How to Calculate AI Chatbot ROI Before You Build

Most chatbot ROI math starts too late. Teams ask about model cost, tool cost, or chat volume after they have already decided to build. That misses the real question: which questions should the bot answer, and how much staff time does that work consume today?

For many Nepali businesses, the best chatbot work is not complex. It is store hours, branch details, delivery status, appointment booking, pricing rules, document requirements, admission dates, return rules, payment steps, and lead capture. These questions come through Facebook, WhatsApp, Viber, website forms, calls, and walk-in follow-up.

Start with the current support load

Pull one month of messages if you have it. If you do not, sample one busy week. Tag each message by topic. Keep the tags plain: price, location, delivery, return, booking, documents, complaint, and sales lead.

Then split the messages into three groups.

• Safe answers: the same answer is right most of the time.
• Collect and hand off: the bot can gather details, then send the case to staff.
• Human only: the answer needs judgment, approval, or private data review.

Only the first two groups belong in your ROI model. If you count refunds, medical advice, loan decisions, or pricing exceptions as bot wins, the model will look better than the real business.

Put a price on repeated questions

Use a rough but honest number. If a support person handles 900 repeated questions per month, and each one takes three minutes, that is 45 hours of work. Multiply those hours by the true monthly staff cost, including supervision and rework.

The expected handling rate should be conservative during the first month. A good pilot may start by handling only 25 to 40 percent of questions while staff review the logs. That is not failure. It is how you avoid bad public answers.

Add missed revenue

Support savings are only one part of the return. A chatbot can also catch leads outside office hours, ask for missing details, route a buyer to the right branch, or keep a customer from leaving because no one replied.

Do not guess wildly. Count only leads that already exist in your inbox or website chat. If 120 monthly leads go unanswered after hours and 20 percent usually convert when staff respond quickly, you have a number worth testing.

Subtract the hidden costs

A chatbot is not free after launch. Budget for source updates, testing, staff review, prompt changes, and channel changes. If prices, policies, admission dates, stock, or branch hours change often, someone needs to own those updates.

Check Nepali language before you count savings

Many support teams in Nepal receive mixed messages: Nepali in Devanagari, Romanized Nepali, English, local place names, abbreviations, and short replies sent from a phone. Test those examples before you assume a bot can handle volume.

Use real messages, not polished demo prompts. Include spelling differences, half-written questions, and mixed scripts. If the bot fails there, your expected handling rate should be lower until the source and prompts improve.

Know when not to build

A chatbot is the wrong first project when the business has no current support load, no clear source answers, no owner for updates, or no handoff path. In that case, start with a better FAQ, contact flow, or lead form. AI can come later.

A sane target for the first month

The first pilot should prove three things. The bot answers a fixed group of questions correctly. Staff can review and fix weak answers. Customers can reach a person when the bot reaches its limit.

If those three things work, expand the topic list. If they do not, fix the source and handoff before spending more.