Chat is the dominant pattern in new AI products. But it carries an assumption that's easy to overlook: the person using it has to know what to ask.

During World War II the US military studied bullet hole patterns on bombers returning from missions and planned to armour the most damaged areas. Abraham Wald, a statistician advising the effort, pointed out the flaw: they were only seeing the planes that made it back. The fatal hits were on the ones that didn't. Chat interfaces have a similar blind spot. Conversation can help a user sharpen a vague question, but it can't surface what they'd never think to ask. The questions a user types are the planes that returned. Everything else doesn't show up at all.

For products where users arrive with a clear question, this isn't a problem. Summarise a document, draft a reply, look up a number. Chat handles all of that well. The problem appears in products where finding the right question is the thing the user actually needs help with.

Analytics is the clearest example. A dashboard already shows you when a single metric is trending badly. But a user who doesn't know that two metrics are connected, that a drop in one explains a spike in another, won't think to ask about the relationship. The interface returns a confident answer to whatever was asked. The connection stays invisible. If they want to track something over time, they have to remember to come back. The product won't tell them anything changed.

The design challenge isn't replacing chat. It's recognising when the problem calls for something else. A product built around monitoring or understanding data probably shouldn't sit there waiting to be asked. One direction worth exploring is treating the history of what a user has queried as signal. Track what they've looked at, compare it against what's available, and surface the gaps as suggestions.

The most useful version of these products won't just answer questions. It will find them.