An AI feature bolted onto a financial product has an obvious failure mode: it sounds confident about something it got wrong, and the person reading the answer has no easy way to tell the difference between a well-sourced fact and a plausible-sounding guess. For most software that is an annoyance. For a treasury or finance team relying on an answer to reconcile activity or explain an outcome, it is disqualifying.
Copilot was built around one rule meant to make that failure mode structurally impossible rather than merely unlikely: it may never state a number without evidence chips linking to the records it came from. If it cannot source something, it says so, instead of guessing.
What an evidence chip actually is
Every figure Copilot surfaces, a payment total, a stuck payment, an outstanding invoice, ships attached to a chip that links directly to the record it was pulled from. Not a citation in the abstract sense, a link to the actual payment, invoice or ledger entry sitting behind the number, the same record you could pull up yourself and check.
That is a deliberate constraint on how the feature is allowed to answer, not a nice-to-have layered on top afterward. An answer without a populated evidence chip is not a degraded answer Copilot is allowed to give anyway. It is not a valid answer at all.
Why "I don’t know" is a feature, not a limitation
The instinct with most AI products is to make the assistant sound helpful in every situation, which in practice often means trained to always produce something rather than nothing. Copilot is built the other way. When it cannot source an answer, it says so rather than filling the gap with something that sounds right.
That matters more in finance than almost anywhere else software gets applied. A wrong answer stated with confidence is worse than no answer, because it can get treated as fact and propagate into a decision before anyone thinks to double check it. A clearly flagged "I don’t have evidence for that" costs a follow-up question. A confident wrong number can cost a lot more.
What this looks like day to day
Ask Copilot why a payout is still pending and it will point at the specific record showing the hold, not describe pending payouts in general. Ask it which invoices are still outstanding and it will list the invoices themselves, not a figure you have to trust. The evidence chip is what makes the answer checkable in the time it takes to click through, rather than something you have to take on faith or re-derive yourself from scratch.
Built for the person who has to explain the number next
The real audience for an evidence chip is not the person who asked Copilot the question, it is whoever that person has to explain the answer to next: a controller closing the books, an auditor asking where a figure came from, a co-founder who wants to see it themselves. Copilot’s answer only stays useful past that first read if it can survive being checked, so it is built to make checking it as easy as asking it in the first place.