Have banks given AI agents responsibility too early?
I think that in certain cases where there is a bit of experimentation going on, we have likely done that. For the industries that we serve, which are highly regulated, we have to make sure that we have the right guardrails in place. We need to make sure there are things like auditability and observability, and that for any agentic operation the human is always in the loop and the human is always on the hook. That is needed not just from a legal liability perspective, it is also good best practice.
The agentic, autonomous processes and the technology being used today can certainly help our operations, but ultimately we have to make sure that a human is always on the hook.
What does agentic AI mean for governance?
Ultimately, what autonomous and agentic operations do today is that you describe the what and the agent figures out the how. Therefore we have to make sure that the how has guardrails on it, and that is how we should design the system.
I do think this combination of traditional technology, which has deterministic, rules-based approaches, together with agentic operations, which have probabilistic approaches, is really about getting the right combination and balance of both. That is where we will see the best outcomes.
How do you guide a workforce with different skill sets in using these tools?
Ultimately it is about culture, and about how we encourage the use of these tools and technologies, both individually and across the whole enterprise, to get the outcomes that we need. Last year was largely about pilots and experimentation. This year it is really about moving from pilots to production. We believe the sweet spot in today’s era is human intelligence plus artificial intelligence.
What advice do you have for moving from pilot to production?
I always think you need the pilots. You need the experimentation, because this is how you scale up, discover and explore what is viable and what is possible. But ultimately everything has to have a business case in the world we work in, and everything has to operate within guidelines and guardrails. The use cases you choose really determine how you can get success. If you choose the right cases that will deliver the right outcomes, that puts you in a good spot.
We also need a mindset where we are redrawing existing processes and existing approaches. Tinkering around the edges will give you a better outcome, but it will not give you a dramatic outcome. We have an opportunity to be very optimistic and to have dramatic outcomes, so we have to start by completely redrawing what we think the outcome looks like and how we might get there.
Guardrails and accountability for agentic AI
Banks are moving from AI experiments to autonomous operations, and highly regulated industries cannot afford to skip the controls. In this FinextraTV interview at Sibos 2026, Akber Jaffer, CEO of Smartstream, explains why auditability, observability and a human who is always “on the hook” are essential for any agentic operation.
For firms exploring Smart Agents agentic AI for financial operations, the principle is simple: describe the outcome, let the agent work out the how, and design guardrails around that how.
Balancing deterministic rules and probabilistic agents
Effective governance does not replace traditional technology. Jaffer points to a combination of deterministic, rules-based systems and probabilistic agentic approaches, and to human intelligence plus artificial intelligence as the sweet spot for today’s operations.
From pilot to production: redraw the process
Pilots remain essential for discovering what is viable, but every use case needs a business case and clear guardrails. Jaffer argues that tinkering around the edges delivers a better outcome, not a dramatic one, and that firms should start by redrawing the outcome they want. Watch the full interview above, or speak with Smartstream about agentic AI use cases.
