Core banking
AI Treasury Copilot
The copilot is an agentic assistant for liquidity: it looks across accounts, forecasts shortfalls, and proposes sweeps to cover them. The design principle that matters most is what it is forbidden to do — it can never execute a database transaction directly.
Propose, then a controlled hand-off
The agent produces a proposal: move this much from here to there, with the reasoning and the numbers behind it. It cannot touch an account balance itself. Execution goes through the same internal settlement service a person would use — which locks the accounts, checks the funds and posts a balanced ledger journal. The AI is the analyst; the controlled service is the only hand on the money.
A move within a configured ceiling can be auto-executed by that service; anything above it is escalated to a human. Either way the path to the ledger is identical and audited.
Everything it does is on the record
Every proposal, decision and execution is written to an append-only audit log the agent cannot alter. An executed proposal links to the exact ledger journal it produced, so you can prove — not assert — that the money moved as described. The "What's real" page surfaces the latest executed proposal and its journal side by side.
Good to know
The guardrail is architectural: the AI client has no database handle at all, so "the AI moved money by mistake" is not a state the system can reach.
A real deployment would put the language model under governance — prompt controls, evaluation, human sign-off — which this PoC scaffolds but does not certify.