Ecommerce teams in Nepal often run support across many channels. A customer asks on Instagram, follows up on WhatsApp, then calls the next day. The first AI project should reduce that repeated reply work.
The best first build usually sits close to the inbox. If a store cannot answer product, delivery, return, and payment questions fast, a more complex AI project will not help much. Start where customers are already waiting.
Start with customer questions
Gather the top questions from chat, comments, and calls. Most stores see the same topics: size, delivery area, payment, exchange, warranty, and stock. Put approved answers in one source before the bot goes live.
Use the exact wording buyers use. Save examples from Facebook comments, Instagram messages, WhatsApp, Viber, website chat, and call notes. Many buyers will ask with short phrases, mixed Nepali and English, or product nicknames that do not match the catalog.
Build a small answer bank from the source your staff trusts now. That might be a product sheet, delivery table, return policy, size chart, warranty note, and store location list. If those files conflict, fix the source before adding AI.
Decide which channels come first
Do not connect every channel on day one. Choose the channel with the highest repeat load and the easiest handoff. For many stores, that is website chat or Facebook Messenger. WhatsApp and Viber can follow after the answer rules are stable.
Each channel has different behavior. Website users may ask product questions. Social users may ask for price and stock. WhatsApp users may send photos, voice notes, or one-word replies. The pilot should match the channel, not a generic demo.
Use AI to sort returns
Returns need rules. The system can collect order number, reason, photo, pickup area, and purchase date. A person should still approve edge cases and high-cost refunds.
Write the return flow as a checklist. Which products can be exchanged? How many days are allowed? Is the item used? Does the customer need packaging? What photo is required? AI can ask for missing details, but staff should approve any exception.
Read stock data every week
AI can turn a stock sheet into a short weekly summary. Which items are low? Which products moved slowly? Which branches need a transfer? Keep the first report small so staff can check it quickly.
Stock answers should come from a current system. If the store updates stock once a day, the bot should say when the answer was last checked or send the buyer to staff for confirmation. Do not let a bot promise stock from an old sheet.
Do not let a bot promise stock, discounts, refunds, or delivery dates unless those answers come from a current system or approved file.
Plan the handoff
A good ecommerce bot knows when to stop. Price exceptions, angry buyers, damaged goods, payment disputes, and delivery complaints should go to staff. The handoff should include the chat summary, order number, product, and what the customer already sent.
Without that context, staff ask the same questions again and customers get annoyed. The pilot should make staff work shorter, not move the same burden from one screen to another.
Test mixed customer language
Buyers may write in English, Nepali, Romanized Nepali, or all of them in one message. Test those examples before launch. If the bot cannot read them, keep the pilot inside staff review first.
Include spelling drift, product nicknames, local place names, and half-written messages. A buyer might write "kalanki samma delivery huncha?" or "COD xa?" or send a product screenshot with no text. Those cases belong in testing.
Measure one metric
For ecommerce, use a metric close to the inbox: first reply time, repeated questions handled, leads captured after hours, or return requests sorted. Review the logs every week.
Pick one old baseline. How long does the team take to answer after hours now? How many return requests need a second message? How many price questions never get a reply? The pilot should be judged against that baseline after staff review.
What to fix after launch
The first logs will show missing product names, weak return rules, outdated delivery areas, and questions staff did not expect. Fix the answer source before adding more channels. A small, accurate bot is better than a broad bot staff have to correct all day.
After the first channel works, add the next repeated queue. For most stores, that means moving from product questions to returns, then from returns to lead follow-up or stock summaries.