1. Traditional financial services organisations were built around human decision makers and defined processes. Which of those structures is the first to fail when AI systems start generating and acting on intelligence across teams, and what does that failure look like day to day?
The first thing to break is the approval chain, specifically the assumption that a human reviews output before it moves. Traditional financial services processes are built around a decision point: an analyst prepares something, a manager reviews it, and it goes forward. AI systems don’t wait for that cadence, though. Instead, they generate at a pace and volume that makes case-by-case human review either impossible or purely theatrical, a rubber stamp on something nobody actually read. Day to day, this looks like backlogs of unreviewed recommendations, or reviewers who start approving in bulk because the volume has outstripped their capacity to actually check anything.
2. Governance is the word everyone reaches for, and it usually means a committee. What does an operating model that genuinely works with AI
look like in practice? Who owns a decision when an AI system made it, and who is accountable when it is wrong?
Governance-as-committee fails because a committee is built to make decisions periodically, not continuously. What works instead is decision rights defined by category rather than by instance: before the system runs, someone specifies which classes of output can be acted on automatically, which require review, and which require escalation, revisited on a schedule rather than argued about case by case.
Accountability follows the same logic: 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. The mistake is treating the AI’s involvement as diluting accountability when it actually should sharpen it, because now the boundaries are written down instead of implicit in someone’s judgment.
3. Decentriq’s own field is confidential computing and data collaboration, which is a particular vantage point on this. What does
that work show you about AI readiness that someone selling models or agents would not see?
Selling models or agents, you see whether the technology works. Selling data collaboration infrastructure, you see whether organisations are willing to let their data be used at all, and that’s a much earlier and more revealing test.
We recently supported a collaboration between a leading global wealth manager and the New York Times, aimed at reaching high-net-worth prospects with precision. The wealth manager was legally not allowed to hand over raw customer data to a publisher under any circumstances, and the publisher had the same constraint in reverse: it needed to protect reader data while still proving the value of its audience. Neither side’s blocker was the targeting model. Rather, 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.
And while this is an example of successfully overcoming this hurdle, it’s one that creates a pattern across financial services generally. The stall isn’t really about the technology, and it isn’t really about the AI either. It’s a signal that the organisation hasn’t resolved its own internal comfort with data use, and no model, however capable, fixes that.
4. If the organisations that get the most from AI are not those deploying the most technology, how would a bank actually know whether
it is ready? Is there a test, or a signal, that separates the firms that are from the firms that think they are?
The technology-deployment signal is close to useless, because everyone can deploy something now. 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. If a project stalls because a model needs retraining or more compute, that’s a normal technology problem with a normal fix. If it stalls because nobody can say who signs off, or because two departments both assume the other owns the outcome, that’s the real test failing.
A useful contrast is a Swiss bank we worked with that had already resolved this before any campaign began. As a regulated institution, it couldn’t hand its customer data to a publisher under any circumstances, so the decision about how first-party data would and wouldn’t be used was settled at the outset, not negotiated mid-project. The campaign that followed, activating first-party audiences with a premium local publisher instead of purchased segments, saw a 129% increase in click-through rate and a 44% reduction in cost per page view, and the bank has since made the approach standard for its future campaigns with local publishers. 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. Firms that are ready can usually point to exactly who owns a given AI-driven decision before the project starts. Firms that think they are ready can point to the technology stack and not much else.
5. What is the most common mistake you see financial institutions making right now on this, and what is the honest cost of getting it
wrong? We would rather have a specific and uncomfortable answer than a diplomatic one.
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.
It doesn’t sort itself out. It usually surfaces as a slow, expensive stall: a system or product that’s technically live but practically idle, generating a return somewhere well south of what was modelled in the business case, while the licence, integration and maintenance costs keep running regardless of whether anyone’s acting on the output.
The honest cost is that nobody notices this kind of failure until much later. A cancelled project gets a post-mortem and a line in next year’s budget as a lesson learned. A stalled one just keeps drawing down cost every quarter with no one accountable for closing it out, because on paper it’s live and technically a success. Multiply that across every business unit that onboarded its own AI capability the same way, and the real cost is a system nobody ever measured against what it was supposed to deliver.
Author Bio
Maximilian Groth is the co-founder and CEO of Decentriq, a Zurich-based data collaboration company.
