What was the cash position at 11:47am? On a snapshot-based model the honest answer is an interpolation between two saved states. Real-time rails turned that guess into a liability, and the fix is to stop storing a position at all and fold an immutable event log into the answer. The event log is not free, and most teams have not put the bill on the budget.
Your real-time rails settle in fifteen seconds. Your treasury org chart was built around an end-of-day cutoff. The gap between them is where the next reconciliation loss lives, and it shows up in three roles - cash positioning, exception ops, and intraday risk - long before it shows up in the tech stack.
The widely cited figure is that 88 percent of AI proofs of concept never reach production. Agentic payments inherit that gap, and the teams that cross it are not winning on model selection. They are winning on payments-specific infrastructure that makes an agent idempotent, governable, and rollback-safe at the rail boundary.
The multi-rail debate is over - most US banks enabling instant payments now run both RTP and FedNow. The cost moved from the strategy deck to the operations floor, where exception handling becomes a decision tree, reconciliation becomes real-time position-keeping across two liquidity models, and routing is a per-transaction decision in milliseconds.
The translator is ready, the mapping is tested, the structured fields are wired - and the project is still not on track, because the customer records are missing structured address data across KYC, core banking, and correspondent files. In November 2026, SWIFT, SEPA, and CHAPS start rejecting unstructured addresses, and most banks are framing the wrong problem.
Payment modernization projects do not stall at the architecture layer. They stall at the security review layer - threat models built for monoliths, questionnaires written for general software, cycles tuned for annual releases. The teams that ship on time put security in the room before design, not after.