Shipped AI into a live bank last year. Honest lessons!

I spent most of last year getting a multi-agent AI system into actual production at a European bank. Real transactions, not a sandbox.

Accuracy wasn't the issue. Auditability nearly killed us. Banks don't ask "is it right" — they ask "can you show a regulator exactly why it did that." We rebuilt half the product around this after the first compliance review.

Also... nobody mentions that the orchestration takes longer than the model itself. Fallbacks, confidence handling, who reviews edge cases at 11pm. The LLM part was maybe 20% of the work.

And pricing. They want fixed costs, you have variable inference costs. That contract conversation took longer than building the thing. Solve that before you get to legal.

Anyone building AI for regulated industries — what caught you off guard???