An online home-goods retailer whose support centre handles 95,000 contacts a month, most of them repeat questions that depend on live order data.
The business
Corvane Home sells furniture and homewares online in North America, about 2.1 million orders a year. Its support centre handles 95,000 contacts a month: 55% by chat and 45% by email. Volume rises to 2.5 times normal in November and December.
| Contact type | Share | Where the answer lives |
|---|
| Order status ("where is my order?") | 38% | Order Management System, live |
| Returns and refunds | 24% | Returns portal and refund policy |
| Product questions (assembly, care, dimensions) | 21% | Product manuals and help centre |
| Account and billing | 9% | CRM and payment provider |
| Complaints | 8% | Needs a person |
Measured baseline
- Average handling time: 7.5 minutes per chat, 11 minutes per email.
- Customer satisfaction (CSAT): 78%. First-contact resolution: 64%.
- 120 support agents at a loaded cost of $32 an hour.
- Refunds over $250 need a supervisor's approval.
Systems and content
| System | Role | Notes |
|---|
| Order Management System | Live order and delivery status | API, 95th-percentile response 250 ms |
| Returns portal | Creates returns and refunds | API; refund rules are configured here |
| Product catalogue | 8,000 products, 2,300 PDF manuals | Manuals change when a product is revised |
| Help centre | 400 articles | Updated weekly by the support content team |
| CRM | Customer history and contact records | Agents work every contact here |
Constraints and targets
- Chat: first response within 3 seconds at the 95th percentile. Email: reply within 4 hours.
- Card numbers are handled by a PCI-certified payment provider and must never enter the model's context.
- The approved run budget for AI is $18,000 a month, including peak season.
- Implementation is estimated at $250,000 one-off, plus $6,000 a month for hosting, monitoring and support.
- System prompt, policies and tool definitions: a stable 6,000 tokens.
- A typical chat runs 4 turns. Each turn adds about 2,500 tokens of conversation and retrieved or tool content, and produces about 300 tokens of output.
People
| Stakeholder | Cares most about |
|---|
| Director of Customer Experience (sponsor) | CSAT, resolution, peak-season coverage |
| Head of E-commerce Engineering | Integration effort, reliability, on-call load |
| Finance business partner | Run cost, payback, headcount assumptions |
| Support team leads | Escalation quality, agent workload |