Snapshot
Who: Integrated steel service center (anonymized) — four sites, shared inventory, high SKU mix.
Problem: Coil tags damaged in the yard, partial picks without photo proof, and nightly reconciliations that never quite matched the ERP.
What we shipped: Edge vision models for tag & stencil OCR at receiving, a reconciliation service that scores WMS vs. scale vs. imagery, and human-in-the-loop queues for low-confidence reads.
Challenge
Operations leaders could see shrink in the financials but not attribute it to a lane, shift, or carrier window. Auditors wanted lineage from weighbridge to bay.
Approach
We instrumented the highest-variance choke points first (receiving + transfer aisles), froze scope for a nine-week pilot, and wired outputs only into systems teams already trusted.
Solution components
Rollout timeline
Weeks 1–2 · Discovery
Shadowed receiving crews, mapped 14 exception types, baseline variance by site.
Weeks 3–6 · Pilot build
Two lanes live with shadow mode (AI suggestions not auto-posting).
Weeks 7–9 · Hardening
Latency SLOs met, security review passed, cutover for auto-post on >92% confidence.
Months 4–6 · Scale
Rolled to remaining plants; integrated weighbridge CSV drops.
“For the first time we could show finance exactly which movements cleared and which ones didn’t—without adding headcount at the gate.”
— VP Operations, anonymized program sponsor
Results & controls
- Write-offs tied to mis-identified coils fell 41% YoY in the pilot corridor.
- Night-shift tally work dropped ~18k hours per quarter across four sites.
- 99.2% of auto-posted transactions required zero rework after month five.
- Board-ready audit pack: immutable event ledger + confidence histograms per site.
Representative scenario for marketing purposes. Company name, figures, and timeline are illustrative composites and not attributable to a single public customer without a signed release.