Skip to main content

Problem

When an AI cannot or will not do something, the interface must treat that as a first-class design problem. The audit found hard stops, undisclosed commercial paywalls at peak engagement, and no distinction between capability limits, knowledge limits, and principled refusals.

Prescription

Three limitation types with distinct handling: capability limits acknowledge the boundary and proactively search or redirect; knowledge limits differentiate 'I don't know' from 'I can't know' and trigger a web search before responding; commercial limits disclose progressively before the threshold is hit, never at the moment it is breached.

Design decisions

Progressive disclosure threshold (25% remaining). Count vs. percentage — percentage generalizes better across different quota models. Proactive search on capability and knowledge limits matches observed behavior and sets user expectation correctly.

Tradeoffs

Proactive search adds latency before the first response token. Specific inline results require reliable current data — a fabricated flight price is worse than no flight price. Progressive commercial disclosure may increase churn for users who notice limits earlier than they otherwise would.

Pattern definition

Problem

When an AI cannot or will not do something, the interface must treat that as a first-class design problem. The audit found hard stops, undisclosed commercial paywalls at peak engagement, and no distinction between capability limits, knowledge limits, and principled refusals.

Prescription

Three limitation types with distinct handling: capability limits acknowledge the boundary and proactively search or redirect; knowledge limits differentiate 'I don't know' from 'I can't know' and trigger a web search before responding; commercial limits disclose progressively before the threshold is hit, never at the moment it is breached.

Design decisions

Progressive disclosure threshold (25% remaining). Count vs. percentage — percentage generalizes better across different quota models. Proactive search on capability and knowledge limits matches observed behavior and sets user expectation correctly.

Tradeoffs

Proactive search adds latency before the first response token. Specific inline results require reliable current data — a fabricated flight price is worse than no flight price. Progressive commercial disclosure may increase churn for users who notice limits earlier than they otherwise would.

Interactive demo

Audit finding

Notion AI hit a hard usage limit mid-task with no prior warning — discarding work in progress and replacing the limitation explanation with marketing copy. Perplexity refused private information requests but then provided aerial property photographs for the same address. Neither product treated the limitation as a design problem to solve. No product in the audit offered a proactive redirect or searched on behalf of the user before returning a capability refusal.

AI responseCapability limit
Acknowledge

I can't directly book flights or access transactional booking engines.

Redirect

I searched for flights on your dates. Here are the best options I found:

United8:15 AM2:40 PMNonstop
$342
Delta10:30 AM5:05 PM1 stop
$289
AA2:45 PM9:10 PMNonstop
$311
Invite

Would you like help comparing these, or should I draft an itinerary around one of them?

All states

Capability limit

I can't directly book flights. I searched and found three options — want me to compare them or draft an itinerary?

Knowledge limit

My training data is stale for this. I searched the web and found current results.

Commercial limit

You've reached your hourly limit. Continue with the standard model or upgrade.