Multi-Agent AI in Banking and Insurance: Orchestration Lags Pilots

What happened

Only 10% of banks and insurers have achieved enterprise-wide orchestration of multi-agent AI, even though 59% are piloting or deploying the technology, according to Capgemini research reported by Finextra. The figures point to a scaling problem rather than a lack of interest in the technology.

The findings suggest financial institutions are moving past early experimentation but struggling to turn isolated pilots into coordinated, organisation-wide systems. Orchestration, the layer that coordinates several AI agents so they can work together on a process, appears to be the capability lagging furthest behind.

Why it matters

Multi-agent systems, in which several AI agents handle different parts of a task and pass work between one another, could change how banks and insurers run processes such as onboarding, claims and servicing. The research indicates that making these systems work across an entire institution, rather than within a single team, is the harder task. Leaders who treat agents as a collection of separate pilots may find that their operations improve in pockets while the end-to-end process stays fragmented.

For experience leaders, the practical implication is that customers do not experience individual agents. They experience the handoffs between them. A customer who must repeat information because one agent cannot see what another has already done will judge the institution on that failure, however capable each component is on its own.

By the numbers

• 10% of banks and insurers have achieved enterprise-wide orchestration of multi-agent AI.
• 59% of banks and insurers are piloting or deploying multi-agent AI.

The Renascence take

Most coverage of this research will frame it as a maturity gap. We read it as a design gap, and the distinction matters for where investment goes next.

Many institutions are building agents around internal tasks and departmental org charts, then wondering why the customer experiences a relay race with dropped batons. Start by mapping the end-to-end journey and assigning a single accountable owner to every handoff, whether human or machine. Measure success by how often a customer has to repeat themselves, not by how many agents are live. Orchestration is a service-design problem before it is a technology problem.