The dominant narrative around AI right now is the speed at which it enables organizations to move. It should also come as no surprise that speed is being conflated with progress while reports abound of the mistakes that AI is making are getting written off as anti-AI rhetoric. The question that matters more, and the one I can’t seem to find a good answer to, is how companies should handle moving fast while allowing for mistakes to happen deliberately.
Researchers have spent enormous amounts of time and money quantifying what happens when AI amplifies bad decisions across an organization, and that work deserves to be taken seriously. So does the pressure C-Suites feel to adapt quickly or get outcompeted, even when fast adoption destroys value along the way.
The companies that get through the next few years in one piece will learn to tell automation-friendly decisions from automation-hostile ones, and build the discipline to act on that difference. Recklessness is the real risk here—not AI.
The pressure to adopt AI has reached a fever pitch. A Boston Consulting Group survey from January puts AI adoption in the top three priorities for CEOs. Companies plan to double their AI spend in 2026, and 94% plan to maintain or increase that spend regardless of whether they see returns this year.
The reasoning is straightforward. In a market that rewards margin expansion through lower per-unit cost without price increases, every company is looking for somewhere to find that math. And the pressure originates from the board. A separate BCG survey found that nearly 40% of CEOs don’t think their boards grasp what AI can actually do, and over half think their boards’ views are warped by market hype. Whether the boards are right or wrong is beside the point. CEOs believe their jobs depend on adopting AI, and that belief is driving their decisions.
The realized productivity gains are thin. In their October 2025 paper, Dritjon Gruda and Brad Aeon, using meta-analyses and systematic reviews rather than anecdotes, challenged what most boards seem to assume. AI produces efficiency gains in some scenarios, but those gains are heavily context-dependent and the outputs come with quality issues. They also found that on open-ended creative tasks, human-AI collaboration underperformed humans or AI working alone.
Anyone familiar with the productivity paradox will recognize Solow’s Paradox. In 1987, the Nobel laureate economist Robert Solow noted that you can see the computer age everywhere except in the productivity numbers. The paradox eventually resolved in the early 1990s, when IT investment finally started showing up as productivity, but the same term gets reused now for digital transformation, and more recently for AI. Whether AI’s returns will eventually show up the same way is an open question. So far the data doesn’t point to a “yes.”
None of this means AI is useless. It means the headline, that AI acceleration equals organizational progress, is not what the data supports. Speed is being mistaken for progress, and right now that mistake is expensive.
Automation amplifying error is older than AI. AI is just running the same playbook at a larger scale and across more domains than previous waves did. The earliest and most dramatic versions came from algorithmic trading, before machine learning, when automated systems were built specifically for speed.
In May 2010, in an already turbulent U.S. equities market, a large mutual fund started executing automated S&P futures hedges set to 9% of the previous minute’s trading volume, with no regard for price or time. Other automated trading systems reacted to the cascade. The combined effect set off a liquidity crisis that erased nearly $1 trillion of market value from the economy over twenty minutes. The regulatory and remediation costs alone were much larger than the $4.1 billion trade that started it.
More recent cases involve decisions that are irreversible or expensive to reverse, with weak or slow feedback loops, and situations where pattern-matching breaks down on novel inputs. The damage in these cases is still being tallied.
Let’s use Klarna as the canonical example. After deploying an OpenAI-powered chat agent in 2024, the company cut headcount by 38%. The agent was handling roughly two-thirds of customer service volume, the equivalent of 700 full-time employees. CEO Sebastian Siemiatkowski projected $40 million in annual savings and pointed to an 82% drop in average call length as proof that AI was going to remake work. He started lobbying policymakers to prepare for the displacement Klarna was demonstrating.
Then the quality issues surfaced. AI handled simple requests well but struggled with multi-step billing disputes, fraud cases, and emotionally charged interactions. Repeat-question volume didn’t drop. Customer satisfaction fell on high-value interactions, which generated negative reviews and churn risk. By early 2026, Klarna decided to reverse course. Their new plan is a hybrid approach where AI triages Tier 1 issues, and assists humans with Tier 2 and 3. But the public messaging about AI replacing the workforce has made recovering headcount harder. Add in the lost institutional knowledge and the cost of training new staff, and the projected $40 million savings has evaporated.
AI-driven speed has improved outcomes in plenty of places. Fraud detection systems process millions of transactions and catch patterns no human reviewer would. Manufacturing QA spots defects faster than the eye. These cases have a shared structure: high volume, low stakes, fast feedback, and humans waiting in the loop for anything ambiguous. The decisions companies need to slow down on are the inverse: irreversible or expensive to reverse, with slow feedback, where novel cases break a model’s pattern-matching.
To be fair, the data and these case studies leave some ambiguity. No company has ever gone bust from adopting AI too slowly. But, as demonstrated in Klarna’s case, they have been burned by adopting too fast. Meanwhile, the competitive pressure continues to mount. Markets reward AI announcements, boards demand strategies, and the long-run upside looks more and more promising.
Block put a sharper version of this in front of investors in February. The company cut more than 4,000 of its roughly 10,000 employees to shift to what they are calling an “intelligence-native” operating model. And wouldn’t you know it? The stock went up 24%. However, early reporting describes plummeting morale, product quality slipping, and high-level executive exits. Whether that one ages well remains to be seen.
The point is: this choice isn’t binary. The question is not whether to adopt AI but how. And specifically whether the adoption process can delineate between decisions that can be safely sped up from decisions where speed itself is the risk. Companies need to build that distinction into their operating model.
Building dedicated capacity to deliberate irreversible decisions is not a new idea. Several industries solved versions of this long before software showed up.
Nuclear aircraft carriers, air traffic control, and power utilities run nearly error-free despite the catastrophic potential built into their work. Karl Weick and Kathleen Sutcliffe, in Managing the Unexpected, identified five principles that make this possible: preoccupation with failure, reluctance to simplify, sensitivity to operations, commitment to resilience, and deference to expertise. That last point is critical. In these organizations, decision authority moves to whoever has the most relevant knowledge during a critical operation, regardless of rank.
Building teams to deliberate over critical automation decisions is just one piece of the answer, however. The other is giving them a working method for that deliberation that doesn’t get bent by political pressure. This, in my opinion, is where “human-in-the-loop” breaks down.
Human-in-the-loop creates automation bias. Humans over-rely on automated recommendation, and AI assessments delivered before the human reaches a conclusion will anchor the human’s judgement, which can be really bad when the AI hallucinates.
So what does a working method look like? Cybersecurity teams run one already: red teams.
Red teaming works as a deliberation method when the organization is choosing what to automate, because it replaces speculation with adversarial, evidence-based prioritization. Automation choices in the SOC carry real consequences, and red teaming forces a structured argument about which functions, once automated, hold up against people trying to break them.
The full set of requirements for these deliberation methods probably warrants its own essay. But the general criteria are clear enough. The method should require multiple reviewers on irreversible decisions, so no single perspective drives the call. The organization has to reward override decisions, so reviewers feel safe making the right call when the stakes demand it. And the architecture has to enforce deliberation at the system level rather than leaning on policy alone to make it happen.
The cost of admission to the AI era is not the wholesale abandonment of judgment, despite what the dominant narrative implies. Companies that treat irreversible decisions as proof-of-concept exercises will keep learning what Klarna learned, and what Block is learning now. Markets reward the announcement, but the bill comes due later.
Leaders don’t have to choose between adopting AI and adopting it carefully. Solow’s Paradox will resolve the way it has before, through the slow, unglamorous work of figuring out which functions actually benefit from automation, which ones only appear to, and which ones quietly accumulate debt that surfaces years later. Sorting that out is the work itself, not a delay tactic.
Boards demanding AI strategies, CEOs whose tenure feels contingent on those strategies, and analysts rewarding announcements over outcomes are all forms of pressure. Pressure isn’t deliberation. The companies that come through this period in good shape will be the ones that built the discipline to know when speed is an asset and when it’s a liability.