Client and context
The client is an independent auto parts distributor operating in three LatAm countries. Under NDA we don't publish the name, but the profile is representative of the market: ~12,000 active SKUs, 60% OEM / 40% aftermarket mix, five strong vehicle brands (Toyota, Nissan, Chevrolet, Hyundai, Ford) and a sales channel dominated by WhatsApp (78% of orders enter via chat before hitting the ERP).
The customer-facing team was 11 reps taking orders manually: read the VIN or the customer's description, open the manufacturer catalog in a browser tab, check supersessions by hand, quote in Excel, then go back to the chat. A typical cycle took 12-18 minutes and ~30% of after-hours inquiries were lost.
The problem: expensive silent misorders
Before the project, 6.3% of monthly volume ended in a wrong-part return (misorder). The cost wasn't only logistical: each misorder consumed two calls, a courier visit, ERP order reopening, accounting adjustment, and in 1 out of 5 cases the customer migrated to a competitor.
Root cause wasn't lazy reps — it was structural: OEM catalogs in separate tabs, no automatic supersession (SSPL) validation, and zero connection between WhatsApp chat and real ERP stock. Reps were quoting what 'should be there', not what was there.
- Misorders: 6.3% of monthly volume (~480 returns/month).
- Average cost per misorder: USD 47 (logistics + time + ERP rework).
- Average quoting time: 14 minutes.
- Lost after-hours inquiries: ~30% of nightly inbound.
- Reps doing catalog work instead of closing sales.
The solution: AI agent + official EPC + ERP
We deployed AutoParts AI Agent as a customer-service layer on WhatsApp Business API. The agent receives the message (text, photo of the part, VIN or part number), decodes the VIN against the official OEM catalog, validates SSPL supersessions, queries stock and price in the ERP in real time, and returns a formal quote with confirmed availability — all in under 2 minutes.
When the case needs judgment (body panel with partial damage, non-standard substitution, customer credit), the agent hands off to a human rep with full context pre-loaded: decoded VIN, OEM candidates, stock per warehouse and customer history. Reps went from taking orders to closing complex sales.
- VIN decoding against the official OEM catalog (not scraped).
- Automatic SSPL validation before quoting — kills misorders at the source.
- Live ERP connector: stock, tier pricing and customer credit.
- Multi-number WhatsApp Business API, warehouse-based routing.
- Human handoff with full context whenever judgment is needed.
Implementation in 9 weeks
The project ran in four phases. Weeks 1-2: ERP integration (catalog, stock, tier pricing, customers) and normalization of the 12,000 SKUs. Weeks 3-5: connection to the official OEM catalogs of the five brands and SSPL rule configuration. Weeks 6-7: closed pilot with 200 B2B customers on a single warehouse. Weeks 8-9: rollout across the three country operations.
The client put a full-time product owner and two integrators from their IT team. From EITS: a solution architect, an integrations engineer and an AI trainer. The agent's learning curve was accelerated using 18 months of WhatsApp transcripts the client already had archived.
Results after 4 months
We measured against the quarter prior to implementation. All figures are from month 4 post full rollout, not from the pilot.
The KPI that moved business economics the most wasn't rep time savings — it was the misorder drop. At 480 avoidable returns/month × USD 47 unit cost, direct monthly savings exceeded USD 8,500, not counting NPS impact and B2B customer churn.
- Misorders: 6.3% → 3.9% (−38%).
- Average quoting time: 14 min → 4 min (−71%).
- Quotes resolved without a human: 0% → 67%.
- Orders generated outside business hours: +22%.
- Sales team throughput: 2.1× (more quotes/rep/day).
- B2B NPS (quarterly survey): +18 points.
- Implementation payback: 5.2 months.
Before vs. after — 4 months post-implementation
| Metric | Before (Q4 2025) | After (month 4) | Δ |
|---|---|---|---|
| Misorders / monthly volume | 6.3% | 3.9% | −38% |
| Average quoting time | 14 min | 4 min | −71% |
| Quotes without human intervention | 0% | 67% | +67 pp |
| Orders generated outside business hours | Baseline | +22% | +22% |
| Throughput per rep / day | 1.0× | 2.1× | +110% |
| B2B NPS | +24 | +42 | +18 pts |
Lessons for other distributors
Three lessons we keep seeing in other parts-distribution deployments. First: ROI lives in misorders, not in time savings. If you sell only 'faster rep', the business case stays small; real savings come from cutting returns, order rework and silent churn.
Second: without official EPC and live SSPL, the agent just automates the mistake. A chatbot wired to a scraped database quotes the wrong parts faster. Third: human-AI handoff must carry full context or the rep disables it in a week — the agent has to be the rep's assistant, not a hostile filter.