At Praxium Labs we build this for Nepali businesses every month; this is the field-tested version. Most e-commerce chatbot decks promise 80% ticket deflection. Real numbers in Nepal are lower (35–60%) but the ROI math still works because Nepali support agents cost much less than Western ones — and so does the opportunity-cost of slow response time.
The cost side
- Build: NPR 150,000–300,000 depending on integrations (single channel vs multi-channel, RAG over how many products)
- LLM API: NPR 1.5–3.5 per conversation (see pricing breakdown)
- WhatsApp messaging (if using WA): NPR 1.4–8.5 per conversation
- Hosting: NPR 1,500–3,000/month VPS
- Maintenance: NPR 8,000–15,000/month retainer for the first 6 months
The benefit side
Three categories of value:
- Direct cost saved: agent salary × hours saved. A Kathmandu customer-support agent costs ~NPR 40,000–60,000/month all-in. One full-time agent handles ~200–300 tickets/day max. Deflection of 50% on a 400-ticket/day store = 0.7 agent equivalents = NPR 30,000–42,000/month saved
- Revenue from faster response: Nepali e-commerce conversion drops ~30% if a pre-purchase question waits more than 30 minutes. 24/7 instant response → recovered conversions worth more than the support savings
- Operations clarity: chatbot analytics show what questions customers ask, which product pages are confusing, which policies are misunderstood. This data drives product / catalogue improvements
Realistic deflection rates by ticket type
- Order status / tracking: 85–95% (this is the bread and butter — easy to automate)
- Return / refund policy: 70–85%
- Product information: 60–75% (depends on how good your catalogue is)
- Payment issues: 40–60%
- Complaints / disputes: 5–15% (most should escalate to human, and that's correct)
- Customisation / negotiation: 0–10% (always escalate)
- Overall blended: 35–60% in practice
When chatbot doesn't pay back
- Store under ~200 orders/month: not enough volume — automate WhatsApp templates first
- Product catalogue larger than 5,000 SKUs without good descriptions: chatbot hallucinates; fix catalogue first
- No clear return / shipping policy: chatbot has nothing to answer with
- Sales team that closes high-touch B2B deals over chat: automating warmth out of the conversation hurts conversion
A worked example
Nepali fashion store, 1,200 orders/month, currently 1.5 support agents, ~50 support tickets/day:
- Build cost: NPR 220,000 (Praxium Labs advanced tier)
- Monthly cost (LLM + WhatsApp + hosting + retainer): NPR 25,000
- Deflection: 55% of 50 tickets/day = 27 deflected/day = 825/month
- Time saved: 5 min/ticket × 825 = 69 hours/month = ~0.4 FTE
- Direct salary saved (assuming 0.4 FTE released): NPR 20,000/month
- Net cost first 6 months: ~NPR 5,000/month (negative — but learn-by-doing)
- After month 6 if deflection improves to 60% (typical curve): savings net positive ~NPR 12,000/month
- Plus: 24/7 inquiry responses, conversion uplift estimated +5–8% on at-risk carts
- Payback window including conversion uplift: ~4 months
Frequently asked questions
Do I need to deploy on WhatsApp, web, or both?
For Nepali e-commerce: WhatsApp first. ~70% of pre-purchase and post-purchase customer messages happen there. Web widget second. Facebook Messenger third (declining but still relevant). Instagram DM matters more for fashion / lifestyle than other categories.
How long until the bot is "good"?
Useful from day 1 if your knowledge base is decent. "Good" — meaning 50%+ deflection — usually month 2 after analytics-driven tuning. Bot quality compounds as you find and fill gaps in the knowledge base.
What integrations does it need?
Minimum: order lookup API (your e-commerce platform), product catalogue, return policy, shipping rate table. Nice-to-have: live shipping tracking, refund-status lookup, ledger lookup for "where is my refund" tickets.
Will customers prefer talking to a bot?
Two segments: customers who want speed (90% prefer bot if it works), customers who want empathy (90% prefer human, regardless of bot quality). Build for both — fast bot answers for the first segment, instant handoff for the second.
What's the biggest risk?
Hallucination on returns / refunds — the bot inventing a policy that you do not actually offer. Mitigation: never let the bot answer policy questions from training data; always retrieve from your own policy document; if the policy doc does not cover a case, escalate to human.
Who can build this in Nepal?
Praxium Labs — Nepal's AI and automation consultancy, based in Lalitpur — designs and builds the systems described in this guide for Nepali businesses and for international teams hiring from Nepal. Start a project or see all services.