The most revealing thing about AI in the boardroom may be how often it now shows up as a leadership problem rather than a technology one. Of 900 CEOs, 78% believe a failed AI strategy could cost them their job. This was not another statistic in the market. It made me think – how accountable are CEOs for what AI does to the company and its customers? With AI-specific roles now load-bearing the AI sack, why are the Chief Executives tying their tenures to AI’s scorecard?
Into that climate, I reached out to three CEOs from very different fields: enterprise software, banking infrastructure, and cybersecurity, and asked them about platform risk, governance visibility, and personal exposure. Their answers describe a pattern: ambition is intact, control is not, and each leader has built a different theory of how to close that gap.
Florian Douetteau, co-founder and CEO of Dataiku, talks about AI bets that become load-bearing architecture. Jouk Pleiter of Backbase describes the risk of deploying intelligence on top of fragmented systems that were never ready for it. Benny Czarny at OPSWAT frames the problem through the lens of security, where AI can quickly become more noise than control if it is applied without discipline. Together, they sketch a more complicated picture of AI leadership than the one that usually appears in public: we are way past the question of whether companies are “doing AI”, and the flashlight is now on whether they can still account for what it is doing to the business.
“A wrong AI bet is buying excitement instead of value.”
Benny Czarny, CEO at OPSWAT
The Wrong Bet
Asked what a wrong AI bet looks like in practice, Douetteau immediately widened the frame. The mistake, he said, is usually not choosing the “loser” in the model wars.
“In practice, the wrong bet is more boring and more expensive.”
It is committing early to one provider’s framework, building agents and operating practices around its primitives, and then discovering — sometimes only 18 months later — that the commercial terms, the model quality and even the regulatory environment have all shifted.
What makes that problem so stubborn is that it lives deeper than vendor choice. “They discover the bet was never about the vendor. It was about the operational architecture they built to use the vendor, and they cannot walk away from that architecture without rebuilding the company.”
Pleiter makes a similar point from the banking side, though in more operational terms. The wrong bet, he said, is buying AI capability without investing in AI-readiness. Banks may sign a deal, get a compelling demo and even move quickly into pilot mode, only to find that the model can see only part of the customer, cannot act across channels and cannot produce any evidence trail regulators will accept. In that world, the problem is rarely the model itself. It is the fragmented foundation beneath it. “When you deploy AI on top of fragmented systems, your fraud model reasons over one slice of customer data while your recommendation engine has a different slice.”
ALSO READ: What “High-Risk AI” Actually Means for the Teams Running HR, Finance and Customer Ops
Czarny’s version is the most bluntly commercial. A wrong AI bet, he said, is buying excitement instead of value. In cybersecurity, that can mean embedding a large general-purpose model into a product because it sounds advanced, when the use case really demands deterministic, high-confidence detection. It can also mean dressing up a product with a single chatbot or summary feature and calling that transformation. Or it can mean forcing AI into the business because headlines suggest it should reduce headcount, rather than because it solves a real operational problem.
Taken together, the three answers describe a shift in the way CEOs now think about AI risk. The concern is moving from model selection to organisational consequence: from which tool to which dependency.
“Most banks don’t have that line of sight today. Around 60% of frontline work happens in the handoffs and the whitespace between systems.”
Jouk Pleiter, CEO at Backbase
The Visibility Problem
If the wrong bet is one side of the story, governance is the other. All three CEOs, in different ways, returned to the same frustration: AI is moving quickly through organisations that still struggle to see it clearly enough.
Douetteau’s answer was the most self-aware. Dataiku, he said, uses its own platform internally to build, deploy and govern AI systems, and that experience feeds directly into the product roadmap. But even with that, the pace of experimentation is moving faster than ever. Governance, he suggested, has to keep up with a level of internal build activity that is itself a sign of success. “If a CEO tells you they have full line of sight on every AI system in production, the more likely explanation is that their AI program is less ambitious than it needs to be. That is a real problem, a quieter one than the governance problem but no less serious.”
Pleiter takes a different view. In banking, he said, line of sight is fundamental. If you cannot see what AI is doing, you cannot govern it. And if you cannot govern it, you cannot scale it. Backbase has built a control layer around that idea, with a decision token attached to each AI action so banks can see what policy was applied, what data was used, what the model version was and what outcome followed. That matters because the speed of AI in a bank changes the nature of governance itself. Committee-based oversight is too slow for systems making hundreds or thousands of decisions a day.
Czarny’s answer is the broadest, and in some ways the most alarming. In his experience, many organisations know they are using AI, but far fewer can answer basic questions about where it is used, what data it touches, which decisions it influences or whether the output can be audited. That gap matters even more in critical infrastructure and government environments, where data sovereignty, access control and local deployment can all shape the architecture. His point is simple: “Governance cannot be a document. It has to be part of the system.”
ALSO READ: From Siemens Energy to Bank of America: What “Quietly Advanced” Enterprises are Doing Differently
Dataiku’s recent survey of 900 CEOs worldwide showed 70% are the primary driver of AI strategy, yet only 60% are involved in more than half of AI-related decisions. The same study found that 80% believe their role is at risk if AI fails to deliver measurable business gains by the end of 2026. The numbers are useful, but the conversations give them texture: CEOs are being held accountable for systems that are increasingly distributed across teams, platform providers and operating layers.
“CEOs of public companies have an exposure that runs in quarters and analyst calls. As a founder, there is no exit option.”
Florian Douetteau, CEO and Co-founder of Dataiku
When AI Goes Wrong
Accountability is where the three responses begin to diverge more sharply by industry.
In banking, regulators expect clear accountability chains when something goes wrong. When agents in a platform make a recommendation that harms a customer, who should be accountable – the bank, the platform provider, or both? Pleiter was unequivocal.
“Both. But if we zoom out, the banks will be the ones the public holds responsible.”
The platform provider may share the technical burden, but the licensed entity is the one people trust with their money. The bank chose to deploy the agent. The bank set the autonomy level. The bank accepted the risk.
That answer is also a statement about what banks now need from their technology partners. The expectation is not merely that platform providers offer capability, but that they help banks build the evidence trail and decision structure required to defend AI use in front of regulators, boards and customers. The pressure on banks to modernise is real, but so is the burden of proving that the system is safe enough to trust.
ALSO READ: Can AI Really Earn a Seat at the Supplier Negotiation Table?
Czarny’s framing is more cultural, but no less direct. AI should not be something a CEO delegates and revisits once a quarter, he said. It is already shaping product, security, cost structure, sales productivity, hiring and company valuation. He captures the tension between enthusiasm and discipline inside a company trying to move quickly without losing control: “I like to think about it in two groups: the AI saints and the AI vampires.”
“The AI saints are the engineers and leaders who are truly leaning in. They are learning the latest agentic tools, experimenting with Claude, Copilot, local models, automation, test generation, code review, and new ways to ship faster without sacrificing quality. They are not waiting for permission. They are curious, practical, and obsessed with making the company better.
“The AI vampires are more complicated. They are not always people who reject AI. Sometimes they are people who are very excited about AI, which is good, but they push it everywhere before there is enough discipline around it.
“Some AI vampires can become AI vigilantes. They move too fast without enough testing. They may use AI as an excuse to cut too many people too quickly. They may kill products before understanding the customer impact. They may replace judgment with tools. They confuse speed with strategy.
That is where leadership matters, he says. The right culture is not anti-AI. It is also not blind AI. The right culture is disciplined AI.
ALSO READ: This Startup Went from a Team of 20 to 6. Yet, Humans are their Most Valued Asset.
Douetteau, by contrast, is thinking about exposure as a founder, which gives his answer a different weight. “CEOs of public companies have an exposure that runs in quarters and analyst calls. They can deliver AI outcomes or not, and the board will eventually ask.”
As a founder, he said, there is no exit option. He is making a bet on a broader thesis: that enterprises will need an orchestration and governance layer above the model providers and hyperscalers. If that thesis is wrong, the problem does not disappear with a missed quarter. It becomes existential.
That is one of the more telling differences in this story. The language of risk is universal, but the shape of exposure depends on where a CEO sits in the company. Founders are betting their long-term belief in the architecture. Operating CEOs are being judged on execution. In both cases, the accountability is real.
Why This Matters Now
The reason these answers land is that they reflect a wider shift already visible in the numbers. Dataiku’s survey found that 65% of CEOs now worry more about over-investing in the wrong AI providers than under-investing overall, a reversal of the old fear that companies would be left behind if they moved too slowly. Fifty-seven per cent believe insufficient AI explainability could trigger a customer trust or brand credibility crisis. Put alongside the comments from Douetteau, Pleiter and Czarny, the message is clear enough: AI is now being judged on whether it can be governed, explained and defended under pressure.
That is what makes this stage of the AI conversation different from the last one. The question now is who has built enough architectural discipline around AI to know what happens when the system starts making meaningful decisions.
The most interesting part of all these conversations is that none of the CEOs sounded defensive. They sounded alert. They were comfortable talking about the limits of current governance, the fragility of AI-readiness and the risk of expensive wrong turns. That may be the clearest sign yet that the AI maturity curve in enterprise has changed shape. The conversation has moved from enthusiasm to accountability, but the executives at the centre of it are still trying to work out what accountability should actually look like.
And that may be the real story: whether CEOs trust the organisations around AI enough to stand behind the decisions it makes.
ALSO READ: Designing AI‑Ready Public Infrastructure: Global Lessons from India’s Aadhaar Builder
