59% of Financial Firms Are Running Multi-Agent AI. Only 5% Actually Have It Working.

Capgemini's World Cloud Report for Financial Services dropped today with a headline that will fill a lot of executive decks this quarter: 59% of financial institutions are piloting or running multi-agent AI in production. That number comes from 1,100 financial services executives across 14 markets, surveyed over the summer.

Skip to the footnotes.

The same report finds that only one in ten of those institutions has achieved the full orchestration required to run multi-agent systems autonomously across the enterprise. And only 5% — one in twenty firms surveyed — have actually institutionalized the capabilities required to operate agentic AI at enterprise scale.

That's the funnel: 59% running, 10% running well, 5% embedded into how the organization actually works. It's not a laggard problem that patience solves. It's describing the same firms, at the same moment, against the same standard.

The Returns Problem Is Downstream of the Measurement Problem

The number that sits next to the deployment figures is worse. Nearly half of the financial organizations building multi-agent systems within core banking and insurance workflows cannot consistently demonstrate returns on those investments. These aren't experiments being wound down. These are production deployments — agents that passed procurement, got integrated, and went live. The organizations running them can't show that they work.

The conventional read is that multi-agent AI is hard and enterprise deployment is slow. That's true and irrelevant. The gap between 59% deployed and 5% institutionalized is not primarily a technical complexity gap. It's a measurement gap. If you don't have a framework for evaluating whether an agent is performing well enough to extend its scope — to trust it with more decisions, more autonomy, deeper integration — you don't get from "this is running" to "this is reliable enough to depend on."

You get to "we're not sure if it's working," which is exactly what the returns data describes.

What the 5% Did Differently

The Capgemini report identifies a small group — 5% of respondents — that has cleared the institutionalization bar. The report doesn't publish a detailed breakdown of what distinguishes them, but the data points cluster around a few patterns.

90% of firms are using or planning intelligent model routing to manage performance and control token costs. Model routing is a signal that an organization understands agent performance as a managed variable, not a fixed property certified by a vendor before deployment. A firm that routes tasks to different models based on complexity and accuracy requirements is implicitly maintaining performance standards across its agent fleet. It's doing the continuous measurement work that the other 95% mostly skips.

Banks are also increasingly adding AI supervisors — agents that watch other agents, flagging behavior outside expected parameters. That's a runtime control, but it also generates behavioral data: logs of what agents do in production, under conditions the demo never covered, against real edge cases. Organizations that accumulate that data eventually have something to evaluate against. Organizations that don't are flying blind and calling it deployment.

The 5% are probably not smarter than the 95%. They're more likely the ones who decided that "it's running" and "it's working" are different claims that require different evidence.

Financial Services Is the Canary

Financial services is worth watching here not because the numbers are unique, but because the feedback mechanisms are faster and the stakes are higher than most industries.

A multi-agent system that fails to demonstrate returns in a bank doesn't get a quiet budget cut in the next planning cycle. It becomes a line item in a board presentation, a question from a regulator, a risk management conversation about whether the system should be expanded or contained. The scrutiny is higher, so the gap between "deployed" and "working" shows up faster. What takes two years to surface in a retail deployment shows up in two quarters in banking.

If 59% deployed and 5% institutionalized is where financial services is, six months into serious multi-agent deployment at scale, the other industries are behind that curve, not ahead of it. The 59%-to-5% gap is coming for enterprise healthcare, enterprise retail, enterprise manufacturing — but those sectors are slower to publish the data, and the feedback loops are softer.

The Infrastructure Question

One finding from Capgemini is easy to overlook: 52% of banks and insurers are turning to hyperscalers as their preferred partner to govern multi-agent deployments. That's not a build-versus-buy decision. It's an admission that the governance infrastructure for multi-agent systems is not something most financial institutions can build themselves.

The same logic applies to performance infrastructure. If you're running dozens of agents across different workflows, each from different vendors, evaluated against different internal benchmarks with different definitions of "good," you don't have an agent fleet. You have an inventory. The governance layer tells you what the agents are doing. It doesn't tell you which ones are worth extending and which ones are costing you money.

That question — before deployment, and continuously in production — is what the institutionalized 5% is answering. The other 94% isn't there yet, and the route there is not more deployment. It's better verification.

The Number That Matters

The headlines will carry 59%. It's a legitimizing number — more than half of the industry is in, the technology is real, the adoption curve is confirmed.

The number that tells you something useful is the gap between 59% and 5%. That gap is not a waiting room. It's a measurement problem with a measurement solution. The organizations that close it won't do it by running more agents. They'll do it by building the capability to tell the difference between an agent that works and one that runs.

That's a harder capability to ship than the agent itself. It's also the one that determines whether the 59% eventually becomes a productivity force or a rounding error on the balance sheet.

---