Fintech Expert Declares AI Readiness a 'People Problem,' Not a Tech Hurdle

As financial institutions increasingly integrate artificial intelligence into their daily operations, the critical challenge shifts from understanding what the technology is capable of to transforming how firms are structured to effectively utilize it. Maximilian Groth, CEO and co-founder of Decentriq, a company specializing in confidential computing and data collaboration, contends that a fundamental change in the approval chain is necessary. He observes that the organizations deriving the most significant benefits from AI are often not those simply deploying the greatest volume of AI systems.

The discourse surrounding AI in financial services has predominantly focused on its capabilities—its models, speed of improvement, and potential applications. However, a more subtle, yet crucial, question persists: are banks and fintech firms truly equipped to implement this technology effectively? This query gains urgency with the EU AI Act’s Article 50 transparency obligations coming into full effect from August 2, 2026, and as UK regulators explore the reach of autonomous, agentic systems into retail finance. Both regulatory frameworks underscore a central issue: accountability for decisions in which machines play a role.

Groth’s perspective is distinct, shaped by Decentriq’s mission to provide infrastructure that enables two organizations to combine data without either party accessing the other’s raw records. His understanding of AI readiness is thus heavily influenced by firms’ willingness to use their data at all, rather than solely by a model’s performance. While Decentriq has a commercial interest in this argument, the underlying premise—that readiness is about evolving operating models rather than just acquiring tools—merits careful consideration.

When asked which part of a traditional financial services organization falters first when AI systems begin generating and acting on information across teams, Groth pinpoints the review step that underpins most processes. He states, “The first thing to break is the approval chain, specifically the assumption that a human reviews output before it moves.” AI systems operate at a pace and volume that render case-by-case human review either impossible or merely ceremonial, reducing it to a rubber stamp on content that remains unread. This often manifests as backlogs of unreviewed recommendations or reviewers who resort to bulk approvals, overwhelmed by the sheer volume.

For many firms, governance is the immediate solution proposed for this weak link, typically involving a committee. Groth argues that a committee-based approach is ill-suited for the problem. “Governance-as-committee fails because a committee is built to make decisions periodically, not continuously,” he explains. He advocates instead for decision rights defined by category rather than by individual instance. This means that before a system operates, specific classes of output are pre-defined as actionable automatically, requiring review, or necessitating escalation, with these classifications revisited on a schedule rather than debated case-by-case.

Regarding accountability when an AI system makes an erroneous decision, Groth emphatically rejects the notion that the technology dilutes responsibility. He asserts, “If an institution deploys a system and defines the boundaries of what it's allowed to decide, the institution is accountable for the outcome within those boundaries, the same as it would be for a junior analyst acting within their mandate.” He views treating AI involvement as a dilution of accountability as a mistake; instead, it should sharpen it, as the boundaries are explicitly documented rather than implicitly relying on human judgment.

Groth emphasizes that Decentriq’s confidential computing business offers a unique insight that a vendor solely selling models or agents might miss: an organization’s fundamental willingness to utilize its data. He illustrates this with a collaboration Decentriq supported between a global wealth manager and a large publisher, aiming to reach high-net-worth prospects. Both parties faced legal prohibitions against sharing raw customer data. “Neither side’s blocker was the targeting model. It was that the deal could not exist at all until there was a way to combine signals without either party seeing the other’s raw data,” Groth notes. This situation reveals that the impediment is not about the AI technology itself, but rather an organization’s unresolved internal comfort with data usage—a problem no model, however advanced, can rectify.

This leads to Groth’s practical test for AI readiness: since nearly any firm can now deploy some form of AI, the sheer quantity of technology in use reveals little. A more telling indicator, he suggests, is the “friction location”—where a specific AI initiative encounters obstacles. “A better signal is friction location: watch where a specific AI initiative gets stuck, and check whether it’s a technical blocker or a decision-rights blocker,” he advises. If the project stalls due to a need for model retraining or more computational power, it’s a standard technical issue with a standard fix. However, if it halts because no one can identify who has signatory authority, or if two departments both assume the other is responsible for the outcome, that signals a genuine failure in readiness.

He contrasts this with a Swiss bank Decentriq assisted, which had meticulously defined how its first-party data would and would not be used before any campaign commenced. According to Decentriq’s reports (which are not independently verified), the subsequent campaign achieved a 129 percent increase in click-through rate and a 44 percent reduction in cost per page view. Groth clarifies, “None of that came from a better model. It came from the bank already knowing what it would and wouldn’t allow before the technology was ever switched on.” He concludes that truly ready firms can typically identify exactly who owns a given AI-driven decision before a project begins, whereas those who merely perceive themselves as ready often point only to their technology stack.

The most common and costly mistake, Groth observes, is treating AI readiness as a procurement decision. “The most common mistake is treating AI readiness as a procurement decision: buying the capability and assuming the operating model will sort itself out once people see the results,” he explains. This assumption is flawed, often leading to a “slow, expensive stall”: a system that is technically operational but practically idle, generating returns far below initial projections, while licensing, integration, and maintenance costs continue to accrue irrespective of whether its output is acted upon. Groth laments that this type of failure often goes unnoticed until much later. A cancelled project undergoes a post-mortem, but a stalled one perpetually drains resources without accountability for its closure, as it remains technically “live” and superficially a success. The impending EU AI Act’s transparency requirements, effective August 2, 2026, and extended preparation time for higher-risk systems, amplify the urgency of this distinction. The crucial question is whether this additional time is used to establish clear ownership for AI-assisted decisions or merely to acquire more technological capability.