Mid-market industrial distributor (headcount 120, $18M annual revenue). A sector that historically "isn't for AI". Here's how it shipped, what we touched and what we didn't.
The starting point
The admin team was burning 32 hours a week — across 4 people — on repetitive tasks: classifying delivery notes, reconciling invoices against purchase orders, answering recurring questions from branch offices, updating the ERP with data that arrived by email.
Invisible work. Expensive. And above all: work no one wanted to do but everyone needed done.
What we automated (and what we didn't)
| Process | Decision | Reason |
|---|---|---|
| Delivery note classification | Yes | ~100% clear rules, high repetition |
| Invoice-to-PO reconciliation | Yes | High volume, low exception rate |
| FAQ responses to branches | Yes | Top 30 questions cover 80% of volume |
| ERP updates from email | Partial | Edge cases need human review |
| Supplier negotiation | No | Relationship component + high risk |
| Customer credit decisions | No | Regulatory risk |
The stack
- GPT-4o-mini for document classification (unbeatable cost)
- Claude Sonnet 4 for answers that need long context
- Supabase as the orchestration + logs layer
- Direct webhook into the ERP for controlled writes
- Langfuse for observability
How it was measured
Before: a real 2-week stopwatch with the 4 people logging time per task category. After: automatic dashboard counting each document processed vs. the human SLA it was replacing.
| Metric | Before | After (week 8) | Delta |
|---|---|---|---|
| Hours/week on repetitive work | 32h | 4h | -87% |
| Average time per delivery note | 4 min | 12 sec | -95% |
| Manual errors | 4.2% | 0.3% | -93% |
| Cost per process | $5.20 | $0.45 | -92% |
What we COULD have automated but didn't
The client also wanted to automate supplier negotiation. We said no. Not because it was impossible — because it was inconvenient: the relationship layer with strategic suppliers is part of the value. Automating it would have been efficient and expensive at the same time.
Total cost and ROI
- Investment: $12,000 (setup + 8 weeks of implementation)
- Recurring cost: $480/month (LLM + infra + monitoring)
- Hours/year freed: ~1,400h (28h × 50 weeks)
- Equivalent value: ~$52,000/year (at a fully-loaded $37/hour cost)
- Payback: 3 months · Year-1 ROI: 308%
We left it running. The admin team wasn't laid off: they were reassigned to acquisition. In 6 months the company had grown 12% without touching headcount.