Reading a receipt is not recording an expense.

A receipt photo reaches the team. AI recognises the amount, date and supplier. That is a useful part of the work, but it does not prove that the expense is correct, belongs to the selected project or has not been recorded before.

Three steps with different responsibilities

Trust depends on what happens next

An assistant that says “done” after preparing the data creates a false expectation. The team needs to distinguish “I read the document”, “I propose this record” and “the system confirmed it was saved”.

The confirmation boundary matters particularly when an action changes costs, sends a message or updates shared information. This is a process design decision as well as a technical setting.

AI can prepare the proposal. The record needs a verifiable confirmation.

The questions we would ask

• Can we trace the record back to the source document?
• What happens when the same receipt arrives as both a photo and a PDF?
• Which fields need review when extraction is unclear?
• How does the team see that the destination system failed?

Starting with one document type and one defined action makes the workflow easier to check. Expansion makes sense once the exceptions are understood.