Should you entrust your company strategy to an AI? Its upkeep, yes. Its decisions, no.

Lean has many tools: value stream maps, kaizen, standardized work, visual management. None of them says where to point the effort this year. That is the job of hoshin kanri, management by policy, which Yoji Akao codified (1991) and the Japanese standard JIS Q 9023 describes: a yearly Plan-Do-Check-Act on the company's policy itself, a few breakthroughs deployed level by level, kept apart from daily management, which holds the existing business to standard.

Thomas Jackson's Hoshin Kanri for the Lean Enterprise (2006) makes the link explicit in its structure. Value stream mapping appears inside the planning of the midterm strategy; the question "is the organization lean?" is asked in the president's diagnosis, a step of the hoshin cycle; standardized work and kaizen are how the hoshin is institutionalized. The lean tools execute, the hoshin steers them.

"First layer" is our wording, not Jackson's. It says one thing: in the order in which a lean company is built, the policy comes before the tools, because it decides what the tools are for.

A team that drifts from the policy drifts at the pace of meetings. An agent drifts at the pace of requests. Ask it for a campaign, a report, a supplier analysis, and it will optimize the request in front of it, with nothing to tell it whether that request serves this year's three breakthroughs or a fourth that nobody chose.

The answer is the same as in a lean company: give the agent the policy, not more instructions. Once the X-Matrix is readable by every agent, through MCP, each action can be traced up a chain:

• the action serves an initiative (South), owned and dated;
• the initiative moves an annual objective (East);
• the annual objective serves a 3-to-5-year breakthrough (North);
• and an indicator (West) will show whether it worked.

In lean terms, effort that does not serve the policy is waste. An agent that checks the chain before acting stops producing it: when a request traces to nothing, it says so instead of doing it. That is jidoka applied to agents: stop and make the problem visible, rather than keep going.

Sometimes the request is right and the matrix is wrong: a customer, a regulation or an incident has changed the situation. The agent cannot decide that. It logs the request as an idea linked to the closest element, and the next review, or the next round of catchball, decides whether it becomes an initiative. The policy stays the leaders', and nothing seen on the way is lost.

Before you act on my request, read our X-Matrix (get_matrix_data).
1. Name the initiative this action serves, the annual objective it moves and the indicator that will show it.
2. If the action serves none, do not do it yet. Tell me so in one sentence and offer to log it as an idea (add_idea), linked to the closest element if there is one.
3. If it serves one, do it, and say which chain it serves in your answer.
Never change a target, a breakthrough objective or a correlation on your own: propose it, I decide.

The line is the one lean draws between the work that keeps documents true and the judgment that decides. It holds for an assistant as for a coordinator.

• Entrusted to the agent: keeping the matrix up to date; recording each indicator reading with the tool it comes from; preparing the review agenda; flagging what is due, late or off target; checking that a proposal serves the policy; presenting the matrix in the deck, page or report the reader needs.
• Kept by the leaders: choosing the breakthroughs; setting the targets; arbitrating resources in catchball; going to see on the floor; deciding in review, with the decisions recorded as they were said.

On 7 and 8 November 2026, the leadership of a health services company built its 2027 X-Matrix with an AI agent connected to Hoshin over MCP. The company asked us not to name it. Three findings from that build answer the question of this page better than any argument.

The strategy was not where the team could read it. The revenue trajectory, the reference customers, the business model and the impact indicators lived in one executive's personal AI memory. Useful to one person in one assistant, invisible to everyone else. The test the agent proposed for where each piece of information belongs: if a colleague needs it tomorrow without that executive, it goes in the shared strategy.

Alignment needs definitions, not only goals. Without a formula, a source system and an owner for each indicator, two agents compute the same usage rate in two ways. The agent asked for indicator definitions, and for a frozen copy of the matrix at each review, so the basis of a decision stays readable after a figure is corrected.

The agent stopped instead of improvising. Each time it lacked a capability, it reported it from the server rather than working around it. Those reports are how the indicator definitions, the initiative sheets, the frozen reviews and the cadence tools were built into Hoshin the next day.

Confidentiality. An element marked internal never reaches a public page, nor do its correlations; what leaves the organization is an external reading without figures, names or history. For a firm that cannot send client data anywhere, the server also runs on your own machine, with the same tools: see keeping client strategy data in your own environment.

Loss of control. A connection never has more rights than the person behind it, and inviting someone stays a human act in the app. The agent writes; reviews are recorded with their decisions as they were said, and with the state of the matrix at that moment.

Errors. Each reading carries its source and the date it was extracted, is marked provisional while its period runs and final after, and every initiative has a plain page, generated by Hoshin, to check what the agent presents against.

Lock-in. The matrix is reached through MCP, an open protocol, from any agent that supports it; the server is open source (MIT), and the API exports the whole matrix. Your agent can change; the policy stays.

Write the policy first, small: three to five elements per quadrant, from the free X-Matrix template or with your agent. Connect the agents that act for the company to the same matrix. Then give them one rule, the prompt above: before acting, find the chain. Why the format has to be a structured record rather than a document in a prompt is explained in strategic memory; where the agents pay back first in a lean company, in AI in lean management.