The problem
Driver support calls follow a pattern: about 80% are known questions with known answers, asked thousands of times, in Nepali with English mixed in, at all hours, with spikes on weekends. A human call centre for that means 20+ agents for 24/7 coverage, three to six months to hire and train, and even then a 60–70% answer rate is considered good. Drivers who cannot reach support stop driving.
What was built
A dedicated TalkC instance connected to 17 phone lines, a knowledge base covering bonus structures, payment cycles, app troubleshooting, pricing rules and account verification, and a persona that speaks warm, natural Nepali with the fillers drivers use. The AI empathises first with an angry driver, then solves; it never says "your issue has been noted". Frustration or an explicit request transfers to a human with the transcript attached. Anything unresolved becomes a ticket with sentiment and priority.
Timeline: day 1 first conversation, day 2 lines and server, day 3 knowledge base from the existing FAQ, day 4 test calls, day 5 soft launch at 30% of calls, day 10 at 80%, day 14 all calls except explicit escalations.
Results after three months
| Metric | With TalkC | Typical human centre |
|---|---|---|
| Calls handled | 12,000 → 18,000 → 22,000+ a month | capacity-bound |
| Answer rate | 96.7% | about 60% |
| Time to answer | about 500 ms | 20–60 s hold |
| Average call | 62 s | 4–6 min |
| Resolved by AI alone | about 85% | — |
| Tickets auto-created | 3,700+ in 3 months | manual |
| Human team | 3 for escalations | 20 planned |
Estimated saving: about $120,000 a year against an all-human centre, with volume up 80% at flat cost. Weekly call reviews raised the resolution rate by roughly 30% over the three months.
The fleet partner: beyond the phone line
A Kathmandu fleet partner of a global ride-hailing platform runs its driver operation on TalkC. It started with calls (more than 20,000 by June 2026 from 4,500 drivers) and then moved the rest of the work onto the same AI:
- WhatsApp support on the official business number, with the AI replying in Nepali within about ten seconds and bundling several messages into one answer. By June 2026: 41,728 messages, 10,729 replies written by the AI.
- Documents read automatically. Photos of vehicle registration books, licences and citizenship cards are classified and the make, model, plate and owner extracted; app-error screenshots are described for support.
- Registration over chat: a guided conversation collects the driver's details and stores a draft for staff approval.
- Human agents with their own logins, assigned chats, role-scoped inbox and activity logs; the AI can be switched off per conversation by an admin.
- Onboarding pipeline on the platform's partner portal (invite, accept, vehicle, assign), proven end to end on real drivers, and fleet dashboards with drivers, vehicles, earnings and a live map synced on a schedule.
This is what we mean by automating the business, not the phone. How we do it →