Agentic AI Growth Continues Amid Operational and Compliance Warnings

The slow rollout of multi-agent artificial intelligence systems and growing concern over the risks of poorly governed AI models suggest current iterations of the technology may be testing the limits of its effectiveness – even as financial institutions press on with deployments.

In each case, the same factor had been identified at the root of the shortcomings and also as the basis of any solution: data management.

Tech Stacks

Capgemini’s World Cloud Report for Financial Services 2027, published late last week, found that isolated agentic AI trials were giving way to multi-agent systems. Three-fifths of respondents to two surveys said they had put such strategies to the test. So far, however, only a tenth of them had been able to achieve the level of orchestration needed to run such arrays autonomously at scale.

The study, based on the company’s 2026 Global Financial Services Executive Survey and its Global Executive Interviews, which engaged more than 1,100 leaders worldwide, noted that 80% said they would prioritise the upgrade of their infrastructure, recognising that fragmented legacy systems were inadequate for their AI goals.

“To scale successfully, banks and insurers recognise that modernising their underlying tech stack is a prerequisite to succeed in the agentic AI era,” the report stated.

Slow ROI

Organisations’ difficulties in achieving a swift return on their AI investments has seen them switch their development investment focus, the report added.

“In just 12 months, financial institutions have moved beyond experimenting with individual AI agents tackling isolated tasks to deploying multi-agent systems that coordinate actions across entire business lines,” said Ravi Khokhar, Global Head of Cloud for Financial Services at Capgemini. “Yet adoption alone does not create value.

“Successfully scaling these systems requires infrastructure that drives end-to-end orchestration, effective governance and security by design, and clear link into bottom-line impact,” Khokhar said, adding that the companies that are “pulling ahead” are doing so because they have incorporated cloud as an operational foundation into their strategies.

Regulator’s Warning

Data management was central to warnings by the UK’s financial regulator, the FCA, about weaknesses in AI deployments among the nation’s banks. Along with inadequate oversight and gaps in risk assessments, failures in managing data adequately were blamed for exposing institutions to operational and regulatory risks.

In its most recent AI update publication, the regulator pointed to the case of one bank that had allowed erroneous AI-generated code recommendations to proceed into its systems, resulting in a “loss of service” incident that rippled through the enterprise.

As well, the unauthorised use of AI tools by employees led to the leak of sensitive bank information at other institutions. Other shortcomings found by the FCA included incomplete inventories of AI systems, inconsistent risk ratings for AI agents and the breakdown of formal approval procedures.

The overseer issued a raft of good-practice recommendations that included identifying and planning the needs and objectives of any AI project before implementation, and testing applications for biases before their operationalisation.

Importantly, it urged the creation of governance layers for AI that include “escalation routes and compliance controls that maintained accountability, human oversight and risk management across complex workflows”.

Human Inclusion

The criticality of retaining humans in AI processes data was stressed in a separate report by the Group of Boutique Asset Managers (GBAM), which said its members at the smaller end of the market could offer better value to their customers if the technology worked in conjunction with human expertise.

“The one thing we know of our ability is that it is based on superior judgement. That’s what we bring to the party. It’s judgement; it’s the ability to find context in things that can’t be easily read as right or wrong,” GBAM chairman, Tim Warrington, told delegates to the body’s conference in Florence, Italy.

The group includes KAGEN Funds, MAPFRE, Aubrey Capital Management, Nextam Partners, First Avenue, Ranmore Fund Management, Azvalor Asset Management and LarrainVial Asset Management.

Good Data, Bad Data

A common thread in the recent slew of surveys and reports on AI’s utility underline a a common thread; that the most critical ingredient to successful AI adoption is good data.

The fundamental importance of good quality and well managed data was highlighted in a contributed article to Data Management Insight this week, by Gus Sekhon, VP Strategy, FINBOURNE Technology.

In his piece, Sekhon argued that across multiple asset classes, well maintained and controlled data is vital for trust.

“AI can certainly help by speeding up data processing and in identifying any inconsistencies. But it cannot manufacture reliable valuations or resolve conflicting records without reconstructing historical information that was never properly looked after,” he wrote.

“Indeed, deploying AI across fragmented, poorly governed data risks making the problem even worse. Faster analysis of unreliable information is still unreliable analysis and what investment managers need is a more disciplined approach to their underlying data architecture.”