The GenAI Paradox, Part 2: From Boardroom Hope to Operational Reality
The C-suite is betting on Agentic AI as the next silver bullet. On the ground, agents enter infinite loops, burn through API credits, and Klarna's automation-first strategy reversed course.
This is Part 2 of a five-part series. Read Part 1: The Great Divide | Part 3: The Invisible Workforce Crisis | Part 4: Shadow AI and the Singapore Model | Part 5: Trust, Governance, and What Comes Next
Leadership Sentiment: The View from the C-Suite
The FOMO-Defensive Strategy
Despite elusive ROI, investment continues to surge. This behaviour is driven largely by a “Fear Of Missing Out” and a defensive strategic posture. Deloitte’s 2025 survey confirms that investment is growing, driven by the fear of falling behind competitors rather than clear, calculated returns — 85% of organisations increased AI investment in the past year, and 91% plan to increase further. IBM’s 2025 CEO Study found that 61% of CEOs believe competitive advantage depends on having the most advanced Generative AI, creating a self-reinforcing cycle of investment despite the lack of immediate payoff.
What this FOMO misreads is where the scarcity has gone. For two centuries, value flowed to whatever could be made explicit, codified, and scaled — and AI is the final step in that logic: it commoditises the most explicit artifact there is, software itself. The boardroom is racing to buy more of the very thing that is collapsing toward free. The advantage no longer lives in having the most advanced model; every competitor will have the same one next quarter. It lives in the capacity AI cannot supply — the accountable, judgment-laden human who decides what the model is for and owns what it does. The 61% are bidding up the surface. The moat has quietly inverted to the substrate underneath it.
The Shift to Agentic AI as the New Hope
As the limitations of chat-based GenAI interfaces become apparent, leadership attention has pivoted aggressively toward “Agentic AI.”
The Promise. Agents promise to move beyond content generation to task execution — planning, reasoning, and acting autonomously to achieve business goals. This represents a shift from “chatting with data” to “acting on data.”
Adoption Rates. McKinsey reports that 62% of organisations are experimenting with AI agents, though only 23% have begun scaling them.
The “Silver Bullet” Trap. There is a growing danger that leaders view Agentic AI as a silver bullet to solve the productivity stagnation of GenAI. Experts warn that chasing agentic capabilities without fixing underlying data fragmentation and process issues will lead to “sterile, less differentiated experiences” and further pilot failures.
The deeper error is to assume that acting on data removes the human from the loop. It does the opposite. When the model only suggested, a human still orchestrated the workflow and executed each step, and a wrong answer cost a re-prompt. When the model acts — planning, calling tools, shipping outcomes across many steps — the scarce human work does not disappear; it relocates upward and concentrates into three things at once: architecting the autonomous process (which tools, which guardrails, which hand-offs), verifying its compounding output, and owning what an autonomous system did in the company’s name. The agentic regime does not retire the human judgment leaders hoped to automate away. It raises the price of getting it wrong — because the cost of a confidently-wrong autonomous action is categorically higher than a confidently-wrong chatbot answer. The 62% are not buying their way past the judgment problem. They are buying a more expensive version of it.
The Trust and Governance Dilemma
Trust remains a critical bottleneck. Fewer people believe AI companies will safeguard their data, and concerns regarding fairness and bias persist. For executives, the challenge is balancing the pressure to innovate with the rising tide of AI-related incidents and regulatory scrutiny. The “New Gavel” effect predicts that executives will increasingly be held personally responsible for rogue AI actions, shifting AI risk from an IT problem to a board-level liability.
This is not a temporary regulatory inconvenience; it is structural, and it explains why governance is a different kind of work from operation. No jurisdiction accepts “the AI decided.” Accountability — the capacity to own a consequence and answer for it when it goes wrong — cannot be transferred to a system. It is one of the few things that becomes more valuable as machine output grows, because every autonomous action still needs a human whose name is on it. That is why we model AI capability not as one ladder but as four distinct accountabilities — those who build it, those who operate it, those who assure it, and those who direct it. The “New Gavel” lands squarely on the last two: the leaders who set the autonomy policy and the governance specialists who own the controls. An organisation can buy the model. It cannot buy the answerability, and a board that confuses the two is the one the gavel finds first.
The Operational Reality: Agentic AI and Technical Failure Modes
While the C-suite envisions a future of autonomous agents, the operational reality on the ground in 2025 is fraught with technical fragility and integration nightmares.
The Reality of Agent Performance
Agentic AI pilots in 2025 have revealed significant limitations in the technology’s readiness for complex, high-stakes enterprise environments.
Infinite Loops and Cost Spirals. In technical communities, developers report agents entering “infinite loops,” where an agent attempts to fix an error, fails, retries, and burns through significant API credits without human intervention.
Memory and Context Deficits. Agents often lack robust long-term memory. Without the ability to retain context over long execution horizons, they struggle to learn from mistakes or adapt to specific user preferences, rendering them “stateless” and inefficient for continuous workflows.
The “95% Recall” Trap. Chaining multiple stochastic components — where each step has a less than 100% success rate — leads to compounding errors. If an agentic workflow requires five steps, each with 95% accuracy, the total system reliability drops to roughly 77%. In enterprise processes requiring near-perfect execution (financial reconciliation, compliance reporting), this unreliability is unacceptable. This compounding error problem is what we identified in Sovereign Command as a core architectural challenge — and why governance cannot be an afterthought bolted onto autonomous systems.
It is worth being precise about why this fails, because the fix is not “a better model.” The agent’s language model is probabilistic — the same input can yield different output. Traditional software is deterministic — same input, same output, every time. The compounding-error math is simply what happens when you build a process out of probabilistic links and expect deterministic reliability from the chain. The mature response is not to wait for a model that never errs; it is to wrap the probabilistic core in a deterministic harness — explicit verification, guardrails, and structured control that the organisation designs and owns. In our terms, the harness is the roughly 98% of the system you engineer and can defend; the model is the 2% you rent. Put another way: the Institutional Vault — the codified rules, processes, and context an enterprise owns — can execute the judgment a human has already encoded into deterministic logic. What it must never do is judge the un-codified call on the model’s own probabilistic say-so, then close the loop with no one able to answer for it. The Vault can decide; it must never judge. Reliability in financial reconciliation does not come from trusting the agent. It comes from a human-built harness around it that refuses to let an unverified probabilistic step stand.
Klarna: A Cautionary Tale
The experience of Swedish fintech Klarna serves as a primary case study for the risks of aggressive AI substitution.
In February 2024, Klarna announced that its AI customer service assistant was handling the equivalent workload of 700 outsourced agents. Separately, the company reported saving $10 million in marketing costs by using AI for image generation instead of agencies. The customer service AI was projected to drive a $40 million annual profit improvement.
Then came the reversal.
Quality Degradation. CEO Sebastian Siemiatkowski admitted that “cost unfortunately seems to have been a too-predominant evaluation factor,” resulting in “lower quality” outcomes. Customers complained about interactions with what they described as “slop-spinning algorithms.” The word slop is diagnostic: it is what fluent, confident, plausible-but-wrong output looks like to the person on the receiving end. The failure was not that the AI was incapable — it was that the organisation extended it unearned trust, deploying it where no human was doing the work of judging whether the output was actually any good.
The Pivot Back to Humans. By May 2025, Klarna initiated a recruitment drive to bring humans back into the loop, ending a year-long hiring freeze. Siemiatkowski described the new model as an “Uber-type setup” targeting students and rural workers for flexible, remote customer service roles — a human fallback to ensure quality and empathy.
The Klarna trajectory — from triumphant AI-first announcement to public quality admission to rehiring drive in barely twelve months — is not an indictment of AI in customer service. It is an indictment of automation strategies that optimise for cost without governance architecture to maintain quality. Klarna let the codified store judge — handing the un-codifiable call (does this customer feel heard, is this answer actually right) to a probabilistic system and closing the loop with no human able to catch the drift. That is the line the Vault must never cross. The rehiring drive was the company buying back the judgment and the accountability it had tried to automate away — and the “Uber-type setup” of flexible human agents is, read generously, an attempt to rebuild the human layer the cost-first strategy had hollowed out. The lesson is not “humans beat AI.” It is that the capacity to judge AI’s output is precisely the capacity you cannot delete from the loop — and the expensive way to learn that is Klarna’s.
Vendor Fragmentation and Data Moats
Implementation is further hampered by a fragmented vendor ecosystem. Major players like Salesforce, Microsoft, and Workday are building “walled gardens,” creating interoperability challenges. CIOs report that agents built in one ecosystem cannot effectively interact with workflows in another, preventing the vision of a unified, autonomous enterprise.
The strategic risk hides inside this fragmentation. The model layer is commoditising and will be backend-agnostic; the durable asset is the organisation’s own codified knowledge — its rules, processes, examples and context. When that asset is poured into a vendor’s walled garden, the moat the company is building accrues to the vendor, not to itself. The defensible move is to treat the codified store as something the enterprise owns and can carry across backends, and to build the in-house capacity to govern it — because a capability you can buy off the shelf is, by definition, available to every competitor. Adoption can be procured; the capacity to evolve an institution around AI has to be built inside it. That is the difference between renting an autonomous enterprise and owning one.
About the Author
Ethan Seow is a Centre for AI Leadership Co-Founder and Cybersecurity Expert. He is ISACA Singapore’s 2023 Infosec Leader, ISC2 2023 APAC Rising Star Professional in Cybersecurity, TEDx and Black Hat Asia speaker, educator, culture hacker and entrepreneur with over 13 years in entrepreneurship, training and education.
This is Part 2 of a five-part series. Continue to Part 3: The Invisible Workforce Crisis.