A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Payments And Banking: A Bundle For The Shape Of The Answer," Kin Lane, November 6, 2026, https://apievangelist.com/2026/11/06/payments-and-banking-a-bundle-for-the-shape-of-the-answer/
Imran kept the invoice open in one window and the post in another. He worked in accounts payable at a mid-sized distributor, and his team had been trying out a model that read supplier PDFs and drafted payment files. Nobody had yet said out loud what would happen when it got one wrong.
He started with the part he liked. The post says "Money is a string," and explains that a decoder will happily emit 1250.1000000000001 if you let it. He had seen a spreadsheet do almost exactly that to a remittance total once, and he had spent an afternoon explaining it to an auditor. Seeing the amount pinned to a two-decimal pattern felt like someone finally writing down a rule he had been carrying in his head.
Then he got to the section on what the decoder cannot do. The post says "The grammar can’t do arithmetic," and that the transaction count and the control sum can only live in rules. Imran noticed he was uncomfortable. His team had assumed that a structured output meant a correct one. The post's own fixtures said otherwise: the decoder accepted a broken document with five separate faults, and only the rules caught them. Nobody on his team had written a rule. They had a schema and a good feeling.
He pushed back, though, on the retry setting. The post retries once and then goes to a person, and calls that "Fewer retries than the other bundles, on purpose." Imran thought one retry was still generous. If the model gets a control sum wrong, he did not trust it to get it right on a second pass just because it had been shown an error message. He would have sent it straight to a human. He also did not love that the post's date check says nothing about cutoff times or bank holidays, which is exactly where his own files bounced most often. The author admits this, which helped, but admitting it did not fix it.
The last section made him put his coffee down. The post says a correct IBAN that belongs to someone else "isn’t caught by anything here," and that a supplier changing bank details by email is a fraud pattern a model could make worse by extracting the new details perfectly. Imran thought of the three change-of-bank requests sitting in his inbox that week. One of them had come with an invoice already approved.
He did not build a bundle. He did not even open a schema file. What he did was smaller. He opened the shared document where his team described the pilot and added a single line under the risks: before any model-drafted file is released, a person confirms any changed supplier bank details by calling a number already held in the vendor master, not one from the email. Then he sent the line to his manager and asked whether it could go in before Monday's pilot review.