Most demonstrations of AI agents in finance share a quiet assumption: that the goal is to get the human out of the loop. The agent reads the invoice, the agent pays the invoice, and the impressive part is that nobody touched it. For a demo, that is the whole point. For a regulated business, it is exactly the property you cannot ship.
The useful question is not "can the agent do the task?" It is "what happens the day the agent is confidently wrong?" If the answer is "money already moved", you do not have a product. You have a liability with a nice interface.
The pattern: propose, dispose, observe
The design that holds up under that question is older than AI — it is just separation of duties, applied to agents. Three roles, kept distinct:
- Agents propose. A worker agent does the analysis and stages an action — a payment, a reconciliation, a flag. It does not execute. It puts the action in a queue with its reasoning attached.
- Humans dispose. The action waits behind an approval gate. A person with the authority to act clears it, sends it back, or escalates. The agent's confidence is an input, never the authorisation.
- An oversight agent observes. A separate agent watches the proposals and the approvals — including the humans doing the approving — and reports anomalies outside the normal chain of command.
The power is in the seams. The agent that proposes cannot approve. The human who approves is themselves watched. The watcher reports sideways, not up, so it cannot be quietly overruled by the person it is meant to check.
Why the queue is the real product
It is tempting to think the model is the asset and the queue is plumbing. It is the other way around. Models improve every few months and are available to everyone. What a regulated customer cannot get off the shelf is the operating system around the agent: the queue, the role model, the audit trail, the kill switch that lives with a single accountable owner. That system is what makes "we put AI into production" a sentence an auditor will accept rather than wince at.
Why this is an Absolute Group bet
The Group runs its own operations on this pattern and offers it as infrastructure. Worker agents stage actions; humans clear them through an ActionGuard-style queue; WhistleBot watches both and reports outside the chain of command. The bet is that the durable value in agentic finance is not the cleverest model — it is the discipline of the queue around it.