The Problem
A 60-person custom metal fabrication company had a reporting problem that was hiding in plain sight. Every Monday morning, a production coordinator spent 3–4 hours pulling data from their ERP system, their CNC machine monitoring software, and a separate quality control database. She'd paste it into Excel, clean it manually, build a summary table, and email PDF reports to five department heads.
Then on Friday, she did it again for the weekly close-out.
That's 6–8 hours every week—roughly 300 hours per year—spent on a task that exists only because three software systems don't talk to each other.
Nobody questioned it because it had always been done that way. The coordinator was good at it. It worked. But it was entirely manual, error-prone when data was late or formatted differently, and bottlenecked on one person. When she took vacation, the reports didn't go out.
The plant manager knew this was a problem but assumed fixing it required replacing their ERP system—a multi-year, multi-hundred-thousand-dollar project they weren't ready for.
We told them there was a faster path.
What We Did
Week 1: Connections and data mapping
We started by connecting to all three data sources: their ERP via a read-only API endpoint, the CNC monitoring software via its export API, and the QC database via a direct SQL connection. No system replacement, no migration.
We mapped the data each report needed and built a normalized data model that reconciled the different formats and terminology across the three systems. (The ERP called a production order a "work order." The CNC system called it a "job." The QC database called it a "lot." Same thing, three names. Classic.)
Week 2: AI-assisted normalization and report generation
The raw data alone wasn't enough—the original reports included summaries, trend callouts, and exception flags that required judgment. We used an LLM layer to:
- Identify production variances above threshold (and write a plain-English explanation of what likely caused them)
- Flag quality issues that matched patterns from prior rejected batches
- Summarize week-over-week performance in 2–3 sentences the department heads could actually use
The output is a formatted report—same structure as before, just generated automatically.
Week 3: Delivery, testing, and handoff
We set up automated delivery: Monday morning reports run at 5am and land in inboxes before anyone starts their shift. Friday close-outs run at 4pm. The plant manager gets a Slack notification when they're done. If a data source is unavailable, the system alerts the coordinator before the report would have been late—not after.
We spent the final days testing edge cases (missing data, system outages, format changes) and training the team on how to adjust thresholds and add new report recipients themselves.
The Results
The coordinator now spends about 15 minutes per week on reporting—spot-checking the output, not generating it. She redirected the recovered time to production scheduling work that had been backlogged for months.
The reports are more consistent than the manual versions because the AI applies the same rules every time. Human fatigue at 4pm on a Friday doesn't affect the output.
The plant manager's comment: "I expected this to be more complicated. The hard part was just getting the three systems to give us clean data—and you handled that."
Why This Works for Manufacturing
Manufacturing businesses often have exactly the right conditions for AI automation:
- Structured, repetitive data — Production metrics are predictable in format and timing
- Multiple disconnected systems — ERP, MES, QC, and scheduling tools rarely integrate natively
- High cost of manual consolidation — Skilled people spending hours on data entry instead of decision-making
- Clear success metrics — Easy to measure time saved and error rate
You don't need to replace your ERP. You need the systems you have to work together—and AI is now good enough to handle the translation layer between them.
If your operations team is spending meaningful time every week on reports that could run automatically, start with our AI readiness assessment. We'll tell you honestly whether automation makes sense and what it would cost.