AivexaNewsSearch
AI news for builders and product teamsChecked every hour

When Does It Actually Make Sense to Put AI Into Products?

Collected Nov 11, 2026

Savannah Kunovsky, who leads a team of industrial product designers at the San Francisco design consultancy Ideo, told IEEE Spectrum that the industry's rush to add AI to physical devices often runs ahead of any clear reason to do so. In an interview with journalist Stephen Cousins, she described teams arriving at Ideo feeling pressured to demonstrate AI work, sometimes because a CEO has promised the board an "AI story" before an earnings call, or because a competitor has just shipped an AI-powered version of its product. The devices she cites as emblematic of the wave include smart-vision toothbrushes and mood-reading household robots.

Kunovsky's first move with such clients is to step back and ask who the product is for and what those people are trying to accomplish. Once that is settled, she said, the right role for AI — large, small or invisible — usually becomes obvious. She applies a single test to separate a helpful feature from a gimmick: does it help someone do what they are trying to do, judged in the real context of use? She also pushed back on the reflexive turn to large language models, noting that computer vision, sensor fusion and on-device models are part of a broader toolbox, and that choosing the right capability matters as much as deciding whether to use AI at all. Because AI is a new interaction paradigm that asks people to change their behavior, she argues it has to be simple — drawing a parallel to Ideo's predecessor firm giving Apple's first mouse, in the early 1980s, a single button so that people learning an unfamiliar system faced an effortless mechanism.

Her clearest examples of AI degrading a product are cases where it displaced something that already worked. She called Humane's screenless AI Pin, which substitutes an assistant for apps and a display, a worthwhile provocation whose technology was not ready: it asked people to abandon a familiar interface for one that was slower and less reliable. She pointed to everyday versions of the same pattern — a voice assistant that mishears a command and turns off the wrong light, or a customer service bot that traps users in a loop when they want a human. In cars, moving climate and volume controls onto touchscreens and into voice menus means actions once performed by feel now pull a driver's eyes off the road; she noted that Euro NCAP, Europe's car safety rating program, now rewards physical controls for that reason.

An internal client example makes the trade-off concrete. A company wanted a cooktop with no knobs, where the user tells an AI what is cooking and it manages the heat. In testing, the pan began to smoke, the user reached for a knob that was not there, said "turn it down" twice, and pulled the pan off the burner before the AI responded. The knobs stayed. Kunovsky's rule of thumb: anything done instantly, by feel or under stress belongs on a physical control — a volume knob, a light switch, a stove burner, an emergency stop — because a mechanical control is faster and more trustworthy than a system that must interpret intent. This is a useful heuristic for hardware teams: latency and failure modes that would be tolerable in a chat window become unacceptable when the user's hand is already moving.

She was careful not to frame the choice as analog versus intelligent. Analog objects such as film cameras, good kitchen knives and notebooks carry value precisely because of craft and a feeling of full control, but a smart home can offer a voice assistant, an app and a wall switch for the same light, and redundant paths are fine as long as they do not add complexity for its own sake. Good integration, in her account, starts from what people need, then considers technical feasibility and business sense, and only then assigns AI a role: sometimes the core of the experience, as with a natural voice interface, and sometimes invisible background learning of preferences. She cautioned that because LLMs operate on text, the default instinct is to bolt a chat window onto everything, when physical products can communicate through light, sound, vibration, movement and the feel of a control.

The applications she finds most convincing extend what people can perceive, understand or do. Hearing aids can use AI to isolate the voice a wearer wants in a noisy restaurant; apps such as Be My Eyes let blind and low-vision users point a phone at the world and hear a description; live translation in earbuds lets two people without a shared language converse. She described the design target for proactive systems as the attentiveness of a good restaurant server who refills a glass without intruding, and named a coming wave of personal agents — Instinct, Dots and Muse — that work in the background and return with findings. The governing variables there are relevance and timing: how loudly a product speaks should match the stakes, and repeated dismissals should prompt adjustment. That implies telemetry and feedback loops are product requirements, not analytics afterthoughts.

On testing, Kunovsky said generative AI changes what a product definition even is. Traditional software is specified and then verified against that spec; with nondeterministic models you instead define the space the product should operate in. Evaluations, which describe good performance across thousands of situations, and constitutions, written principles and boundaries for the model, together with model software, configuration and runtime infrastructure, form the container the product lives in — work she places in design rather than as an engineering step at the end. Trust, she said, suffers when outputs vary, since people learn to rely on products that behave identically every time. Teams must therefore decide in advance where the product should be consistent and where it may explore, signal to users which mode they are in, and show what it is doing when uncertain.

Override paths she treats as essential, not merely an emergency stop for a home robot or a car. Undoing an action, editing a suggestion or switching a feature off all count, and they must be obvious, fast and available at the moment of need rather than buried in a settings menu. She cited Ethiqly, an AI writing tool developed in her work that helps students organize ideas rather than write essays for them, where teachers approve every piece of feedback before a student sees it. In her framing, easy intervention turns a mistake into a small moment; its absence turns the same mistake into a reason to abandon the product.

Why it matters: Product and engineering teams building AI into hardware or software get a practical checklist — start from the user's task, pick the model type that fits, keep physical or immediate controls where speed and reliability matter, define behavior through evaluations and constitutions, and make reversal one tap away. My inference is that this favors teams able to instrument real-world usage and iterate on timing, since the failures described are interaction failures rather than model-quality failures, and that shipping an AI feature purely for an earnings narrative carries reputational risk that outlasts the launch.

Read at IEEE Spectrum · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Over the past few years, the tech industry has rushed to plug a new microchip into every device on the assumption that if you can add AI, you should. Smart-vision toothbrushes , mood-reading household robots , and a host of physical products are invading consumers’ daily lives, reliant on artificial intelligence to learn, predict, and adjust our behavior. But as anyone who’s struggled to update the firmware in their fridge knows, AI doesn’t improve every product. Some product designers are asking where the line is