. On a kiosk at a busy door, a wrong inference that quietly proceeds is worse than a short question, because it puts a member of staff into an exception path with an audience watching. We have written before that the real test of a check-in system is how fast somebody can sort out the arrival that goes wrong. Inferred input raises the stakes on exactly that path.
The studio read
The direction is right and it is not a fad. Gartner expects 40 per cent of software vendors to have AI built specifically for PC deployment by the end of 2026, up from 2 per cent in 2024, and IDC has NPU-equipped machines at 94 per cent of new PC shipments by 2028. The silicon to run this locally is arriving whether or not anyone asked for it, and Anker shipping a radar sensor with no subscription is a fair signal of where the cost floor now sits.
What we would take from Berlin is narrower than the keynotes. Do not treat inference as a way to shorten a flow on a slide. Treat it as a trade: you remove a question, and in exchange you take on the job of showing your working and being correctable. Ship the correction path in the same release as the sensor. If that path is not funded, keep the question. A form that asks is dull, but it is honest about what it does not know.