A cook in a busy kitchen has three different ways an assistant can help. The assistant can hand them ingredients on request. The assistant can offer suggestions while the cook is improvising. Or the assistant can take over a station entirely and produce dishes while the cook focuses on something else.
All three are useful. None of them is the same kind of help. And a kitchen that does not know which mode it is operating in produces chaos, regardless of how skilled the assistant is.
This is the situation most organizations find themselves in with AI right now. They have a powerful assistant. They have not decided what kind of help they want from it. And the cognitive cost of that confusion shows up everywhere.
Three modes, three different bargains
AI can play three different roles in human work. Each one comes with a different trade-off. Understanding the differences is the start of using AI well.
Mode one: AI as a tool. The human is doing the thinking. The AI is doing a specific task on request. Summarize this email. Draft a first version of this paragraph. Find me the three most relevant data points in this file. The human stays in charge of the decision. The AI just speeds up an execution step.
Mode two: AI as an advisor. The human is still making the decision, but the AI is shaping the thinking. Recommend three options. Walk me through the trade-offs. Point out what I might be missing. The human is no longer just receiving output. They are receiving counsel. The AI is influencing what the human considers, not just what the human produces.
Mode three: AI as a decision-maker. The human has handed the decision itself to the AI. The AI is acting. The human is supervising at most, and sometimes not even that. Auto-approve transactions under a threshold. Auto-route customer service tickets. Auto-decide which candidate to advance to the next round of screening.
The cook can use any of these modes. So can a knowledge worker. The interesting question is which mode they are actually using right now, and whether they chose it on purpose.
The hidden mode shift
Here is what tends to happen. An organization adopts AI in mode one. Everyone is excited. Productivity rises. The use case looks safe.
Then, without anyone deciding to, the same tool starts being used in mode two. Then in mode three. Not all at once. Not officially. Just gradually, as people get more comfortable with the AI and trust its output more. Within months, the same tool that started as a way to speed up execution is now quietly shaping decisions, and sometimes making them.
This is the mode shift, and it is happening in most organizations whether they have noticed or not.
The shift is not bad by itself. The problem is that no one decided. The accountability that should have come with the shift never got designed in. The human capabilities that each mode requires never got built. The cultural conversations about which decisions are appropriate to delegate to AI never happened, because the delegation happened without anyone naming it.
What you end up with is an organization operating in mode three on decisions where it intended mode one, with no plan for either.
The cognitive cost of each mode
Each mode demands a different kind of human capability, and shifts the cognitive load in a different way.
In mode one, the human keeps their thinking muscles in shape. They are still framing the question, evaluating the AI's output, integrating it into their own work. The AI is doing a task. The human is doing the thinking. The cognitive cost is low, and the cognitive capability stays sharp.
In mode two, something subtler happens. The AI is now in the room while the human is making the decision. The AI's recommendations shape the option set. The AI's framing shapes the question. The human is still in charge, formally, but the boundary between human judgment and AI suggestion gets blurry. Whether the decision is really "human" depends entirely on how rigorously the human pushes back on what the AI is saying. Most humans do not push back nearly enough. The cognitive cost of mode two is that judgment slowly outsources itself without anyone noticing.
In mode three, the human is supposed to be supervising. In practice, most humans drift toward rubber-stamping. The AI's decisions come at speed and volume that no human can really audit. The supervisory role collapses. The human becomes a signature on output they did not really evaluate. The cognitive cost of mode three is that judgment exits the system entirely for that class of decision, while the appearance of human oversight remains.
The cognitive cost of mode two is that judgment slowly outsources itself without anyone noticing. The cost of mode three is that judgment exits the system entirely while the appearance of human oversight remains.
How to choose modes on purpose
The healthiest organizations I see are the ones that have started naming the modes explicitly. They have decided which decisions live in which mode. They have built different governance for each. They have invested in the human capability that each mode actually requires.
A few principles seem to help.
Match the mode to the stakes. Low-stakes, high-volume, well-defined decisions can live in mode three with light supervision. Higher-stakes, more context-dependent decisions belong in mode two with serious human counsel. The highest-stakes, most novel decisions stay in mode one, where AI is a tool but humans do the thinking.
Make the mode visible. When you are using AI, name the mode out loud. Are you using it as a tool, an advisor, or a decision-maker? Most people cannot answer this question for their own work. Making it visible is the first step in choosing it on purpose.
Watch for unintended drift. The mode shift is real and it happens by default. Schedule honest reviews. Where did a tool become an advisor? Where did an advisor become a decision-maker? Is that what we wanted, or did it just happen?
Invest in the capability each mode requires. Mode one requires basic AI literacy. Mode two requires sophisticated judgment, the ability to push back on confident-sounding AI recommendations and stay in charge of the decision. Mode three requires designed supervision, not just human signatures. Different modes need different investments.
The question worth asking
Here is a useful exercise. Pick a decision your team or organization made recently with AI in the room. Ask yourself which mode you were actually in. Then ask whether that was the mode you intended.
If the answer is "we were in mode three but we thought we were in mode one," you have just identified a piece of the cognitive cost that is currently invisible in your organization. There is probably a lot of it.
The cook with three kinds of assistants is not facing a technology problem. The cook is facing a clarity problem. They have to decide which assistant is doing what, when, and how the kitchen runs differently as a result. The technology is the easy part. The choice is the hard part.
The same is true for organizations using AI. The technology is here. It works. The question is not whether to use it. The question is which mode you want to use it in, for which decisions, with what supervision, and how you are going to keep the answer from drifting without your permission.
That conversation is one most organizations have not had yet. The ones that have it early will operate very differently from the ones that wait until the drift becomes a crisis.