Match effort to stakes

AI products fail in both directions on stakes. They make trivial things heavy — a confirmation dialog to fix a typo — and consequential things weightless, one breezy chat bubble away from sending the wrong email to the wrong list. The principle is proportionality: the interaction cost a person pays, and the compute and latency the system spends, should track what the task is actually worth.

The clearest expression is explicit modes. ChatGPT lets you pick between fast models and slow reasoning-heavy ones; Perplexity separates quick search from deep research. This is stakes-matching made into a control: the person declares “this one matters,” and the system spends minutes instead of seconds. The design win isn’t the toggle itself — it’s that latency becomes a choice rather than an imposition. People will happily wait ten minutes for work they’d have spent two hours on, and they’ll resent ten seconds for a synonym.

Input should scale with stakes too. Freeform text is perfect for low-stakes exploration and miserable for tasks with hard requirements — nobody should specify a legal jurisdiction or a budget cap in prose and hope the model catches it. Structured inputs, constraints, and tuners let people pin down what must be true and leave the model freedom everywhere else. Midjourney’s parameters and Figma’s property-driven AI actions both understand this: the more the output matters, the more the input deserves affordances beyond a text box.

Output containers carry stakes as well. A conversation is the right container for a throwaway answer; it is the wrong container for anything a person will iterate on, because revisions bury each other in scrollback. Claude’s Artifacts and ChatGPT’s Canvas exist precisely to promote high-stakes work out of the chat stream into a persistent stage where versions accumulate instead of scrolling away. Choosing the container is choosing how much the work is allowed to matter.

The discipline this demands from teams is restraint: resist the flat default. One model, one mode, one ceremony for everything is easier to ship and worse for everyone — power users can’t go deep and casual users can’t go fast. Map your tasks by stakes before you design the interaction, and let the cheap things be cheap. The saved attention is what people will gladly spend on the moments that deserve it.

The best argument against gradients is that they export your routing problem to the user. Every mode picker is a question the person must answer before doing their job, and most people answer it badly — they can’t predict whether their task needs the deep model, so they either default to fast and get burned or default to deep and pay latency for nothing. Apple’s whole design tradition argues the opposite of this principle: one great path, no decisions. And modes multiply surface area — every toggle is documentation, QA, and a support ticket. This critique is why the field is moving toward automatic routing, and where routing genuinely works, it should win. But routers misjudge stakes precisely because stakes live in the user’s head, not the prompt’s text — “draft a note to my landlord” might be trivial or might be a legal dispute. The durable answer is a smart default with a visible override: route automatically, but let the person declare “this one matters” in one gesture. Removing the gradient entirely doesn’t remove the decision; it just makes it silently, and wrongly.

Measure the routing, in both directions. Mode-switch regret is the sharpest signal: how often does someone rerun the same request in a heavier mode after a fast-mode miss (under-spend), and how often does a deep-mode run produce an answer the fast mode matches in eval replays (over-spend)? A healthy product shows escalation-rerun rates in the low single digits and a shrinking share of deep-mode invocations that were provably unnecessary. Watch container promotion too: what fraction of multi-turn iterative work gets moved into an artifact or canvas versus dying in scrollback — rising promotion on long-lived work means stakes are finding their right home. And for structured input, track requirement violations: outputs that missed a hard constraint the user had stated in prose are evidence the input surface is under-built for the stakes. Good looks like latency complaints and “it wasted my time on something trivial” complaints both trending down at once — the signature of proportionality, because a flat product can only fix one at the other’s expense.