Q&A: AI and Trade 3: A field guide for trade professionals

A question for the speakers, if a narrow one is welcome. Nobody has asked yet, so I will go first.

My work is classification. I write the determination and I sign it, and the answer is the small part of it. The file underneath is the real output: somebody should be able to open it years afterwards and follow why. In my market the audit arrives long after the shipment, and often after the signer has left the desk.

So when a tool proposes a classification, or clears a party in a second, what should the desk be writing down? And what would an auditor accept as the reasoning, years later, when the signer cannot reproduce how the tool got there?

Most of my day is checking: an invoice against a packing list, a classification against what we declared on the last shipment, the file put together so someone senior can sign it. I am a year into this, and that checking is how I am supposed to be learning. The people who sign got their judgment by doing years of it. If that layer is the first to go, what should someone at my stage still be able to do by hand, and where is the judgment meant to come from instead?

Aya's line about the file underneath being the real output landed hard, because on my desk nobody but me will ever open it. I am one person on my export side: no compliance department, nobody to overrule me, no budget for either. Last week I had which party carried a tax obligation wrong in my own post here, and I caught it myself only because I went back.

The advice is to open the source a tool cites and confirm it says what the tool claims. If I can judge that, I did not need the tool. So, for the speakers: what tells a desk working alone that an answer is right rather than fluent?