AI · Learning design

When help at work replaces the chance to think

Just-in-time support can improve a result while leaving the person no better prepared for the next decision. Here is the design choice that matters.

Mike Vaughan · April 11, 2025 · 2 min read · updated September 26, 2026


A manager asks an AI assistant how to handle a difficult conversation. It supplies a sensible script. The conversation goes well.

What has the manager learned?

Perhaps a great deal. Perhaps only where to find a script. The successful outcome cannot tell us which. That distinction belongs near the beginning of a learning design, before the discussion turns to formats, platforms, or how quickly people can finish.

Keep the help. Examine what it replaces.

People need reliable information in the flow of work. Nobody benefits from having to memorize an infrequently used procedure or improvise a safety-critical instruction.

But when the objective is judgment, look closely at the work the support removes. Does the person still identify the problem? Compare explanations? Decide which evidence deserves weight? Consider who might be affected? Or does the system do all of that and ask the person to approve the answer?

There is a difference between removing an unnecessary obstacle and removing the very thinking someone needs to practice.

An example: change the order of the help

Consider a fictional manager preparing to discuss a missed deadline. A support-first design gives a script immediately. A practice design could ask the manager to first write down two plausible explanations for the missed deadline and one question that would help distinguish them.

Then offer the guidance. Ask what it changed. After the conversation, return to the initial explanations: which held up, which did not, and what would the manager ask earlier next time?

This adds cognitive work deliberately. It does not require making the real conversation harder or withholding essential information.

What the research can and cannot tell us

A 2025 survey of 319 knowledge workers found that higher confidence in generative AI was associated with less reported critical thinking. The study concerns self-reported effort and behavior. It does not establish that using AI causes a lasting loss of thinking ability.

That limitation matters. “All in-flow learning erodes thinking” goes further than this evidence supports. The practical concern is narrower: routinely outsourcing a decision may remove opportunities to develop the judgment that decision requires. Treat that as a design risk to investigate, not a universal diagnosis.

Check the second performance

A supported success is useful evidence about supported performance. To investigate learning, introduce another relevant situation. Record what help is available. Look for what the person recognizes and does without the earlier answer in front of them.

If the job normally includes tools, also examine how the person uses and checks them. Independence from every tool is not the only meaningful outcome.

Take this into a design review: Find one place where your program supplies an answer. What could the learner predict, explain, or decide before receiving it?

Try the launch decision to experience that sequence.

Concepts in this piece

  • What-to-think vs. how-to-think · Knowledge supports judgment. Practice gives people opportunities to use it when the answer is not already supplied.
  • Richness and Reach · Map cognitive demand and participation. Classify the actual experience, not its delivery format.

Bring a real problem

What does your learning need to change?

A role people are struggling with. A decision that keeps going wrong. A capability you need to see before giving someone more responsibility. Start there.

Talk with Mike

Find an idea in the writing

Try

AI-generated guidance based on the writing here. Each question is answered independently. Questions and responses are stored to help identify gaps in the writing; do not enter confidential or personal employee information.