# Three things every leader must do to hold the line against AI decision-making > Source: > Published: 2026-07-08 16:54:25+00:00 [When the leader of a legal-tech startup began using ChatGPT](https://futurism.com/artificial-intelligence/bosses-obsessed-with-ai), he started by doing what most of us do – embracing [the new technology’s ability](https://www.fastcompany.com/91523806/how-ai-is-quietly-exhausting-you-and-what-to-do-about-it) to take over some of the smaller, more [tedious parts of the job](https://www.fastcompany.com/91557241/ai-is-making-work-better-it-may-also-be-making-people-even-lonelier-ai-work-mental-health-loneliness). But within a few months, the [AI](https://www.fastcompany.com/section/artificial-intelligence) model had gone from an assistant that helped with emails to the [highest source of authority](https://www.fastcompany.com/91421353/what-happens-when-ai-does-everything-your-ceo-used-to-do) in the company. The first sign that the train was running off the rails came when he instructed his staff that they were to consult the AI before every meeting to discuss their ideas. Then he started making structural choices about the company based on conversations with the chatbot, [hiring](https://www.fastcompany.com/section/hiring) and firing on the AI’s say-so and pivoting the firm’s entire focus from one practice area to another based on whatever the model had told him that week. The last straw came when he used the AI to generate a document he called “The Bible,” a constantly changing handbook that staff were expected to study so that they would never have to ask a fellow human a question ever again. It is tempting to file this kind of story under “tales of office madness” and move on. But milder versions can be found in an increasing number of workplaces. In ordinary companies around the world, managers who are nobody’s idea of a crank are handing over parts of their thinking to AI models. In some cases, the cognitive offloading remains limited to mechanical tasks or to activities that don’t require deep human judgment. But in other cases, leaders are reaching much too quickly for the ease an AI answer can give. A considerable amount of ink[ has been spilt](https://www.fastcompany.com/91417456/danger-ai-dependency) over the last few years on the important question of how, and to what extent, AI poses a threat to its users’ hard-won skills. But this is not the whole of the danger, and for those who run organizations it may not even be the more important part. The deeper threat is that, even if we keep the skill, we may lose the will to use it. It helps to be clear about what AI models are and are not good at. A frontier model has absorbed something close to the sum of recorded human knowledge: the sciences, the great works of art, the accumulated wisdom of the world’s business thinkers. This makes it extraordinarily powerful. But the power is rooted in generally accessible knowledge. What large language models have far less access to is the view from inside specific human situations. You can give a model context, and a good one will make intelligent use of it. But it is often the case that the most important things you know about your business have never been written down. They are not in the training data and never will be, because they live in your individual experience: the trade-offs you would accept and the ones you would refuse; how a decision will actually land with this board and these customers; the thing you know to be true but have never quite put into words. That knowledge is yours and applying it is what turns an AI model’s generic competence into something genuinely useful to you. Strip it out, and what you are left with is fluent, plausible, and more or less identical to what the same prompt would have produced for any of your competitors. This is why judgment is so hard to delegate to a machine, and also why delegating it is so tempting. The model’s answer arrives fast, reads well, and asks nothing of you. Giving in to it without reflecting seriously on whether or not it is right is a failure of will in its purest form. The dangers of cognitive offloading apply to everyone in the workforce. But they take on an additional dimension for those in leadership positions: leaders are not only at risk of surrendering their own judgment to the machine; their decisions can force whole organizations to make the same dangerous move. And leaders are particularly vulnerable to the forces that can lead to poor decision-making in this area. Aaron Levie, the founder and CEO of Box, put his finger on the issue recently when he argued that[ senior leaders are “uniquely prone”](https://x.com/levie/status/2058582370253701432) to overestimating what AI can do. The reason for this, he suggested, is because they tend to sit so far from the actual work. Leaders see the polished demo, the prototype, or the generated contract, and conclude that the job is done. For many, the long tail of effort that stands between the impressive first output and the visually similar but much more tightly engineered final result is invisible. It is precisely the same distance that can make leaders overestimate AI capabilities that can also blind them to what they are taking away when they insist that their workers use AI for certain tasks. This is how the individual problem becomes an organizational one. A leader who has stopped exercising his or her own judgment will tend to build a company that does not ask anyone else to exercise theirs either. Worse, the company may even start seeing human judgment as a point of friction to be minimized. At one startup,[ a sales strategist watched](https://futurism.com/artificial-intelligence/bosses-obsessed-with-ai) as his founder first handed his own judgment to an AI model and then insisted that everyone else set aside what they were hearing directly from prospective customers whenever it failed to match the model’s strategic vision. The greatest risk of cognitive offloading at the top of an organization is that it doesn’t stay at the top for long. Protecting judgment across an organization is a matter of design, not demand. You can’t tell people to think for themselves while building a company that rewards them for not bothering. Here are three steps you can take to protect the place of human judgment. The irony of cognitive offloading is that it costs you the most when the technology works best. A model that is fast, fluent, and usually right is exactly the one you stop checking. Each act of deference seems reasonable until you find that your organization has handed over not just its menial thinking but its will to make its own judgments. The technology is not the problem here, and using less of it is not the answer. What matters is keeping hold of the thing it cannot supply – the ability to look at what the machine has produced, decide whether it is right, and own the call yourself.