What is catchball in Hoshin Kanri, and what can AI do in it?

The primary sources treat this as a principle of the method. The order below is our own cut of it, and an organization can run it in a workshop, in one-to-one reviews or across several weeks.

• The level above throws: an objective, why it matters, and the target it would like.
• The level below catches: it checks the objective against its capacity, its other commitments and what the shop floor shows.
• It throws back: the initiatives it proposes, the target it can defend, and what it needs to reach it.
• Both levels settle: the target, the conditions, or a difference that goes up to be decided.
• The result is written down where the strategy lives, then the same exchange starts one level lower.

The figures here are invented to show the mechanics. A group wants to cut customer non-conformity by 20% this year. A plant answers: 15% with the current team, 20% with two more quality technicians from the second quarter, and it flags that the same technicians are already promised to another priority.

That answer is worth more than a yes. The group now has three things it did not have: a defended target, a condition, and a conflict between two of its own priorities. Whichever way it decides, the decision belongs to the managers concerned.

None of these is an information problem, which is why no tool fixes them on its own.

• The stamp: people learn that pushing back has a cost, so they catch and submit. The plan looks aligned until the first review.
• The missing record: the exchange stays verbal, and three weeks later two levels remember different commitments.
• The skip: under time pressure, the plan is sent by email and called deployed.

The useful work is around the conversation: before it and after it. The risk is inside it, because a fluent agent can turn a reluctant "fine" into clean wording. A catchball that suddenly reads as harmonious is a warning sign, since disagreement is the signal the process exists to surface.

• Before a round: gather the initiatives already touching the objective, their status, and, if last year's matrix or data exists, what happened to similar targets.
• After a round: draft the record of the target, the conditions and the open conflict, for the people concerned to correct.
• Across levels: show which objectives from above no site has taken on.
• Not the agent: choosing the target, writing a team's counter-proposal as its own judgment, smoothing a disagreement, or declaring that a round is settled.

We are about to discuss this proposed objective with the level above: [objective and proposed target].

From the strategy data I give you, list:
1. The initiatives already linked to this objective, with status and owner.
2. The KPIs linked to it, with current value and target.
3. Any other objective competing for the same owners.

Then write the questions our team should settle before answering.

Rules: use only the data provided, and mark anything missing as missing. Do not propose a target, do not state a cause, do not write our position for us.

Hoshin has no catchball module, and nothing here runs the conversation. It holds what the conversation produced: the group matrix, the site matrices attached to it, and each site element pointing to what it deploys from above.

An agent connected through MCP reads the current state with `get_matrix_data`, shows what nobody took on with `get_cascade` (cascade belongs to the Enterprise plan, and everything is unlocked during the beta), and can record a counter-proposal as an idea linked to the element with `add_idea`. A write token can also change the matrix, so an agreed target should be written once the people concerned have confirmed it. A read-only token is enough to prepare a round.