Worked examples. What a decision architecture diagnostic actually finds.

Situation

A working AI recommendation system was being scaled across an operations footprint after a successful pilot in two divisions. By the third division, throughput collapsed and the executive sponsor was being briefed that the rollout was an "adoption problem". A capital request for further scale was pending.

What the graph showed

WorkLattice mapped the decision chain the AI sat inside, from input through recommendation through human ratification through outcome. Eight decision-holders across the three divisions were interviewed and a decision graph was constructed: authority, accountability, override routes, and control points.

The structural pattern

The pattern that surfaced was not adoption. Two divisions had a clean delegation chain that let middle managers ratify model outputs in under a day. The third division had an inherited delegation override - a steering committee with informal veto rights - that no documentation reflected. The AI was producing recommendations at machine speed and the override layer was absorbing them at committee speed. Queue length grew until the operating layer stopped trusting the recommendations and reverted to pre-AI process.

The operating evidence

The evidence sat in escalation logs and committee minutes: median ratification time in division three was 11 working days versus 0.8 working days in the other two. The override layer had pre-existed the AI by years; it was load-bearing for unrelated governance reasons. The model was not the problem. The structure was.

The sponsor decision

Sponsor decision: conditional scale. Division three was paused. The structural condition (override-layer ratification speed) was identified as the gating constraint. A six-week remediation was scoped to re-tier the override routes for AI-generated recommendations specifically, leaving the legacy governance untouched. Scale to subsequent divisions was made contingent on a measured ratification-speed metric, not a calendar date.

What it changed

The diagnostic gave the sponsor a structural reason to pause that the operating layer accepted. "Adoption problem" would have been wrong, expensive, and resented. The actual constraint was identifiable, specific, and remediable.