How a Tier-1 cut declaration rejections by 89%

This is a walkthrough of a deployment programme across a nine-plant components supplier over one fiscal year. The operator is not named — we are under confidentiality — but the sequence is described exactly as it ran, including the parts that went badly.

The starting position

Declaration rejections were running at roughly 14% of submissions. Each rejection cost between three days and two weeks of cycle time, and rejections were concentrated in the plants with the newest programmes rather than the oldest, which nobody expected.

Weeks 1–3: baseline and the hard week

The first fortnight was data access, not modelling. Week three was the hard one: the extraction was working, and it surfaced that two plants had been maintaining substance data in a local spreadsheet that had diverged from PLM eighteen months earlier. The programme nearly stopped there, because the immediate finding was that the problem was bigger than scoped.

Weeks 4–8: validation before submission

Rulebook checks were put in front of the submission step rather than after it. The rejection rate began falling in the first month, mostly from the reconciliation and completeness categories, which were the cheapest to catch.

Month 2–3: the plateau

Improvement stalled around 6%. The remaining rejections were customer-specific rulebook differences, which the shared validation set did not encode. This is the point most programmes are declared finished and quietly stop improving.

What cleared the plateau

1. Per-customer rule profiles rather than one universal template.
2. The supplier declaration backlog worked through in extraction rather than by chasing, which removed the oldest and worst records from the population.
3. A named reviewer per plant for the exception queue, with queue depth reported weekly alongside the rejection rate.

Where it landed

Rejections finished the year at roughly 1.8%, an 89% reduction. Audit preparation dropped from around eighteen person-days per facility to four. Neither number came from a better model; both came from moving the check earlier and giving the exceptions an owner.