The back office, run by digital agents.
Agents that watch the platform, reason about what to do, and draft the work — then wait at a human approval gate. The judgement is automated; the irreversible act stays with a person.
Operations shows what's happening. The models say what's coming. The agent layer does the next thing: it drafts the response — the settlement to run, the follow-up to send, the exception to resolve — and puts it in front of a person to approve.
This is deliberately not "AI that moves money." Every agent is read-and-propose, never read-and-execute. It can assemble a payment, reconcile an exception or chase a missing number, but the button that actually moves a dollar is pressed by a human, every time.
Watch, reason, draft — then wait for a person.
A shared inbox of proposals
Every agent action lands as a proposed item in one queue — the settlement to run, the return to handle, the invoice to chase — waiting for a decision, not already done.
Transparent reasoning
Each item shows its work: the data it read, the rule it applied, the number it's acting on. An approval is an informed decision, never a leap of faith.
A human approval gate
Nothing irreversible executes without a person. Approval policy — dual sign-off, per-tenant thresholds — applies to an agent exactly as it does to an operator.
Read the whole platform
Agents can see across payments, the ledger, tenant health and the forecasts from Layer 2 — scoped to what the asking user is already permitted to see.
Draft, don't execute
An agent assembles the work — a payment order, a reconciliation, a message — and stages it. The money act is the one thing it always hands back to a person.
Fully audited
Every proposal, every approval and every rejection is logged — the same audit trail as any admin action, so an agent is never a side door around the controls.
Automate the judgement. Keep the person on the trigger.
The hard, repetitive judgement of a treasury back office — what to settle, what to chase, what to flag — is exactly what an agent is good at, and exactly what burns out an ops team. Automating it is the point.
But moving money is irreversible, and the platform's whole design is that a number it can't verify is a number it refuses to act on. Agents inherit that discipline: they propose, a person disposes, and the audit trail records both.
See → predict → act.
Agents are the acting layer on top of live operations and the predictive models beneath them.
Request access → See the whole platform →