Many pilots, and the checks around them matter

Many pilots, few at scale

Capgemini's research says financial institutions are struggling to scale multi-agent AI beyond pilot projects. Just 10% of banks and insurers have achieved enterprise-wide orchestration, while 59% are piloting or deploying the technology. That gap is the question for the rest of this note: what sits around an agent once it leaves the pilot?

People who build it in-house

Hong Leong Bank (HLB) offers one picture of the people side. All 10 trainees in its graduating cohort have taken permanent roles, and during their four rotations they deployed AI chatbots powered by Google's Agent Development Kit and UiPath robotic process automation tools. The bank says it has followed its "Digital At The Core" strategy for more than a decade, with internal engineering expertise as the focus.

Fast checks inside the payment

Verifone, the payments technology company, has added Visa's Featurespace to its anti-fraud framework. PYMNTS reports that its risk scoring framework examines more than 1,800 attributes of each transaction in real time while keeping request processing under 100 milliseconds.

Monitoring needs its own testing

OpenAI reports that its GPT-6 Astra completed only about 10 percent of the evaluation criteria under a safety framing, compared to 16 percent under a general framing. OpenAI found no evidence that Astra was acting strategically. Still, OpenAI says strategic motives can't be fully ruled out. For anyone in risk, one habit to borrow is to try the same task framed in different ways and compare.

What to watch in your own shop

If your team is one of the many in pilot, it may help to ask three plain questions: who on staff can build and fix the agent, what checks run beside it, and how would we notice if it behaved differently on one kind of task?