Ask, don't search: receipts in plain language

Watch someone hunt for a receipt in a traditional expense tool. Filter by date range. Filter by merchant. Filter by amount. Export. Ctrl-F. Five minutes later they give up and re-ask the restaurant.

The problem is that the tool's mental model — tables, filters, columns — has nothing to do with the user's mental model, which is a story: 'that taxi to the airport, sometime in March, around three hundred dollars.'

Semantic search inverts the interaction. You describe what you remember, in whatever language you remember it in, and the system does the translating. '機場的士' matches a receipt whose vendor is 'City Cabs' and whose note mentions Chek Lap Kok — because meaning, not keywords, is doing the matching.

The underrated benefit is for the accountant, not the employee. Questions that used to arrive as interruptions — 'can you find the Wework invoice from Q1?' — now get answered by the asker in seconds. The finance team's inbox measurably quietened.

Search was the last feature where the chat interface proved itself. If you can ask a colleague a question in a sentence, you should be able to ask your books the same way.