Your bank says it is "ISO 20022 ready" for the November 22 cutover. Check the docs and it is usually a translation layer - MT to MX at the SWIFT boundary, MX back to MT inside. That is compliant, and it quietly throws away the structured data that was the whole point.
Subscription billing has two bad options today - store the card or chase a manual approval every cycle. Variable recurring payments add a third: pre-authorized variable amounts with limits enforced by the rail. The UK is building it now, and the lesson for everyone else is about regulatory cycles, not technology.
The CFO sees "save 3 percent on transactions" and the math looks obvious. Treasury sees continuous settlement at 3 AM and a reconciliation process built for batch. Before you switch to account-to-account, count the operating cost the savings slide leaves out.
Payments runs on documents - checks, wires, invoices, KYC files - and the OCR that reads them chokes on handwriting and damage. DeepSeek-OCR processes a page as roughly 100 visual tokens instead of a thousand text tokens, and it is open source. The headline accuracy comes with a compression ceiling worth knowing.
Train GPT-2 on Shakespeare and ask it the capital of France - it forgets Paris. That is catastrophic forgetting, one of several failure modes that surface the moment fine-tuning meets production, especially in regulated banking where "the model flagged it" is not an answer.
The vendor demo makes fine-tuning look like installing software - upload data, train, deploy. The real distribution of outcomes is brutal, most of the work is data preparation, and the model you trained for wire fraud can quietly forget how to read ACH. For financial services, guessing is not an option.