AR · July 2026 · 8 min read
Ticket volume falls for three reasons. Only two of them are good.
Which ticket types actually automate, how much to expect, and the metric that hides whether it worked.
A support queue that halved because the AI worked looks identical, on every dashboard you own, to one that halved because customers stopped bothering.
Three mechanisms, one number
Volume drops when the AI resolves something, when the customer self-serves, and when the customer gives up. Deflection metrics cannot tell them apart, which is why the number is popular with vendors and unhelpful to you.
What actually automates
The reductions come from a narrow band of ticket types. Knowing which lets you predict your own result better than any vendor benchmark.
- Order and delivery status. Highest volume in retail, close to fully automatable if the agent can query the order system.
- Password and account access. High volume, low complexity, but the exceptions are exactly what reaches support.
- Policy questions. Returns windows, shipping times, what is covered.
- Onboarding and setup. Predictable, documented, and expensive to leave unanswered.
If your queue is mostly these, a high deflection figure is plausible. If it is technical diagnosis or account-specific disputes, treat any headline number as marketing.
What the vendors claim
| Platform | Claim | What it counts |
|---|---|---|
| MavenAGI | 93% | Questions answered autonomously |
| Ada | 83% | Support queries resolved |
| Freshdesk Freddy | 80% | Routine tickets specifically |
| Tidio Lyro | 67% | Common customer questions |
| Gorgias | 60% | Repetitive support tasks |
| Intercom Fin | 30-50% | Conversations resolved |
A threefold spread for what sounds like one capability. Only Freshworks qualifies what it is counting, which is worth noticing.
Five things that reduce volume before you buy anything
- Fix the top thirty help articles. Most deflection comes from content, not from the model.
- Add an order status page. Removes the single largest ticket category without any AI at all.
- Fix the email that generates the most replies. Usually a shipping or billing notification.
- Publish your returns policy where people look for it.
- Add a status page if you have outages. Incident tickets spike and are entirely deflectable.
Teams that do these first get more from a cheap agent than teams that skip them get from an expensive one.
The cost inversion
Removing tickets is supposed to save money. On per-resolution pricing every ticket the AI handles is billable, so cutting 10,000 tickets at $0.99 costs $9,900. The saving exists only if that is less than the human time it replaced, and at volume it frequently is not.
Every percentage above is a published vendor claim, quoted as such. We do not benchmark platforms.
Frequently asked
How much can AI reduce ticket volume?
Published claims run 30% to 93%, measuring different events. Realistically, expect reductions concentrated in order status, access issues and policy questions.
What is a good ticket deflection rate?
The number is close to meaningless without knowing what triggers it. Ask instead for resolution confirmed by the customer.
What is the difference between deflection and resolution?
Deflection means the ticket did not reach a human, which includes customers who gave up. Resolution means the problem was solved.
Which tickets does AI remove best?
Order status, password and access issues, policy questions and onboarding. Documented, repetitive, system-answerable.
Can I reduce tickets without buying AI?
Yes, and you should try first. Fixing your top help articles, adding an order status page and rewriting the email that generates the most replies all reduce volume for free.
Does lower ticket volume mean happier customers?
Not necessarily. Volume falls when customers give up as readily as when they are helped, and deflection metrics do not distinguish the two.
How do I measure whether it worked?
Track confirmed resolution separately from deflection. The gap between them is your real error rate.
Why did my costs go up when tickets went down?
Per-resolution pricing. Every ticket the AI handled was billable, so removing them from the human queue moved the cost rather than eliminating it.
How long before ticket volume drops?
Deflection on documented questions shows within days. Anything requiring integrations or content work takes weeks.
Should I aim for 90% deflection?
No. Optimising hard for deflection produces excellent numbers and quietly loses customers who could not reach anybody.
Tools mentioned
Full reviews, pricing tiers and where each one breaks.
Fin by Intercom
The most polished autonomous agent on the market, attached to the pricing model buyers complain about most.
Ada
Multi-channel agent that predates the current wave, with the enterprise footprint that implies.
Tidio Lyro
One of the few genuinely cheap autonomous agents, bundled with live chat and a basic help desk.
Gorgias
The Shopify-native default, with the deepest order, return and refund actions in this category.
You can also look into
How to Set Up AI Customer Service - Complete Guide (2026)
The order of operations that decides whether your AI agent helps or just annoys people into leaving.
How AI agents reduce ticket volume, and the number that hides whether it worked
Vendors report ticket volume reductions between 30% and 93%. The spread is that wide because a ticket that never reaches you is not the same as a customer you helped.
How to reduce customer support costs with AI (and the pricing model that undoes it)
Every guide on this topic promises 30-40% savings. None of them mention that the most common AI pricing model charges you more the better the AI works.
Somebody claims 93% autonomous resolution. Here is what that number can hide.
Vendors publish autonomy rates between 30% and 93%. The range is that wide because they are not measuring the same thing.
WRITTEN BY AR · UPDATED 2026-07-29
I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.