From the loom to AI: stopping at the defect is still the key to continuous improvement

Toyota Industries dates the Toyoda power loom to 1896 and notes that it was equipped with a weft auto stop mechanism. Toyota’s own history describes the device: it “automatically stopped the machine when the weft thread in the shuttle broke or was exhausted”, That history gives 1897 for the completion of the loom; the two sources differ by a year. On the loom of 1905, it describes a warp halting device that shut the machine down when a warp thread broke.

In 1924 the Type G automatic loom changed its shuttle without stopping, and kept the weft-break and warp-break auto-stops. Toyota writes that Sakichi’s design concepts “live on even today in the Toyota Production System as the origin of jidoka”, which it defines as “automation with a human touch”.

A loom that stops produces no defective cloth, but that is the smaller benefit. The larger one is what the stop sets off. Toyota describes it this way: the equipment detects the abnormality and stops, the andon lights up to notify people, and “people need only respond when there is an abnormality”. Nobody has to watch the machine to catch the defect, and each stop points at a cause to remove.

This is why stopping at the defect is the engine of continuous improvement and not a safety device. An alarm that nobody treats teaches nothing. A stop that someone goes to see, traces to its cause and turns into a better standard makes the next defect less likely. The stop is worth what follows it.

A management system has its own broken threads: an indicator with no reading for the month, an initiative with no owner, an annual objective that no initiative carries, a review that was held and never recorded. They are abnormalities against a standard, and until now someone had to go looking for them.

An AI agent connected to the system can detect them and say so, and stop where the decision begins. On our public demo, an agent asked about a certification deadline read the initiative sheet, where the owner had noted that material traceability depends on a scheduling tool not yet chosen. It set that dependency against the customer’s deadline, noted that the initiative has no KPI of its own, and put the choice of the tool to the leadership committee. It decided nothing. The trace is published with our four replayable scenarios. How to write the stop into an assistant is the subject of jidoka for AI agents.

Lean excellence roadmaps often take the form of a grid: a set of themes, each split into items, each item graded on four or five levels. We propose to describe the last level of every item the same way: the team owns the practice, an agent watches it against its standard and signals gaps, and people decide. Four examples, written as behaviors an auditor can observe.

• Strategy deployment: before each review, the gaps are on the table without anyone hunting for them (a KPI with no reading, an objective with no initiative), and the committee spends its time deciding.
• Problem solving: the assistant refuses to name a root cause without data, lists what is missing, and the team goes to see.
• Daily management: a deviation from the standard is raised the same day with its context, and the team leader chooses the countermeasure.
• Standard work: when practice drifts from the standard, the question is put to the team: fix the practice, or improve the standard.

First, the agentic level is only reachable from the one before it. An agent can signal a gap only where a standard exists, the data is kept and the review takes place. Placed on an immature practice, it measures nothing: this is the argument of lean before AI, applied to the roadmap.

Second, a level is described by what can be observed, never by the tool. “Uses AI” is not a level of maturity. “The gap is signaled before the review, and the team changes by itself what the agent watches” is one, and it can be audited.

Hoshin Kanri covers one item of such a roadmap, strategy deployment: the X-Matrix, the review cycle, and an agenda that tells an agent what is due (a reading that is missing, an initiative with no update, a review to record). The agent prepares; the committee decides.

The levels described here are a proposal, not a standard. If your roadmap already has a grid, the test is simple: take one item, write its top level as a behavior, and check whether an agent could signal the gap today. If it cannot, the level below is not yet in place.