Support Ticket Deflection With an AI Chatbot

Support teams call it ticket deflection: a question answered before it became a ticket. Most small businesses do not have tickets. They have a Gmail inbox, a WhatsApp Business number, a phone that rings during a consultation, and an owner who answers all three. For them, deflection means something plainer: messages that never had to be written, and replies you never had to type.

An AI chatbot on your website can take some of those messages away. How many, and whether it is really the chatbot doing it, is something you can measure in a month with a notebook or a spreadsheet. This post covers what to count, what a bot like TailGPT can and cannot take off your hands, and how to raise the number once it is running.

What does ticket deflection mean for a small business?

It means a customer got their answer on your website, so they never emailed, messaged or called you about it. The deflected question is the one that never arrives, which is why it is easy to claim and hard to prove.

Take a physiotherapy clinic in Indore. Every week the front desk answers the same messages on WhatsApp: are you open on Sunday, do you do home visits in Vijay Nagar, what does the first consultation cost, do I need a doctor's referral. All four answers are on the clinic's website. The receptionist answers them anyway, between patients, often more than once a day.

If a chatbot on the website answers those questions for the visitors who ask there, some of those WhatsApp messages stop. Not all of them: some patients will always message because messaging is what they do. But the ones who were already on the website, looking, have a faster route than waiting for a reply.

That is the whole idea. What makes it worth doing is the next part: knowing which messages stopped, and whether they stopped because of the bot.

How do you measure deflection honestly?

Count your repeat questions for two weeks before the chatbot goes live, then for two weeks after, and compare the same categories. Nothing cleverer than that is needed, and anything cleverer tends to flatter the bot.

Before (two weeks). Keep a simple sheet. Every time an email, WhatsApp message or call comes in, note four things:

1. The date.
2. The channel: email, WhatsApp, phone, Instagram, walk-in.
3. The topic in two or three words: "delivery to Pune", "return window", "Sunday hours".
4. Whether your website already answered it: yes, no, or partly.

The "yes" rows are your deflectable questions. That share is the ceiling for what a bot reading your current pages can take, before you write a single new page.

After (two weeks). Put the chatbot live and keep the same sheet, with the same topic names. Then compare topic by topic.

Make the two periods comparable. If you sold twice as much in the second fortnight, you should expect more messages, not fewer. Divide by something that tracks your business: messages per 100 orders for a shop, per 100 appointments for a clinic, per 100 enquiries for a coaching institute. Avoid comparing a quiet fortnight with Diwali week, an admission season or a sale.

Here is what that looks like for an online saree shop in Surat. The figures are illustrative.

Read it row by row. Delivery, cash on delivery and returns fell, because the shop's pages answer them clearly. Stitching barely moved, which says the sizing page is not doing its job. "Where is my order" and complaints did not move at all, and should not have: no website page can answer them. Across the first four rows, 45 repeat questions became 25, so about 44% of those repeats went away. That is the shop's deflection rate, measured on the questions a bot could take.

Two weeks is short, so treat the first result as a direction, not a verdict. Run the sheet again a month later, after you have fixed a few pages.

Which questions can an answer bot deflect?

The questions whose answer is the same for everybody and is written on your website. TailGPT answers only from the pages you ticked and the documents you uploaded, so the rule is simple: if a page says it clearly, the bot can say it.

These are the questions the post on the support questions your website should answer by itself lists and explains how to write. If your pages do not answer them yet, start there; a bot cannot deflect a question your website does not answer.

A follow-up such as "and how long does that take to Kerala?" works too, because TailGPT gives the bot the last three exchanges of the conversation.

Which questions can it not deflect, and where should they go?

Anything that depends on one customer's order, account or situation. TailGPT has no order lookups, no access to your shop system or calendar, no live chat inbox and no hand-off to a person. It cannot take an order or a booking, and it collects no names or email addresses. So these stay with you:

• Order status. "Where is my parcel?" needs your courier tracking, not your website.
• Account-specific questions. "Has my refund gone through?", "Can you move my appointment?", "Send me last month's invoice."
• Complaints. A damaged parcel or a bad experience needs a person and, often, an apology.
• Negotiation. "Can you do a better price on ten pieces?" is a decision, not a fact on a page.

The bot still has a job with these: telling the visitor exactly where to go. It can only do that if your website says it. Write a short paragraph on your contact page and make sure that page is ticked:

For questions about an existing order, WhatsApp us on the number below with your order number. We reply Monday to Saturday, 10am to 7pm. For damaged parcels, email us a photo within 48 hours of delivery.

Two things to know about how the bot passes this on. Answers are plain text with no clickable links, so it names the contact page in words rather than linking to it. And it can give a phone or WhatsApp number only if that number is written on a ticked page; a number that appears on none of your pages is removed from the answer.

A visitor sent to the right place is not a deflected question; they will still message you. But they message once, with their order number, instead of asking the bot, then emailing, then calling.

How do you raise the deflection rate?

Fix the pages the bot could not answer from. Tweaking the bot's settings rarely moves the number; writing the missing sentence on your website almost always does.

Two places in the TailGPT dashboard tell you what to fix:

• The count of unanswered questions. The dashboard lists the questions your pages could not answer, with how many times each was asked, beside an answer rate. Watch whether the rate rises or falls from week to week.
• Show chats. On the Chatbots page, Show chats opens every conversation a chatbot has had. Set When to the last 7 days, search for the words from your sheet ("stitching", "COD", "Sunday"), and read what visitors asked and what the bot said. The post on Show chats explains every filter and figure.

For each miss, decide which of three things went wrong: the page exists but is not ticked, the page exists but does not say it clearly, or there is no page. What to do when your chatbot says it does not know goes through each one. Then:

1. Fix the page on your website, or upload a document (PDF, Word or text) that answers it.
2. Re-read the website from the Websites page so the bot sees the change. There is no automatic refresh. As of 4 October 2026, Starter includes 2 re-reads a month and Growth 8, so save the week's fixes and re-read once.
3. Check the same topic in Show chats the following week.
4. Keep the sheet running and compare the next fortnight with your baseline.

In the saree shop's case, the stitching row is the obvious next target. A sizing page that lists blouse measurements in words, not just in an image, is likely to move that row more than anything else the owner could do this month. The full routine for keeping answers current is in how to keep your chatbot's answers up to date, and how it works walks through choosing pages and uploading documents.

What should you not count as success?

Anything the bot reports about itself. Deflection is measured in your inbox, not in the chat window. Five numbers look like success and are not:

• The number of chats. Many people who ask the bubble would never have emailed you. A busy chatbot proves visitors like asking, not that your inbox got lighter.
• Questions used this month. TailGPT counts a question the pages did not cover as one of the month's questions, the same as one it answered well. Only refused questions are free: attempts to give the bot orders, requests to write code, and too many questions too fast. So the usage figure measures traffic, not answers.
• A miss count near zero. It may mean your pages are complete, or simply that few people have asked yet. Read a few conversations in Show chats before you believe it.
• Fewer messages in a fortnight when the bubble was off. If you reached your monthly question limit or a payment failed, the bubble stopped appearing. Your inbox may have been quieter for reasons that have nothing to do with answers. Note any such days on your sheet.
• A shorter inbox with angrier messages. If people are writing in to check what the bot told them, or to complain that it could not help, the count may fall while the work gets harder. Read the messages, not just the tally.

What does count: repeat questions per 100 orders, appointments or enquiries, falling topic by topic, while the questions that remain are the ones that genuinely need you. A receptionist who answers twelve WhatsApp messages a day instead of thirty, and spends the difference on patients, is the result worth having.

If you have not built a chatbot yet, the free trial lets you add your website, have up to 500 pages read and ask 20 test questions in the dashboard before choosing a plan. Run your two-week baseline sheet while you test.