Agentic lean: where to start, and where the agent must stop

Agentic lean differs from using an AI assistant in lean because the agent runs a loop instead of answering a question. An assistant drafts an A3 when someone asks. An agent reads the records every morning, compares them with the standard and the targets, writes what it found where the team will see it, and stops when something needs a person.

The model is jidoka, not autonomy. Sakichi Toyoda’s looms did not automate the weaver’s judgment; they automated the stop when a thread broke. An agent in a lean management system should do the same: take the repetitive reading and drafting, detect the abnormality, stop and say what is missing. Our guide on jidoka for AI agents gives the stop rules as prompts.

A plant is ready for agentic lean when the loop it hands to an agent is already stable and written down. Kaizen Institute (June 2026) puts it in one sentence: “Standard work converts those insights into the repeatable operational logic that AI agents need to function reliably: a system that changes arbitrarily from day to day cannot be automated or optimized.” Its conclusion: “not AI replacing lean but lean creating the conditions that AI requires.”

Four checks tell whether a loop is ready, and each has a clear “not yet”. The longer argument, and the four questions to ask before automating any task, are in lean before AI.

• A standard exists for the loop: the same report, in the same format, every day.
• The data is recorded where an agent can read it, and a missing day shows as missing.
• Someone owns the gap: a named person reacts when the reading is off.
• The targets are written: without a target, the agent can describe the day but cannot say whether it matters.

The first loop to give an AI agent in a plant is the daily reading of production and quality: production against plan, OEE split into availability, performance and quality, downtime ranked by minutes and by frequency, scrap. It is repetitive, it already has a format, and its value is easy to name: the people in the morning meeting spend their time on the gaps instead of on compiling them.

The second loop is the problem report. An agent can draft an 8D from the non-conformity, the history and the measurements, as long as it stops at the steps that need something to have happened in the plant. Our guide to the AI-assisted 8D report gives the prompt.

What comes last, or never: anything that acts outside the plant or on the equipment. Anthropic writes that “Claude for Manufacturing doesn’t operate equipment”, and that its skills are built so that “the person responsible can review and approve the work before it reaches a customer or supplier.”

An agent in a lean management system must stop in five places, and each stop is written into its instructions, not left to its judgment:

• When data is missing: absent is not zero. A day without the downtime export stays a day without data, and a monthly figure says how many complete days it rests on.
• Before the steps that need reality: Claude for Manufacturing drafts an 8D from D1 to D5 and D7 and leaves D6 and D8 empty, because verified corrective actions and closure need something to have actually happened.
• Before anything leaves the plant: a quote, a purchase order or a corrective action report goes to a customer or a supplier only after a person approves it.
• When the target is missing: without one, the agent reports the reading and asks for the target instead of inventing one.
• Before a decision: the agent prepares the options and the evidence; the person in charge decides, and the agent records what was decided.

An agent that reads the plant every morning sees the gap to yesterday and to the standard. To say whether a gap matters, it needs what the plant decided to improve this year: the annual objectives, the KPI targets and the improvement initiatives behind them. Yoji Akao’s Hoshin Kanri (1991) separates the two: daily management holds the standard and reacts to deviations, hoshin kanri decides what must improve and deploys it.

Claude for Manufacturing covers the daily side well. The strategy side has to exist somewhere the agent can read: in hoshin kanri it is the X-Matrix, where each objective is linked to the initiatives and the indicators that measure it. hoshin.app keeps the X-Matrix as data and exposes it to any agent through MCP, so the same agent that read the plant can read the targets. Where AI fits in that layer is the subject of AI in lean management.

Only three things should go up from the plant to the strategy layer, each with its source:

• The reading of a strategic indicator for the period, with its definition, the days it rests on and where it came from.
• One improvement idea per repeating pattern, such as the same defect coming back after a fix that did not hold, never one per non-conformity.
• Evidence for the site’s maturity, proposed as a draft that a person validates.

An agent in agentic lean needs the smallest rights that let it do its loop: read the plant records, write readings and drafts, never validate. Every reading it writes keeps its source, so a person can check where a number came from. In hoshin.app a reading carries the tool that produced it, a reference to the source record and the extraction date, and validating a maturity assessment is signed in the app by a person, never by an agent’s token.

The review is part of the loop, not a formality. A person who approves everything the agent writes without reading it is not a review; the agent should give the reviewer little to read, with the doubtful points first.

We tried this on November 13, 2026, on a copy of our demo plant, with synthetic records in the exact format of Claude for Manufacturing’s morning report: 22 production days in October, two of them missing the downtime export, and a quality register where the same bore defect comes back after a corrective action that did not hold. The agent ran in Claude Code, connected to hoshin.app, with the plant records in the same account.

The agent wrote a monthly reading of 646 PPM on the quality indicator, resting on the 20 complete days and naming the two others. It wrote no downtime cost and no loss figure, because the days were incomplete and the cost rate was a placeholder. It opened one improvement idea for the repeated defect, not one per non-conformity. It proposed level 3, not 4, on the maturity item “What happens when a defect appears?”, because the fix had not held. The trial also found a defect on our side, fixed the same day: the hosted server did not keep the source of a reading.

What was not tested: a real plant, the Claude for Manufacturing skills producing the records themselves, plant records shared from another account, and an unattended run publishing its dashboard, which Anthropic’s own documentation says has not been verified yet.

Four mistakes recur, and each has a simple countermeasure:

• Automating a report nobody reads: the agent makes waste cheap. Name the reader and the decision before automating.
• Turning missing data into zeros: the monthly figure looks complete and is wrong. Keep absent as absent and say how many days a figure rests on.
• One ticket per event: the idea inbox floods and nobody sees the pattern. One idea per repeating pattern, with the events as references.
• Filling the X-Matrix with shop-floor data: it becomes a second dashboard and nobody reads it as a strategy any more.

In the lean maturity grid we propose, the agentic level is the top one, reached only from the level below: the loop already works with people, and the agent takes the repetitive part of it. The evidence for that level is not that an agent runs, but that the team acts on what the agent flags. Our guide on the lean excellence roadmap describes the grid and what its top level means for each item.