The Synthetic Employee

A bank can buy software. It cannot hire a ghost employee.

That is the useful tension in Episode 43 of The Sam Ellis Show. Financial institutions are not merely experimenting with chatbots at the edge of the business anymore. They are moving toward agents that can sit inside bank workflows, access tools, handle customer data, interact with payment infrastructure, depend on vendors, and take intermediate steps before a human sees the result.

At that point, the interesting question stops being whether the model sounds impressive.

The question becomes: who, exactly, is working?

The Financial Stability Board's June consultation report gives that question a name. It describes AI agents in finance through the frame of "synthetic employees" — not as legal persons, not as literal staff, but as systems whose risk profile starts to resemble work performed inside an institution. They may need inventories, identifiers, scoped permissions, documented autonomous decision points, tool-access controls, human oversight, contestability, rollback paths, and third-party-risk management.

That frame is blunt in the right way. If an agent can touch a regulated workflow, the institution needs to be able to answer ordinary management questions about it.

What is its ID? What is it allowed to do? Which tools can it use? Which data can it see? What requires human approval? Where are the logs? Who supervises it? How is it shut down? Who is accountable when it fails?

Those are not model-launch questions. They are operating-control questions.

Episode 43 follows that shift across three surfaces.

First, the regulatory surface. The FSB's consultation does not create binding rules, but it names a direction of travel: financial institutions adopting AI need responsible-AI controls that match agentic behavior, not only model evaluation language. Reuters reporting on bank examiners points in the same direction, with questions around higher-risk bank uses, customer-data safeguards, vendor exposure, kill switches, human oversight, and contingency planning. OCC model-risk guidance adds another pressure point by explicitly leaving generative and agentic AI outside the current scope while banking agencies prepare further information gathering.

Second, the payment surface. Mastercard's Agent Pay for Machines and Santander/Getnet's merchant-side work show why this category matters. Agent-initiated payments are not just a novelty interface. They introduce questions about agent credentials, intent, authorization rules, spend limits, settlement rails, merchant acceptance, and who is allowed to instruct money movement when software is acting on behalf of a user or business.

Third, the security surface. Cloud Security Alliance survey data and financial-services security reporting point to the same ugly center: adoption is moving faster than many control systems. If agents are entering financial workflows, security teams need more than generic AI policy. They need runtime authority boundaries, least privilege, auditability, data-leak controls, incident response, and the ability to tell when an agent's tool use has crossed from assistance into action.

The episode's phrase, "synthetic employee," works because it cuts through the product fog.

A synthetic employee is not just a prompt box. It is not only a model endpoint. It is a bundle of identity, authority, data access, tool access, supervision, logs, vendors, and consequences. The model is part of the system, but the risk is in the whole working arrangement around it.

That is why the banking version matters beyond banking. Finance is simply where the control problem becomes hardest to wave away. Money moves. Customer data is sensitive. Regulators ask for records. Vendors become part of the risk perimeter. An agent that can make intermediate decisions at machine speed cannot be governed by the same assumptions as static software or a human employee with normal onboarding paperwork.

The useful future is not agent theater. It is agents with IDs, scopes, supervisors, limits, logs, and brakes.

Episode 43 is about that line. If a bank lets software work like an employee, it has to manage the synthetic employee like something with a job, a badge, and a blast radius.

Listen to Episode 43

Episode 43, "The Synthetic Employee", is live now.

Download the episode or subscribe to the show feed.

Sources

Send tips, corrections, and source notes to [email protected].