The short answer
The fastest way to reduce support tickets is to stop causing them — fix the three pages that generate most of your repeat questions, then deflect what is left. An AI assistant helps twice over: it answers the repetitive questions immediately, and the transcripts tell you exactly which pages are failing. Deflection rate on its own is a misleading metric, because an assistant that answers everything confidently and wrongly scores brilliantly on it.
Every support team has the same distribution: a small number of question types generate most of the volume. The work is finding which, and being honest that most of them exist because something upstream is unclear.
Start by reading, not by deploying
Take the last two hundred tickets and sort them by what the person actually wanted. Not by tag — tags are chosen by whoever closed the ticket and reflect the answer rather than the question. You will end up with something like:
- Answerable from existing content. The answer is on your site and they did not find it. Pure deflection candidates.
- Answerable, but the content does not exist. The question is reasonable and nobody wrote the page.
- Caused by a broken thing. A confusing checkout, a misleading email, a form that fails silently. No amount of answering fixes these.
- Genuinely needs a person. Judgement, an exception, an upset customer.
Only the first two are deflection work. The third is a product bug wearing a support costume, and the fourth should never be deflected at all.
Fix the causes first, because it is cheaper
If forty tickets a week ask where an order is, the answer is not a chatbot that recites the tracking policy — it is a tracking link in the dispatch email. Deflecting a question that should not have been asked is a worse outcome than eliminating it, and it costs more.
Do the causes you can fix in a week first. Then deflect.
Where an assistant genuinely earns its place
- The long tail. Fifty questions asked twice a month each are not worth a help-centre article but are trivially answerable from content you already have.
- Out of hours. A question answered at 11pm is a question that does not arrive as an email at 9am, and it is answered while the person still cares.
- Search that actually works. Most site search is keyword matching. Someone asking “can I bring my dog” needs to find a paragraph that says “household pets are welcome”, and only retrieval by meaning does that.
- Triage. Collecting the order number and the problem before a human arrives saves the first two messages of every conversation.
The metric problem
Deflection rate — conversations that did not become tickets — is the number every vendor reports, including us. It is also trivially gamed, and not deliberately: an assistant that answers confidently and wrongly produces a spectacular deflection rate, because the customer goes away. They come back three days later, angrier, and that ticket gets counted as a new one.
Watch these instead, together:
- Ticket volume on the specific topics you targeted. The only number that directly measures the thing you wanted.
- Repeat contact rate. Did the same person come back within a week? This is what catches confident wrong answers.
- Escalation rate, and whether escalations resolve faster than cold tickets. If they do not, handover is broken.
- The unanswered list. Below.
The most valuable output is a list of your own failures
Every week, the questions the assistant could not answer are a ranked list of content that should exist and does not. It is better than any keyword tool, because these are real questions from real prospects in their own words, at the moment they were deciding.
Work down it. Each page you write removes a question class permanently, improves the assistant, and — as it happens — gives you a page that can rank and be cited.
This is awkward for us to recommend. We bill by conversation, and acting on that list reduces conversations. We would rather say it than pretend the incentive does not exist.
What not to do
- Do not hide the route to a human. It converts a resolvable problem into a complaint, and in India it is close to the “forced action” dark pattern regulators now name explicitly.
- Do not deflect angry people. An assistant cannot tell that a question matters enormously to the person asking. Route frustration to a person early.
- Do not let it guess. Configure it to say when it does not know. One honest “I do not have that” is worth ten plausible paragraphs.
- Do not launch and leave. The unanswered list is only valuable if somebody reads it.
A realistic first month
- Week 1. Categorise two hundred tickets. Fix the two worst causes.
- Week 2. Deploy against your existing content. Read every transcript.
- Week 3. Write the top five missing answers from the unanswered list.
- Week 4. Compare topic-level ticket volume against the month before, and check repeat contact rate has not risen.
If you want the mechanics of getting one running, how to add an AI chatbot to your website covers the setup and the four decisions that matter.
