How to Reduce Support Tickets With AI (What Actually Works)
Deflection is the wrong target. The five ticket types AI absorbs well, the ones it must never touch, and the bigger win: tickets that should not exist at all.
There are two ways to reduce support tickets, and only one of them gets discussed. The first is answering more questions without a human: that’s the chatbot everybody thinks of. The second is removing the reason the ticket existed at all, which is less exciting and usually worth more. A good support automation project does both, and it starts by reading your tickets rather than by choosing a tool.
Read fifty tickets before you build anything
Export the last fifty or a hundred conversations and sort them by what the person actually wanted. Not by product area — by intent. You will almost always find that a small number of intents account for most of the volume, and that they are duller than you expected: where is my order, how do I change this setting, do you support X, how do I cancel.
That list is simultaneously your build specification and your test set. Keep it: once the bot exists, those same questions are the regression test you re-run every time you change the knowledge base. Guessing at what customers ask produces a bot that answers questions nobody has, which is the most common way these projects quietly fail. There’s more on drawing that boundary in what an AI chatbot can and can’t answer.
The five categories AI absorbs well
- Status lookups. Where is my order, when does my subscription renew, has my payment gone through. These are the highest-volume tickets in most businesses and the least valuable use of a human. They need the bot connected to real data, not just to your help articles — a bot that can only quote the shipping policy is useless here.
- How-do-I questions. Everything already documented somewhere nobody reads. A bot is a much better search interface for your own help centre than your help centre’s search box.
- Policy and eligibility. Returns windows, delivery areas, what’s included in a plan. Factual, stable, easy to keep correct — provided there is exactly one source of truth and the bot reads it.
- Pre-sales questions. Often filed as support, frequently the most commercially valuable thing in the queue. Answering these instantly, at 2am, in the visitor’s language, is where a support bot starts paying for itself twice.
- Triage. Even when the bot shouldn’t answer, it can classify, tag, gather the missing details and route to the right person. A ticket that arrives complete saves a human two round-trips, which is a real saving even though your ticket count doesn’t move.
The ones it must not touch
Refunds, billing disputes, complaints, cancellations where the person is upset, anything medical or legal, and anything where being confidently wrong costs real money. The test I use is simple: if a wrong answer would require an apology from a human, a human should have given the answer. Emotional conversations are also a hard no — not because the model can’t produce sympathetic sentences, but because a customer who realises they explained their frustration to a robot is angrier than when they started.
The corollary matters as much: the bot must be able to say “I don’t know” and hand over cleanly, carrying the full conversation with it so nobody has to repeat themselves. A bot that admits uncertainty and passes you to a person is trusted. One that bluffs gets screenshotted.
The bigger win: tickets that shouldn’t exist
When you sort those fifty tickets, some clusters are not questions at all — they are symptoms. Thirty “where is my order” messages a week usually mean your shipping confirmation is unclear, arrives too late, or doesn’t include a tracking link. Answering them faster with AI treats the symptom forever. Fixing the confirmation email makes them stop.
This is the part that gets skipped because it isn’t an AI project, and it is routinely the larger number. A proactive notification, a clearer form label, a status page, an onboarding email that answers the question before it’s asked — these remove tickets permanently instead of absorbing them. In eCommerce especially, most of the support load is downstream of order plumbing. Do this pass first; whatever survives it is the honest scope for the bot.
Deflection rate is a vanity metric
“The bot handled 68% of conversations” tells you almost nothing. It counts conversations that ended, and conversations end for bad reasons too — people give up. I’d watch three numbers instead: how many tickets still needed a human, how satisfied people were after a handoff, and how often the bot answered something it should have escalated. That third one only shows up if somebody reads transcripts every week, which is the least glamorous and most valuable habit in the whole project.
Read the failures, not the successes. Every question the bot got wrong is a missing piece of your knowledge base, and the fix takes ten minutes.
What it costs
A custom chatbot trained on your own content starts around $900, and the version wired into live data with human handoff and a review workflow is $2,800-tier work, with 30 days of support after launch — see pricing for the detail. Running costs are usually the surprise: for most businesses the model bill lands in the low tens of dollars a month rather than the hundreds, and the worked token arithmetic is here. That’s the structural difference from a platform that bills you per resolution — the comparison with Intercom and Tidio goes through it properly, including the cases where you should just use theirs.
The chat bubble in the corner of this site is the same thing I build for clients, handing over to me when someone asks for a human. Ask it something awkward. Or book a free intro call and bring your last fifty tickets — that conversation is more useful than any demo.
Key takeaways
- Sort fifty real tickets by intent before choosing anything. That list is both the spec and the regression test.
- Status lookups, how-do-I, policy, pre-sales and triage are the safe wins. Money, complaints and emotion are not.
- Some tickets are symptoms. Fixing the cause removes them permanently; AI only absorbs them.
- Ignore deflection rate. Track tickets that still needed a human, satisfaction after handoff, and wrong answers caught in transcripts.
