Metacognition at Work: How Better Thinking Improves AI-Era Decisions

It is 4:45 p.m., and the strategy memo looks finished.

The argument is clear. The language is polished. The market summary is convincing. An AI tool helped the team turn scattered notes into twenty confident pages before the meeting, and everyone wants to approve the recommendation and go home.

Then someone asks, “What decision is this actually for?”

The room goes quiet.

The memo compares three markets, but no one agreed whether the company is choosing a launch country, a research priority, or a budget allocation. The work looks ready. The thinking is not.

That moment is metacognition at work. Not abstractly “thinking about thinking,” and not doubting everything. Metacognition is noticing the condition of your thinking and changing the process when confidence has moved ahead of evidence.

In an AI-assisted workplace, that pause matters because fluency arrives earlier than certainty. A document can become impressive before the underlying question, assumptions, and accountability are clear.

Metacognition is a steering function

Researchers often describe metacognition through two connected activities: monitoring and control. You monitor what you know, how confident you are, and whether the current approach is working. You then use that judgment to control what happens next: spend more time, seek different evidence, change strategy, ask for help, or stop.

The relationship is imperfect. People can feel certain and still be wrong; they can notice confusion and respond with an unhelpful method. Experimental research on metacognitive monitoring and control shows that the link between subjective confidence and subsequent behaviour is complicated, not a flawless internal alarm. Koriat et al., Journal of Experimental Psychology: General

This is why metacognition is not simply self-awareness. “I feel uncertain” is awareness. “I feel uncertain because our forecast depends on an untested retention assumption, so I will request cohort evidence before approving the budget” is metacognitive control.

It is also different from overthinking. Overthinking circles the discomfort and keeps generating possibilities. Useful metacognition is bounded: it identifies what would make the thinking adequate, changes the process, and returns to the decision.

Why polished AI output can create false readiness

Generative AI is not the only source of overconfidence. Seniority, beautiful slides, consensus, deadlines, and a confident speaker can all make weak thinking feel complete. AI makes the problem more visible because it can produce a coherent surface so quickly.

A 2025 CHI study surveyed 319 knowledge workers and collected 936 examples of generative-AI use. Participants described critical-thinking work shifting toward verification, integration, and task stewardship. In the analysis, greater confidence in AI was associated with less reported critical-thinking effort, while greater confidence in their own ability was associated with more. This was a self-report survey, not proof that AI universally reduces critical thinking or that confidence directly caused the difference. It does show why calibration—matching confidence to the task and evidence—deserves attention. Lee et al., CHI 2025

Field evidence also warns against one global judgment about AI competence. In an experiment with 758 consultants, AI improved performance on tasks inside its demonstrated capability boundary, while performance worsened on a task outside that boundary when people used it. The tasks, participants, and model were specific, so the percentages are not a promise for another company. The durable lesson is that the boundary is “jagged”: success on one kind of work does not establish reliability on the next. Dell’Acqua et al., HBS research summary and paper

Back in the 4:45 meeting, the team’s mistake is not using AI. It is treating the quality of the memo’s surface as evidence that the decision beneath it is sound.

Use the Thinking Checkpoint before commitment

The checkpoint has four questions. Use it when a decision is consequential enough that being wrong would be expensive, difficult to reverse, or harmful—and when the work suddenly feels “done.”

It should take minutes for an ordinary decision, not become a second project. If a question exposes a real gap, assign the smallest action that would close it.

1. What decision are we making?

Write the decision as a choice that belongs to a named owner and has a time horizon.

“Analyse European expansion” is not a decision. “The executive team will decide by 30 September whether to fund a six-month validation in one European market” is.

The sentence reveals whether the evidence in front of you is even relevant. In the strategy meeting, much of the memo compared long-term market attractiveness. That might inform a later launch decision, but it could not by itself select the next research investment.

If the team cannot finish the sentence, pause the recommendation. More analysis of an unnamed decision usually creates more material, not more clarity.

2. What are we treating as true?

Name the few assumptions that must hold for the recommendation to make sense. Do not list every uncertainty in the universe. Look for the assumptions carrying the weight.

The memo assumes the existing offer will transfer with limited localisation, that distributor interviews represent buyer demand, and that a competitor’s public pricing reflects actual transaction economics. AI helped organise the available material; it did not verify those assumptions.

Now mark each one:

• Supported by direct evidence
• Plausible but not yet tested
• Contested or contradicted
• Unknown

This step separates a fact from a useful guess without pretending that all uncertainty can be removed.

3. What would change our mind?

Teams often ask for evidence that supports the preferred answer. Metacognition asks what result would make that answer weaker.

For the market memo, perhaps the team would delay if interviews reveal a mandatory product adaptation, if the contribution margin falls below an agreed threshold, or if a small paid test attracts interest but not qualified demand.

Define that evidence before commissioning more research. Otherwise every new signal can be explained away after the fact.

This question does not demand certainty. It creates a stopping or changing condition. If no conceivable evidence would alter the recommendation, the meeting is defending a commitment, not examining a decision.

4. What must remain human-owned?

AI can gather, compare, draft, challenge, simulate, or recommend. It cannot absorb the organization’s accountability simply because its output is useful.

Name the person who owns the decision. Then identify the judgments that cannot be delegated: whether the evidence is sufficient, which stakeholders bear the downside, what trade-off the organization accepts, and when the decision must be revisited.

This is consistent with the NIST AI Risk Management Framework, which calls for defined roles and responsibilities in human-AI configurations and oversight. NIST provides a risk-management structure, not a guarantee that adding a person to a workflow makes it safe. Human review only helps when the reviewer has context, authority, time, and a clear responsibility. NIST AI RMF Core

In the meeting, the strategy lead remains responsible for the recommendation. The research owner is responsible for the evidence request. The executive team owns the risk appetite and funding decision. “The model suggested it” belongs to none of those roles.

The checkpoint must change something

Reflection has earned its place only when it changes the work.

After the four questions, choose one outcome:

• Proceed because the decision, assumptions, disconfirming evidence, and owner are adequate.
• Narrow the decision to what the evidence can support.
• Test one load-bearing assumption.
• Change the method because the current evidence cannot answer the question.
• Escalate because the trade-off or risk belongs to someone with different authority.
• Stop or delay because the decision is not ready or no longer worth making.

This prevents metacognition from becoming a performance of sophistication. The output is not “we had a thoughtful discussion.” The output is a changed decision path.

For the strategy team, the checkpoint converts “approve European expansion” into something smaller: fund two weeks of customer and localisation research in one market, with a predefined condition for continuing. The polished memo becomes an input, not the decision itself.

Make uncertainty safe enough to say aloud

A checkpoint fails if the person who notices a problem is punished for slowing the room down.

Leaders do not need to celebrate doubt endlessly. They can make three behaviours normal: naming confidence levels, asking what would change the recommendation, and thanking the person who catches a faulty premise before execution. They can also model a changed mind without treating it as a loss of authority.

The language matters. “I do not like this” invites a debate about taste. “I think our confidence is ahead of our evidence because this assumption is still untested” gives the team something to examine.

That is especially important when AI is involved. People should be able to question an AI-assisted result without becoming “anti-AI,” and use one without being assumed careless. The standard is not loyalty to or suspicion of the tool. It is whether the thinking process fits the decision.

If you want to explore the adjacent risk of letting tools carry too much cognitive work, see cognitive offloading. For the broader shift in knowledge work, read how AI changes reading, writing, and thinking at work.

Return to the decision, not to endless reflection

It is now 5:02 p.m. The team has not approved the expansion recommendation. It also has not scheduled a philosophical retreat about decision-making.

It has named the actual decision, exposed three assumptions, defined evidence that could reverse the choice, and assigned human ownership. The next step is smaller than the original recommendation and more useful: test the assumption carrying the most risk.

The AI-assisted memo did not become worthless when the team questioned it. It became properly placed.

That is what metacognition contributes at work. It gives you a brief moment outside the momentum of your own thinking—long enough to see whether the confidence is deserved, change the method if it is not, and then move again.

Sources and scope

• Koriat, Ma’ayan, and Nussinson: monitoring and control in metacognition
• Lee et al. (CHI 2025): generative AI and self-reported critical-thinking effort
• Dell’Acqua et al.: field evidence on the jagged technological frontier
• NIST AI Risk Management Framework Core

The cited studies illuminate particular samples, tasks, models, and methods. They do not prove that one checkpoint improves every workplace decision or that generative AI has one uniform effect on thought. The Thinking Checkpoint is a practical decision aid, not a substitute for domain expertise, evidence, governance, or accountable judgment.