Will AI make weaker problem solvers? A lean answer

Lean problem-solving methods are designed to build people as much as to fix problems. Mike Rother’s Toyota Kata (2009) describes improvement as a routine practiced daily, with a coach, until it becomes a habit; the A3, as John Shook showed in Managing to Learn (2008), works the same way through the dialogue between an author and a mentor. In both, the learning comes from answering the coach’s questions and going to see.

That is what AI can short-circuit. At the Lean Summit 2026, Art Smalley summed it up: “AI is a tool; you need to do the thinking for it to lead to kaizen.”

AI weakens problem solvers in three ways, all tempting because they save time on the day:

• It answers instead of challenging: the assistant writes the 5-whys, the root cause and the countermeasures, and the author edits a finished page instead of thinking.
• It produces the plausible cause: the analysis reads well, nobody spots the jump, and nobody goes to the floor to test it.
• It takes the formative cases: Bertrand Duperrin (October 2026) warns that if AI absorbs the tasks through which experience is built, an organization can become more efficient while learning less.

An AI assistant strengthens problem solvers when it is set up as a coach rather than an author. Tyson Heaton and Art Smalley (Lean Enterprise Institute, July 2026) recommend stating the coaching style in the prompt (directive, guided and neutral, or question-only) and naming the method, so the model does not average several. For the root-cause step they add choosing the type of analysis yourself, 5-whys, fishbone, scatter plot or design of experiments, because a model left to choose reaches for the 5-whys nearly every time.

Used that way, the assistant sorts each step of the author’s reasoning into fact, assumption or logical jump without proposing a root cause, says what to go and see, and prepares the mentor’s questions. The author still does the thinking; it just gets challenged sooner and more often.

The Lean Enterprise Institute lists the guardrails it writes into its own problem-solving prompts: “Analyze only the data I give you. Do not invent numbers. Do not state the root cause. Ask me for what you do not have. Do not blame the operator. Explain the probable cause-and-effect mechanism.” Among the things to avoid, it adds: do not propose training as the countermeasure.

Our own lean prompts add one rule from jidoka: treat missing data as missing, not zero, and list it, so that a gap becomes a question instead of a guess.

A five-step routine keeps the learning with the team and needs no new tool:

• The author drafts the current condition and the analysis first, without AI, from the data and a floor visit.
• The assistant challenges the draft: facts, assumptions, logical jumps, what to go and verify, and no root cause of its own.
• The author goes back to the floor with those points and revises the analysis.
• The mentor uses the assistant to prepare questions only, then holds the coaching conversation in person.
• Once a month, the team looks at its own capability, not its speed: which analyses held up, which countermeasures worked.

Whether AI builds or erodes skills is a management decision, not a property of the tool. Art Smalley (February 2026) recommends treating standards as enablers rather than constraints, for example requiring a coaching check before a problem-solving report is submitted, building feedback loops for both people and prompts, and designing escalation paths that turn bad AI output into learning rather than blame.

Our guide to AI for A3 problem solving gives a prompt for each box of the A3, including the challenge of the 5-whys and a question-only prompt for the mentor.