Summary
The most consequential use of artificial intelligence (AI) may happen before anyone automates a workflow. It happens when a person begins thinking with it.
A leader uses AI to pressure-test a strategy before a meeting. A marketer asks it to find weaknesses in an argument. An analyst asks it to identify patterns across research, and an executive works through several interpretations of an ambiguous problem before deciding.
After enough repetition, AI can become part of how that person approaches difficult work. The change is subtle, because the human still owns the laptop, attends the meetings, and makes the final call. Part of their cognitive process now runs through interaction with an external system.
That can expand human capability. It can also change which capabilities people practice, where judgment develops, and what happens when the system disappears. Organizations adopting AI need to understand both sides at once, because AI can raise the cognitive capacity available to a person while quietly reshaping the human capability underneath it. That makes personal AI use an organizational design question, not only a productivity one.
AI Is Moving Up the Cognitive Stack
Early workplace AI use concentrated on production: write this email, summarize this document, reformat this report. Those uses remain valuable, and they leave most of the underlying thinking process intact.
More sophisticated users engage AI earlier in the work. They ask what they are missing, what assumptions they are making, and what would make a strategy fail. They ask AI to find contradictions across documents, generate alternative explanations, and challenge their reasoning. At that point, AI is participating in problem definition, synthesis, and evaluation, activities that sit much closer to judgment than to drafting.
Microsoft Research supplies direct evidence of the shift. A 2025 study of 319 knowledge workers, conducted with Carnegie Mellon University and drawing on 936 real-world examples of generative AI use, found that higher confidence in generative AI was associated with less critical thinking, while higher confidence in one’s own ability was associated with more. The same research found that AI does not simply remove thinking; it shifts the nature of critical thinking toward information verification, response integration, and task stewardship. OECDOECD
AI redistributes thinking rather than eliminating it. The organizational question is whether that new distribution produces stronger human judgment or merely a faster path to a plausible-looking answer.
Cognitive Capacity and Cognitive Capability Are Different
This distinction matters more than it first appears. Cognitive capacity describes how much usable attention and mental effort a person has available. Cognitive capability describes what the person can actually do with it.
AI can help with the first. A leader who once spent two hours synthesizing 20 documents might get a useful first synthesis in minutes, and the saved capacity can move toward testing assumptions, speaking with customers, or making a sharper decision. That is a genuine gain.
A different outcome is just as possible. The leader stops reading much of the source material, reviews the AI’s summary, and moves on. The immediate task gets easier, but over time the leader may encounter fewer chances to notice weak evidence or build firsthand familiarity with the subject. The organization has gained capacity. Whether it has preserved capability is a separate question, and one most AI rollouts never ask.
Sterling Phoenix already treats cognitive capacity as an organizational concern through the Cognitive Operating System, registered as the practical operating system for protecting and allocating human cognitive capacity, and through Leadership Cognitive Throughput, which addresses the finite cognitive capacity of leaders as complexity and decision volume increase. AI adds one new question to that existing work: what should people do with the capacity AI releases? That may matter more than how many hours AI saves.
Thinking With AI Can Make a Skilled Person More Capable
The upside deserves equal weight. Experienced people can use AI to extend the reach of expertise they already have. A knowledgeable executive can test more scenarios before committing resources. A strategist can compare more competing explanations, and a subject-matter expert can interrogate a larger body of information than they could alone.
The human still supplies something the model does not: responsibility for the outcome, along with organizational context, professional judgment, and knowledge of consequences. AI expands the analytical surface around those capabilities rather than replacing them.
This is why a productivity framing alone falls short. If an experienced leader uses AI to evaluate five plausible interpretations instead of one, the value shows up as better judgment rather than faster output, and that benefit rarely fits neatly into a time-saved calculation.
Expertise Changes the Quality of the Collaboration
AI does not erase the value of expertise. In many situations, expertise determines whether someone can recognize a weak AI answer in the first place. A novice and an expert can receive the identical confident recommendation: the expert notices that an assumption violates operational reality, while the novice sees only polished reasoning.
Recent research on cognitive offloading sharpens this concern. A 2026 study published in Frontiers in Psychology distinguishes dependent offloading, where a person delegates core thinking to AI, from autonomous offloading, where AI scaffolds thinking the person still directs, and finds that the manner of AI use, not merely its frequency, determines the downstream effect on the person’s own cognitive development.
That distinction produces a tension no adoption metric can resolve on its own. AI can make experienced people more capable, and the same system can make it easier for inexperienced people to produce work that only appears experienced. One amplifies existing judgment. The other can conceal its absence, and the two can look identical from the outside until something goes wrong.
The Skill Risk Appears When AI Removes the Practice That Built the Skill
Expertise usually develops through repeated exposure to problems: people attempt tasks, make mistakes, encounter exceptions, and refine their mental models through feedback. Some work is inefficient precisely because someone is still learning how to do it, and automation can remove that work along with the inefficiency.
Consider a junior analyst who historically spent hours reviewing source material, building comparisons, and preparing a recommendation for a senior colleague. AI can now perform large parts of that process much faster. The efficiency gain is obvious, and the developmental consequence is harder to measure. If the analyst receives a completed synthesis and checks it for errors, they are performing a different cognitive task than constructing the synthesis themselves, and the organization may be trading tomorrow’s expertise pipeline for today’s throughput without ever deciding to make that trade.
Review Does Not Automatically Preserve Judgment
Many AI policies rely on one safeguard: a human reviews the output. That requirement sounds reassuring and says little on its own about whether the human can perform a meaningful review. Real review requires enough knowledge to detect a problem, enough attention and time to look closely, and enough independent judgment to disagree with a plausible, machine-generated answer.
Sterling Phoenix’s Organizational Judgment System exists for exactly this problem. The registry defines it as the architecture for building, distributing, challenging, preserving, and renewing the human judgment that AI-enabled work requires, through components including the Judgment Demand Map, Challenge Architecture, Exception Library, and Judgment Debt. AI-supported thinking belongs partly inside that architecture, because the organization has to preserve the capability required to supervise the system it has deployed. Otherwise, “human review” becomes a formality rather than a control.
Personalized AI Deepens the Relationship
The previous article in this series examined why one person’s AI can know more about their work than someone else’s. Persistent instructions, accumulated conversations, project material, and connected knowledge make an AI environment increasingly useful to a particular person, and that accumulation changes cognitive collaboration too.
A deeply contextualized AI requires less explanation before useful work can begin. The user can start farther into the problem, refer back to earlier reasoning, and compare current thinking against prior decisions. Humans have always used external cognitive infrastructure, from notes and spreadsheets to colleagues and databases. AI differs because the external system can now respond, synthesize, challenge, and generate alternatives on its own, which makes the relationship more dynamic and the dependency question harder to ignore.
What Can the Person Still Do Without the AI?
This is one of the most useful tests an organization can apply. Remove the AI temporarily and ask what remains. Can the person still recognize a weak argument, reconstruct the important logic, or identify when evidence is missing? Can they challenge an AI recommendation, teach someone else how the work should be done, or explain why a final decision was made?
The answer does not need to be everything. Organizations already depend on technology in ways few employees could reproduce manually; few executives could rebuild their enterprise resource planning system with paper and pencil, and that dependency is not irresponsible on its own. The real issue is minimum viable human capability: which capabilities must remain strong because they provide judgment, recovery, accountability, or continuity if the system fails or changes.
Sterling Phoenix’s Organizational Capability Resilience addresses this territory directly, built to preserve and renew human, knowledge, process, decision, and technology capability as AI-enabled work changes, through components including Minimum Viable Independence, Capability Continuity Envelope, Capability Debt, and Recovery Architecture. Thinking infrastructure gives that logic one more place to apply.
Dependency Can Accumulate Quietly
AI dependency can be hard to notice because it develops one task at a time. First the AI drafts, then it summarizes, then it researches and compares, then it proposes options and critiques them. Eventually the human’s role narrows to selecting among AI-produced possibilities.
That arrangement may be entirely appropriate for the work involved. It may also mean the organization has changed a person’s job without ever deciding to redesign it. The employee now performs more orchestration, verification, and exception handling, while the old capabilities receive less practice than they once did. AI implementation should therefore include work design: knowing what cognition has moved to the system and what remains with the person is not a detail, it is the design itself.
The Human Work Often Gets Harder, Not Easier
Removing routine cognitive work does not necessarily make the remaining work easier. AI handles predictable cases, so humans inherit the ambiguous ones. AI completes straightforward analysis, so humans investigate the anomalies it cannot resolve. AI generates recommendations, so humans decide whether those recommendations deserve trust, often at higher volume than before.
This concentrates human work around judgment, which is precisely the relationship Sterling Phoenix’s architecture already recognizes between the Organizational Judgment System, the AI-Era Leadership Capacity System, and Leadership Cognitive Throughput, the last of which specifically protects finite leadership cognitive capacity as complexity and decision volume increase. AI can reduce cognitive load at one layer of the organization while increasing it somewhere else, and that redistribution needs to be designed rather than discovered after the fact.
AI Fluency Should Include Knowing When to Think Without AI
Most organizations teach employees how to use AI. Mature organizations will also need to teach people when not to. Independent thinking can establish a person’s initial judgment before the model shapes the framing, test whether someone actually understands the problem, and preserve practice in capabilities the organization considers essential.
This does not require rejecting AI. It requires using independence on purpose. A leader might form a preliminary assessment before asking AI for alternatives. A junior analyst might complete part of a problem manually before comparing their work against the model’s. A team might periodically run an exercise without AI to test what capability remains. The right design depends entirely on which capability the organization actually needs to keep.
Five Questions for AI-Supported Thinking
Leaders do not need to monitor every employee’s thought process, but they do need to understand how AI is changing cognition in roles that matter. Five questions provide a starting point.
- What cognitive work has moved to AI? Look past tasks to interpretation, synthesis, comparison, and problem framing.
- What cognitive work remains human? Determine whether people still own meaningful judgment or mainly approve AI-generated reasoning.
- Which human capabilities must remain strong? Identify the expertise required for review, exception handling, and recovery, and protect its opportunities for practice.
- Where is AI increasing cognitive demand? Count verification, escalation, and exception work, since time saved upstream often reappears as burden downstream.
- How would we know human capability was declining? Waiting for a major failure is an expensive way to find out; error detection, review quality, and performance when AI is unavailable are cheaper signals.
These questions move the discussion from AI adoption toward organizational capability, which is the harder and more durable question underneath it.
The Best Outcome Is Cognitive Leverage, Not Maximum Use
The objective should not be maximum AI use. It should be stronger combined performance from the human and the system together. Sometimes AI should perform nearly the entire cognitive task. Sometimes it should provide evidence, generate alternatives, or challenge a human conclusion, and sometimes the person should think first and consult AI second. The right arrangement depends on consequences, expertise, developmental needs, and how much independence the organization requires. That is work design, and it is also capability strategy.
Leaders Need to Decide What Kind of Organization AI Is Creating
AI can make individual employees extraordinarily capable. A senior professional with deep expertise, strong judgment, and sophisticated AI use can process more information and explore more possibilities than the same person working alone. Organizations should want that advantage, and they should also ask what is happening underneath it.
Are employees developing stronger judgment because AI exposes them to more perspectives, or are they becoming better only at verification? Are junior employees still developing the capabilities required to become experts? Can people recognize when the system is wrong, and does the organization retain enough independent capability to recover when technology, vendors, or models change? Those questions will matter long after today’s model rankings are forgotten.
AI is becoming part of how people think through work. That development can create real cognitive leverage. The operating responsibility is making sure the human capability required to use that leverage stays strong enough to carry it.
