Lean before AI: how not to automate waste

Automating a task that should not exist makes it cheaper to repeat, and therefore harder to question. Bertrand Duperrin (October 2026) puts it in one line: “AI can make waste so inexpensive that we lose any incentive to eliminate it.” KPI Fire, a strategy-execution software vendor, calls the hidden cost of layering automation on non-standardized processes “transformation debt”: without a standard, an agent has no logic baseline, and the errors it automates propagate.

Lean practitioners say it in their own words. At the Lean Summit 2026, Grainger’s CTO Jonny LeRoy warned: “You need to rethink your processes before sprinkling them with AI pixie dust.” The Lean Global Connection 2026 programme includes sessions titled “Before You Automate, Stabilize”, “AI Readiness: Are We Automating Chaos?” and “Lean Before AI”, in which the Instituto Modelo de Cardiología in Córdoba, Argentina, presents its choice of standardizing processes before automating them.

Reporting is where AI pays back fastest in a lean management system, and where the trap is closest. Duperrin lists “a report that’s never read” among the activities that remain waste even when automated. An assistant that drafts the A3, compiles the KPI bowler and writes the review deck in minutes frees hours for the gemba if those documents were needed; if they were not, it makes producing three decks instead of one almost free, and the review drowns in them.

The test is simple: before automating a report, name the person who reads it and the decision it serves. A report that serves no decision should be stopped, not automated; a report that serves one should shrink to what that decision needs, and only then be drafted by an assistant.

Four questions decide whether a task is ready for AI (value, standard, flow and measure), and each has a clear “not yet” answer:

• Value: who uses the output, and for which decision? If nobody can say, stop the task before automating it.
• Standard: is there a current best-known way to do it, written down? If every person does it differently, the assistant will reproduce the variation, not the method.
• Flow: is this task the constraint? Tom Peperkamp (Lean Enterprise Institute, February 2026) shows that AI sped up coding, which was rarely the bottleneck, and created a queue of review instead of faster delivery.
• Measure: how will you know the result is better, and who decides what better means? Define it before the assistant optimizes for something else.

A strategy needs stabilizing before agents read it, just as a process does before it is automated. An AI agent that reads a strategy scattered across a slide deck, a chat history and a folder of notes will reason on the most recent or most repeated text, not on what the leadership team decided. It needs a stable, bounded form: objectives, initiatives and KPIs with explicit links, owners and statuses.

That is the role of the X-Matrix in hoshin kanri, and the reason hoshin.app exposes it to AI agents through MCP: the matrix is the standard the agents work from, and its completeness score shows where the standard itself is incomplete.

Lean before AI means starting small and standardized, not waiting for maturity. Art Smalley (Lean Enterprise Institute, February 2026) describes three predictable management mistakes: locking everything down, mandating AI without guidance, and handing out tools without shared routines. The lean answer is one task, one written prompt with guardrails, one person who checks the output, and a regular look at whether the task got better or just faster.

The reporting layer of the management system fits that start, because its value is easy to name and its waste easy to see. Our guide to AI prompts for lean gives ready-made prompts for that layer.