AI Has a Cost Most Companies Aren’t Accounting for Yet. Here’s What They’re Missing. 

When someone talks about the true cost of AI, most of the words that pop into the head are token spend, GPU costs, model licensing, cost of having outdated infrastructure (and the associated cost of migration), training costs for people, etc. But they leave out one of the most basic and challenging aspects of AI cost.  

I run a data and AI company. I’ve been telling enterprise clients for more than a decade that their AI problem is a data and governance problem, not a technology one (technology is and will always be an enabler). I still believe that. But I’ve started to think our industry skipped a step. 

Let me call it a thinking problem for individuals and a governed change management problem at an enterprise level. Right now, it’s not about whether or not people will accept AI but more like how we ensure that the change is meaningful and that it is also beneficial to us as humans.  

What AI does to your thinking 

Many people remember this research by MIT’s Media Lab on the cognitive debt of using an AI assistant. The findings were and still are of an alarming nature. The report basically showed that people using only ChatGPT showed the least EEG (brain activity), with brain-only participants exhibiting the strongest. It clearly demonstrates what happens to a skill you stop exercising. Even if the skill is as simple as writing an essay. The neural pathways go quiet when you stop using them, and they don’t switch back on command.  

Around the same time, researchers from Microsoft and Carnegie Mellon University published a report on the impact of generative AI on critical thinking. The report labels the shift this way: “The more confidence someone had in the AI’s answer, the less critical thinking they did. The more confidence someone had in their own ability, the more they engaged. Trust in the tool and trust in your own judgment moved in opposite directions.”

It’s an individual-level cost, and it shows up in judgment long before it shows up on a token bill. But now scale this individual cost in small but impactful numbers, and lo and behold, you might have AI confidence due to low self-judgment — which is not the ideal case scenario.  

Regression to the mean, but the mean is a machine. 

2 different ways to look at it — why the mean is a machine

One, when the effort shifts from creation or execution to verification and AI response integration, the mean is essentially the AI baseline that has been created. 

Two, the artificial hive mind, where LLMs independently converge on their outputs with open-ended questions.  

So, I’d argue now that the mean has shifted from a natural average of human ideas to whatever the model considers the most statistically likely response. The more first AI drafts and first AI recommendations we pick, the more likely they’ll nudge their own thinking a little closer to that machine average.  

Do that enough times, across enough people, and the organization quietly outsources its variance, dissent and its weird, unlikely, correct-anyway ideas to a system built to predict the next most probable word. 

Sum of people = enterprise  

This is the most obvious way to state what everyone knows. Once you cascade all the sums of individual choices, it eventually starts converging into a machine mean. Does it mean that as the CEO of a company that works on AI with every breath, I want everyone to stop working on it? Obviously not.  

I’m not walking back what I’ve spent my career arguing. Enterprises need clean data, real governance and operating models built for AI, and most are years behind on all three. It’s just that I have changed my mind about the order of operations. I used to talk about data foundation as the starting point. I now think it’s the second step, not the first. 

Acknowledgement of the change is the zeroth step. Bolting AI onto existing thinking is the enterprise version of accepting the first AI draft. So, the first step is deciding, deliberately, how your people are allowed to think alongside AI. Governance built on top of employees who’ve stopped questioning machine output turns into a checklist that makes leadership feel safe while the judgment underneath it has already eroded. 

What I’d actually do about it 

Here are a few things I’ve started pushing inside my own company and with clients who’ll listen: 

• Have a devil’s advocate or a human disagreement embedded into any process where AI ranks options or drafts a decision. In some cases it’s me, in others it’s the marketing team or the industry leaders — but there is someone who asks “why?” and “why not?” 

• There cannot be an “AI gave this” answer. No AI-generated recommendation goes to a client or the board or to a meeting without someone who wasn’t in the room independently re-deriving the answer before seeing the AI’s version. AI is not accountable; the person making it is. 

• Shipping fast is not rewarded. Trials and tests are rewarded only with progressive improvements and ample quality testing. Then you’ll reward them for editing, challenging or rejecting AI output. 

• Tokens are not a symbol of productivity. You can track them, but don’t essentially equate them with the best work or the most productive performance. 

Now that you’re investing in people, you can invest equally hard in the data foundation and governance layer. Then, it will actually get used by people still capable of questioning it, instead of people who’ve quietly stopped noticing they’ve stopped thinking.