top of page

Multi-agent AI in banking: 59% experiment, 10% orchestrate, and that is what matters

Nov 10
4 min read

On 5 November 2026, the Capgemini Research Institute published its World Cloud Report for Financial Services 2027. Its finding fits in two numbers: 59% of financial institutions are already experimenting with or deploying multi-agent AI, but only 10% have enterprise-wide orchestration in place.


Between those two numbers lies, in our view, the real project of the next two years for a bank or an insurer. Not the agents themselves, but whatever coordinates them, constrains them and makes them auditable.


What the 5 November report says


The study rests on a survey of 1,100 financial services executives across 14 markets, run between June and July 2026, complemented by interviews conducted until September. It therefore captures the actual state of programmes, not intentions.


  • 59% of financial institutions are experimenting with or deploying multi-agent AI.

  • Only 10% have put enterprise-wide orchestration of those agents in place.

  • 90% use or plan intelligent routing between several models.

  • 52% of banks and insurers name hyperscalers as their preferred partners for governing multi-agent deployments.

  • Nearly half cannot consistently demonstrate the return on investment of agentic AI.


The sentence that sums it up comes from Ravi Khokhar, Capgemini's global head of cloud for financial services: adoption alone does not create value. In other words, multiplying agents without a shared architecture produces demonstrations, rarely results.


The report also notes a shift in funding criteria. The share of executives who tie the AI budget to business impact is expected to rise by 141% by 2028, while those who aim only for efficiency would fall by 42%.


Why orchestration is a governance matter before it is a technical one


A single agent that drafts summaries or prepares a KYC file is a tool. Ten agents handing over to each other between the front office, compliance and the back office form a process. And a process, in a bank, falls under internal control rules.


The first question is therefore not which orchestration framework to choose, but who is allowed to trigger what, with which data, and who validates the output. That is exactly the segregation-of-duties logic your auditors already apply to people.


The regulatory framework pushes in the same direction. FINMA Circular 2023/1 on operational risks requires an inventory of critical processes and their dependencies. An agent acting on a core system belongs in that inventory, like any other component.


On the European side, the DORA regulation has required since January 2025 that critical ICT providers be documented and that the resilience of the chains depending on them be tested. Orchestration entrusted to a single hyperscaler, as 52% of respondents intend, becomes a concentration risk in its own right.


FINMA made the point again in August 2026. In an interview reprised on 15 August, its chair Marlene Amstad stressed that shared dependence on a few technology providers can create risks that go beyond a single institution. The institution remains responsible, even when the provider runs the tool.


A Geneva private bank and its twelve agents


Consider a typical situation, drawn from what we observe in Geneva. A private bank with 400 staff let its teams experiment for eighteen months. The outcome: twelve agents in production, owned by five teams, running on three different platforms.


Three agents prepare portfolio reviews for relationship managers. Two pre-fill onboarding files. Four handle complaints or document requests. The last three live within IT and monitor data flows.


Each agent, taken on its own, is useful and reasonably well controlled. But nobody can answer a simple question from the audit committee: if the portfolio review agent relies on a data point corrected by the monitoring agent, who is accountable for the final recommendation?


The cost of this dispersion is measurable. In this kind of set-up we estimate it at the equivalent of two full-time positions spent reconciling activity logs, answering compliance requests and maintaining redundant connectors to the core banking system.


The answer is not to stop everything. It is to define a single orchestration layer, with an agent register, a data contract per agent and a shared audit trail. In our scenario, three of the twelve agents are merged, two are retired, and the remaining seven move under the register in less than a quarter.


The four workstreams of controlled orchestration


Moving from scattered experimentation to the 10% who orchestrate does not call for a gigantic programme. It does require four workstreams, handled in the right order.


  • The agent register: every agent has a business owner, a data perimeter, a list of authorised actions and a criticality level aligned with the FINMA 2023/1 inventory.

  • Segregation of duties: no agent may prepare, validate and execute a sensitive operation on its own. A human or an independent second agent steps in at least once.

  • The shared audit trail: every interaction between agents, models and core systems is logged in one place, in a format compliance can read without help from IT.

  • Value measurement per use case: each agent carries a business indicator, such as processing time, manual rework rate or cost per file, reviewed quarterly. This is what half of the institutions surveyed lack.


Model routing, which 90% of institutions plan, fits naturally into this layer. It lets you reserve the most expensive models for the tasks that justify them and keep sensitive data on models hosted in Switzerland or in the EU.


On hyperscalers, our position is nuanced. Relying on their orchestration tooling is reasonable. Handing them the definition of the rules, the ownership of the logs and the segregation-of-duties logic is not. Those three elements must stay inside the bank.


What we take away


The Capgemini report of 5 November confirms what we see in the field: banks no longer have an adoption problem with agentic AI, they have a coordination problem. The 59% who experiment have done the most visible part. The 10% who orchestrate have done the most useful one.


For a risk function or an IT department, the good news is that orchestration looks like what you already know how to do: inventory, separate roles, log, measure. The difference is that the actors are now agents, and they act fast.


Our conviction: a bank that sets up its agent register and its shared audit trail in 2027 places itself in the 10% by that very act. One that waits will discover its coordination problem during an audit, at the worst possible moment.


To locate your institution on that path, the 39 Advisory AI Barometer gives you, in fifteen minutes, a reading of your maturity in agent governance. And if you would rather talk it through, our team in Carouge is at your disposal.


 
 
 

Comments


bottom of page