I was one of the panelists at SPGI-FINOS Community Networking Panel: Agentic AI – and like all other FINOS events, I really enjoyed presenting about the topic of Agentic AI.

Me representing both a bank and FINOS gave me a rare and powerful dual lens here – to balance the practical enterprise adoption concerns (due to security, scaling, etc) with the open collaboration and interoperability advocacy.

My major core message was, that Agentic AI is not just about autonomous capabilities, it’s about trustworthy delegation in a regulated context. MCP, being the HTTP of AI, gives us a framework for this delegation to be standardized, secure and interoperable, across teams, tools and institutions.

Representing us, a bank, I shared:

• Risk & accountability – Agentic AI has to be traceable, auditable, explainable; cannot be a black box
• Delegation with oversight – Agentic AI always should operate under role based constraints, compliance layers, escalation protocols, just like any junior team member would be
• MCP is the Rosetta Stone for AI behaviors – with it, banks gain the uniform way to describe what agents can do, how they are governed, how they interop, across depts and firms

When I represented FINOS (and I put up a FINOS Ambassador hoodie – I would not go for the neighbor for some theatrics), I spoke about the other angle:

• Avoiding vendor lock in – future of agents cannot be tied to single vendor – FINOS pushing for the open governance, interchangeable agent skills, cross platform compatibility
• MCP enabling financial grade AI agents – just like FDC3 from FINOS standardized inter app comms, MCP standardizes agent governance and orchestration if we do so – opening the door for agent market place, p&p, composable trust and more
• Collaborative dev & shared protocols – banks do not need to reinvent governance wheels in silos, as FINOS do offer sandboxing, and hopefully soon sandbox for shared MCP schemas, safety patterns, audit templates and more: helping the industry move faster together, safer.

I touched some practical example, like:

• Agent skills via manifest saying “this agent can only draft emails, not send them”
• Versioned agent behavior, to track it like we do track code changes
• Cross firm collaboration for example for a KYC agent

I closed it with explaining how Agentic AI is not a product, it’s a new (junior) teammate. MCP helps us ensure that this teammate plays by the rules, explains its decisions, works well with others. Through FINOS, and their https://air-governance-framework.finos.org/ platform, we make sure those rules aren’t just good for one bank, they are good for the whole industry.

I would like to thank Olivier Poupeney for thinking about me and moderating us, Adam Dikker for organizing, Sudeep Kesh for bearing with me on the stage, Ryan Morris for his presentation and demos about S&P Global ‘s AIxtra platform, and Frank Tarsillo ‘s opening remarks.