Buenos Aires, Argentina
The hard part of agentic AI in a bank is not the model. It is the boundary.
Agentic systems inside organisations governed by ISO, SOX and ITIL, where autonomy has to be bounded and auditable.
The constraint is the point
Agentic systems are straightforward to demonstrate and difficult to put into production inside a regulated institution, because the interesting property — that the system decides on its own — is exactly the one the regulator, the auditor and the risk committee need bounded.
The useful question is not what the model can do. It is what the system is allowed to decide without a human, how that is evidenced afterwards, and who owns the outcome when it is wrong.
How I approach it
- Bounded autonomy
- An explicit account of which decisions the system takes alone, which it proposes for review, and which it never touches — written before deployment, not after the first incident.
- Traceability by design
- Every material decision reconstructable after the fact: inputs, version, rationale and reviewer. If it cannot be evidenced in an audit, it cannot be in production.
- Human review where the regulator expects it
- Placed deliberately at the points that carry regulatory or customer harm, rather than spread evenly as a precaution that slows everything and protects nothing.
- Inside the existing control frameworks
- ISO, SOX and ITIL for operations, risk and service management; TOGAF and the C4 model for architecture and how it is documented. Decisions have to arrive in a form the organisation can adopt, audit and defend.
Where this applies
The cases that justify the effort are rarely the ones in the vendor demo. In practice they are operational: claims triage at an insurer, document and evidence handling in compliance work, reconciliation and exception handling in payments, and the manual process that sits between two systems nobody wants to replace.
The value there is measurable as operating cost, which is also the only form in which a board will agree to fund it.
Current work
As Partner, Management Consulting at COA LATAM, I lead regionalisation strategy across fintech, banking and insurance, including the adoption of agentic AI frameworks with guardrails in regulated, multi-country environments.
Contact
If you are building, rebuilding or rescuing a technology organisation, write to me at hernan@hernancurras.com. A first conversation costs nothing and usually takes thirty minutes.
Also on LinkedIn.