The 90-minute rule for AI tools

We have a rule on every AI tool we ship: if a human user can't get to a result they trust in under 90 seconds, we've built the wrong thing.

It sounds arbitrary. It isn't. Here's where the number comes from and why we've kept it.

The default failure mode

Most AI features fail the same way. The model is impressive, the demo looks great, and then the actual workflow looks like this:

1. User opens the tool, thinks for a second about what to type
2. User types a prompt that approximately describes what they want
3. Model produces an answer
4. User reads the answer carefully, because they don't fully trust it
5. User edits the answer, or re-prompts, or gives up

By step 5, the AI tool has cost more time than just doing the task by hand. The user remembers this. They stop using the tool. The feature dies.

Why 90 seconds

Ninety seconds is roughly the upper bound of how long a competent person will spend on something before their attention breaks. Below 90 seconds, the friction of using the tool is invisible. Above it, the tool starts feeling like work — and a tool that feels like work isn't replacing work, it's adding it.

We came at the number from a few angles. Empirically: tools that hit the rule get used daily, tools that miss it get used twice and abandoned. From research: 90 seconds is roughly the median task-switching cost in knowledge work. From the constraint itself: it forces design decisions that produce better products.

How the rule changes design

When you're trying to ship a tool that completes a real task in under 90 seconds, every part of the design changes.

You ruthlessly remove configuration. No "select a model." No "tune the temperature." No 200-word system prompt for the user to write. Defaults that work, options only when they unlock real value.

You preload context. The tool already knows who the user is, what they were just looking at, what they typed last week. It doesn't need to re-ask.

You constrain the output. Structured fields beat a paragraph the user has to read. A draft they can ship beats a "starting point" they have to rewrite. Three options to pick from beats one option they have to evaluate.

You design for trust, not power. The user needs to know, fast, that the output is trustworthy. That means visible reasoning, clear sources, confident defaults, and a one-click "this is wrong" feedback path.

What the rule rules out

Some things we've declined to build because we couldn't hit the rule:

• A general-purpose research assistant for a sales team. Couldn't hit the rule — the user always had to read and verify. Killed the project, rebuilt as a lead-enrichment automation that didn't need a UI at all.
• An AI document drafter for a legal-adjacent workflow. Couldn't hit the rule — review was always slower than rewriting from a template. Replaced with a template + smart-fill, no AI in the loop.
• A chat-based interface to a complex internal system. Couldn't hit the rule — even good answers required follow-up. Replaced with a structured form that the AI populated in the background.

Each of these would have been a successful demo and a failed product. The rule saved us from building them.

The non-AI version

There's a sneakier version of this rule: if a deterministic version of the tool can hit the 90-second rule, it usually beats the AI version on cost and reliability.

Most "AI dashboards" should be regular dashboards. Most "AI categorization" should be a rules engine with an AI fallback for the long tail. Most "AI search" should be regular search with semantic re-ranking for the top results.

We use AI where the problem is genuinely open-ended. The rule's real purpose is to keep us honest about when it is.