Technology design has spent decades treating friction as the enemy. Each extra click is a leak. Each pause is a conversion risk. Each unanswered question is an opportunity for abandonment.

AI intensifies this ambition. The ideal interaction appears to be immediate intention fulfillment: describe the desired state and watch the system produce it.

Yet some friction is the structure that keeps a person involved in their own decision.

Bad friction and productive friction

Bad friction wastes effort without improving understanding or control. Re-entering the same data, navigating hidden settings, waiting for an arbitrary process, or deciphering unclear language deserves removal.

Productive friction creates a moment in which the user can notice consequence, uncertainty, or responsibility.

The distinction depends on purpose. A confirmation screen before deleting a file can be protective; a confirmation before every harmless action becomes noise.

Friction for learning

When a student requests a complete solution, the interface could offer stages: identify the concept, reveal one hint, compare an attempt, then show a worked answer.

The user should remain able to access the answer. The design simply makes the learning path visible.

This is better than moralizing refusal because it preserves agency while offering a structure that supports practice.

Friction for consequential decisions

A system drafting a sensitive message could ask one question before producing final language: “What outcome do you want after the recipient reads this?”

A recommendation affecting another person could require the user to state what evidence might change the decision.

A memory feature could display which past information influenced a suggestion.

These pauses are not delays for their own sake. They expose the hidden dependency between input, model, and action.

Friction for consent

AI systems increasingly operate across email, files, calendars, and communication channels. Permission granted once can become invisible during later actions.

Productive friction restores context at the moment of consequence. Before sending, sharing, deleting, or publishing, the interface should show what data was used and where the result will go.

Consent is stronger when it is connected to a specific action rather than buried in initial setup.

Friction should be adjustable

A novice and an expert may need different levels of interruption. A low-stakes personal draft and a public announcement do not deserve the same review.

Good systems allow friction to scale with risk, reversibility, and user competence.

Possible controls include:

• “Ask me before using remembered personal information.”
• “Require review for external actions.”
• “Hide full solutions until I attempt the problem.”
• “Show uncertainty when sources disagree.”
• “Add a delay before irreversible actions.”

The user should understand why the friction exists and be able to change it where appropriate.

The business tension

Product teams measure speed, completion, and repeated use. Productive friction can make these metrics worse in the short term. A user who pauses may produce fewer actions.

But a system that removes all resistance can also create overreliance, accidental disclosure, shallow learning, and decisions no one can explain.

Trust is difficult to measure because it becomes visible mostly after it is broken.

Leave room for the user to arrive

The deepest purpose of friction is temporal. It allows the person’s judgment to catch up with the system’s capability.

AI can generate language, plans, and decisions faster than people can emotionally or morally inhabit them. The interface should not confuse technical completion with human readiness.

Some pauses are not obstacles on the way to the experience. They are the part of the experience in which the user remains present.

Friction should correspond to consequence

A system should not slow every action. The irritation would teach users to bypass safeguards. Friction belongs where speed can conceal a meaningful transition: publishing a claim, sending an emotionally charged message, deleting a long record, accepting medical or financial guidance, or allowing a model to act beyond the current screen.

The pause can be brief but specific. Instead of a generic “Are you sure?”, ask the user to inspect the recipients, source, irreversible effect, or assumption that matters. Useful friction returns attention to the decision.

Measure preserved agency

Product teams usually measure completion and retention. They can also measure whether users understand what happened, can reverse it, and can continue without the system. Does the interface show uncertainty? Can a person compare an assisted and unassisted route? Does the product make dependency visible before it becomes distressing?

These measures may sometimes conflict with engagement. That conflict is informative. A tool that serves agency should be willing to become quiet when it has done enough.

The deepest design challenge is not adding obstacles to a fast machine. It is deciding which human capacities the speed was supposed to serve.

Why I added an approval step after working so hard to remove clicks

Product design often rewards the shortest path. I spent time simplifying navigation, reducing duplicated controls, and making AI available inside the workspace where a decision was being made. Then I deliberately added a step between the assistant’s recommendation and the database.

The reason was trust. An AI could understand a request such as “move this to tomorrow,” but the exact meaning depended on context. Did “this” refer to the task, the focus block, or the milestone? Did tomorrow mean the due date or the planned execution date? A fast mutation could be technically correct and personally wrong. I changed the interaction so the assistant proposed a specific action, showed the fields that would change, and waited for approval.

That friction was productive because it appeared at the boundary of consequence. Elsewhere I removed friction aggressively. A simple greeting no longer needed the entire agent pipeline. Repeated context did not need to be re-entered. The user could stop generation without hunting for a control. The lesson was not that more friction is safer. It was that friction should move from routine effort to moments requiring judgment.

Aethel needed a similar redistribution. Readers should reach an essay easily, but publication should not be easy for me. Before a page becomes public, I should verify the sources, identify the personal experience behind the argument, inspect repeated phrasing, and accept responsibility for corrections. The reader receives a smooth page because the author encountered the resistance earlier.

A friction map for product decisions

I use a four-zone map. Green actions are reversible and low consequence; they should be immediate. Yellow actions change organization or planning; they deserve a visible preview and easy undo. Orange actions affect commitments, shared data, or long-term memory; they require explicit approval. Red actions are destructive, sensitive, or difficult to reverse; they need confirmation, clear scope, and sometimes an alternative path.

The map prevents two common mistakes. One is adding confirmation dialogs everywhere until users click through automatically. The other is treating conversational ease as consent. Productive friction must be rare enough to remain meaningful and specific enough to explain what is at stake.

There are contexts where even an approval step is too slow, such as emergency systems or accessibility features designed for immediate support. Those cases need pre-authorized rules, bounded automation, and reliable audit rather than constant interruption. The boundary is not the number of clicks. It is whether the user understood and governed the delegation.

The best friction I have designed does not feel like punishment. It feels like a moment when the system returns authorship: “Here is what I think you asked for. Here is what will change. You decide.” In an age of effortless generation, that sentence may be one of the most respectful things an interface can say.

Build a risk ladder instead of a universal warning

Generic confirmation dialogs become invisible through repetition. Productive friction should reveal information proportional to consequence. A private rewrite may need none. An external message should show recipients and final text. A financial or medical claim should expose sources and uncertainty. A destructive action should show scope, recovery, and what cannot be undone.

A simple ladder uses four levels: reversible private action, reversible external action, consequential recommendation, and consequential execution. Each level adds a different safeguard rather than merely another click.

This approach aligns with risk-management principles without turning ordinary use into bureaucracy. The user encounters friction where a new fact could still change the action.

Measure whether agency survived

Teams often evaluate friction by completion time and abandonment. Add three post-action questions: Could the user explain what happened? Could they identify the evidence or rule behind it? Could they reverse or challenge it?

A fast flow that fails these questions may have reduced effort by removing understanding. A slower flow that passes them may still be badly designed if the same understanding could be achieved more clearly. Productive friction is not delay for its own sake; it is information placed at the moment of responsibility.

NIST’s AI Risk Management Framework treats governance, measurement, and management as connected functions. At interface scale, that means a warning should correspond to a known risk and a real user action, not serve as decorative evidence that the company was cautious.

Let the user choose protected modes

People should be able to declare contexts where generation is limited: study mode, confidential rehearsal, source-only research, or human approval required. These modes transform a personal intention into a stable interface rule.

The best friction is often chosen in advance, before urgency and convenience dominate. It gives the future user a boundary their present self decided was worth protecting.

Give every product a friction budget

Friction can protect agency, but too much of it drives users toward shortcuts or trains them to dismiss warnings. Teams should budget it. Reserve the strongest interruptions for actions with irreversible or distributed consequences, and remove decorative confirmations from low-risk flows.

Review the budget with evidence: which warnings changed behavior, which merely delayed it, and which risks still reached users without a meaningful chance to respond? A confirmation that almost everyone clicks through is not automatically useless, but it should justify its place by revealing information or offering a real alternative.

This keeps safety work from becoming a collection of visible obstacles. The goal is a smaller number of better-timed moments in which the user can still choose differently.