There already has been a robust conversation about the new cost of AI, literally. Most recently since Anthropic raised its enterprise rates, but really, even before that, the cost of AI financially, let alone emotionally and structurally. But the new headache is the difficulty of managing token use, and undoubtedly that is only one of the many headaches to come with AI at the enterprise level. Not because of any gloom and doom, just the growing pains of organisational level adoption of new technology.
But for better or worse, the concern about tokens and other financial constraints of AI may be a symptom. A serious one, but still a lagging indicator and possibly not the real problem. Instead, the high usage may be that AI has shifted the dynamic of organization processes in ways we are still digesting.
To be clear, I use AI daily, I think it is equal parts powerful and useful. I’ve spent years of my professional life in the technology sector and if anything, I am probably biased towards being more bullish on that type of innovation. The counterbalance is my understanding of human psychology, which includes noticing that AI might be both aiding us in one way while unwittingly removing some natural human guardrails. We don’t always know what collateral benefits we enjoy until the primary reason for something is removed. We didn’t know that cell phones would let us be endlessly contactable…and remove the built in respite that came with not being endlessly contactable.
The cost of tokens is going up. But the real cost might come earlier.
In a well-run company, senior leaders handle exceptions, not the routine. The hiring freeze is the example. The org institutes a freeze. The CFO exists to override it in the rare, exceptional case — a critical role that has to be filled despite the freeze. What the CFO should not be doing is approving every individual hire. That’s the healthy pattern: the rule handles the volume, the senior person handles only the exceptions to the rule.
The failure mode is not new: in companies that grow too fast, that discipline breaks. (A lot. I’ve seen it.) Decisions that should be settled two levels down start floating up, and senior leaders end up making decisions about things that should never have reached them in the first place. Soon the executive team is making decisions that consume them due to volume not by criticality. They are making routine sign-offs because they can, but they may underestimate the multiple benefits of people further downstream making them.
The number of decisions requiring their approval exceeds the hours they have in the room together. Why? A couple of things: AI makes leaders feel they can make decisions faster than the prior model, which is decisions made further down the chain, so they are making less critical decisions than they should. The other issue is that, in general, AI multiplies the volume of analyses and options flowing upward, so without governance, that same floating-to-the-top failure happens faster. Everyone is armed with more options via AI than they should. The natural limits of human capacity were, frankly, doing more work than we realized.
The bad thing about humans is they don’t have endless capacity, so we can’t generate endless ideas. The good thing about humans is… we don’t have endless capacity. In the past that capped the research and planning stage. People ran out of ideas. (It all sounds so quaint now, doesn’t it?)
But it turns out that cap might have been a protection built in that saved us from a fatigue that we never tested. And related, when people ran out of their own ideas, they asked other humans to weigh in. Those humans act as a QA, pointing out the inconsistencies and the blind spots. That matters because every human has a limited perspective and is prone to confirmation bias based on those exact limits. Another human widens the lens. They bring a distinct perspective, and by definition that perspective is unknowable to the first human. Which means it’s also unknowable to the AI that they use, because everyone’s AI inherits our bias.
Leaders can make more down-the-line decisions now with AI. They probably shouldn’t.
So the new discipline is to hold the line. Taking more decisions up the chain was always inefficient. Now it’s also shedding human-based QA, diversity of thought, and the role that junior members of teams have held. The entry-level work wasn’t just for them to learn — it was also because they did the first careful pass at establishing the information that fed up to the decision-makers. Taking that from them depletes their own development, just as lower-level decision-making taken from junior managers depletes their decision-making development.
So it’s important to learn to keep that decision-making efficiency, because it’s better for the org and it reduces the risk of fatigue, capacity limit, token depletion by teatime. Because it both reduces actual capacity and increases poorer well-being — because we sense it. It’s disorienting. We start chasing a solution to overwhelm or brain fog when it’s not a new problem. We just removed an old guardrail.
