The Problem
This regional manufacturer had a process problem disguised as a technology problem.
Their operations team — small, skilled, and chronically overloaded — was spending more than 20 hours per week on work that shouldn't require a human: manually compiling quality-control reports from sensor data, processing supplier intake emails one by one, and chasing inventory discrepancies through spreadsheets.
Leadership had heard about AI. They'd read the think pieces and sat through the vendor demos. But every AI vendor they'd spoken to wanted to "transform their entire operation." They didn't want transformation. They wanted their team back.
What they needed wasn't an AI platform. They needed a partner who would actually study their workflows, identify where AI would create real value, and build something that worked — without the enterprise overhead or the 18-month timeline.
The Brief: Show us exactly where AI can save time, build only what will work, and prove ROI before we commit to more.
What We Built
Phase 1: AI Readiness Assessment (3 weeks)
Before writing a single line of code, we embedded with their operations team. We mapped five core workflows end-to-end, interviewed the people who actually did the work, and quantified the time cost of each manual process.
The output: a prioritized opportunity map with three high-ROI automation targets:
- QC Report Generation — 4 hours of manual work per cycle, structured data already available from sensors
- Supplier Intake Processing — 40+ emails per week manually sorted, categorized, and routed
- Inventory Anomaly Detection — discrepancies caught late, after the problem had already compounded
The assessment paid for itself before the pilot began. They cancelled a $12,000/year software subscription that the audit revealed was redundant.
Phase 2: AI Pilot (10 weeks)
We targeted the two highest-ROI opportunities: QC reporting and supplier intake.
- LLM-powered QC Report Generator: Ingests structured sensor data from their existing systems, applies their quality standard logic, and produces a formatted QC report automatically. What took 4 hours now takes under 15 minutes — with higher consistency and a full audit trail.
- Supplier Intake Automation: Incoming supplier emails are parsed, classified, and routed to the right team member with pre-populated context. Manual email processing reduced by 80%.
- Model: Built on GPT-4o with company-specific fine-tuning and a human-in-the-loop review step to maintain compliance standards
- Integration: Connected to their existing ERP via API — no data migration required
- Governance layer: Full logging, explainability dashboard, and rollback capability built in
The Results
"The assessment alone was worth the engagement. We finally had a map instead of a maze." — VP Operations
| Metric | Before | After | Change |
|---|---|---|---|
| Hours/week on manual reporting | 20+ | ~2 | −90% |
| QC report generation time | 4 hours | <15 minutes | −94% |
| Supplier email processing | Manual (40+ emails/wk) | 80% automated | −80% |
| Time-to-ROI | — | 6 months | Positive ROI |
Additional outcomes:
- Inventory anomaly detection added to Phase 3 roadmap (currently scoping)
- Operations team redeployed from data entry to higher-value work
- Client signed a Phase 3 Enterprise AI Program agreement
- Zero disruption to production operations during deployment
Why It Worked
The Assessment phase was the unlock.
Most AI implementations fail because they skip the discovery work. Vendors rush to build because building is what they sell. We do the opposite: we spend the first three weeks trying to talk clients out of building things that won't deliver ROI.
In this case, the assessment revealed that two of the five processes we evaluated weren't good AI candidates — the data wasn't clean enough and the edge cases were too complex. We scoped the pilot around the three that were. That discipline is why the pilot succeeded: we only built what we knew would work.
We also built for operations teams, not for data scientists. Every interface was designed to be used by people who needed to do their jobs, not manage an AI system. The human-in-the-loop step wasn't a concession to risk-aversion — it was a deliberate design choice that built trust with the team faster than any demo could.