It’s the question every CX leader is asking right now, usually with a mix of hope and suspicion: can AI actually handle my customer support tickets, or is it going to hallucinate its way through my most important customer relationships?

Here’s the honest answer, and it’s the same one I give every leader who asks me over coffee: yes — but only if you plan it properly and run it on a structured framework. AI is not a switch you flip. Deployed carelessly, it will happily compound bad customer experiences faster than any human team ever could. Deployed properly, it will quietly take a huge chunk of work off your team’s plate and free them to do things that actually grow the business.

At Gravity CX we’ve now deployed AI capabilities for over 50 Australian and New Zealand companies running on Zendesk. So this isn’t theory. It’s what we see working, and failing, in real support operations every week. Let me walk you through what “properly” actually looks like.

First, what does “handle a ticket” even mean?

Part of the reason this question is so hard to answer is that “AI handling a ticket” isn’t one thing. An AI agent really does three things in sequence: it understands the customer’s intent, decides on the best next step, and acts on that decision. Sometimes acting means drafting a reply for a human to approve. Sometimes it means resolving the whole conversation end-to-end without a ticket ever landing in a queue.

That distinction matters, because the realistic answer to “can AI handle my tickets?” is: it depends on which tickets, and how much autonomy you’ve earned. AI is brilliant at the high-volume, well-understood questions — “where’s my order,” “how do I reset this,” “what’s your returns policy.” It is not, and should not be, left alone on the ambiguous, high-stakes, emotionally charged ones. The skill is knowing the difference and building your rollout around it.

The proof: 2,000+ conversations a month, and three people moved into sales

Let me make this concrete with a client. We deployed a Zendesk AI Agent for The Bart Group that now handles more than 2,000 conversations every month — answering product questions and helping customers find answers on their own, autonomously.

The interesting part isn’t the deflection number. It’s what it let them do with their people. Because the AI was reliably absorbing that volume, The Bart Group moved three full-time support team members into sales roles. Nobody was replaced. The work that didn’t need a human anymore stopped needing one, and those humans went and did something more valuable for the business.

That’s the outcome I want leaders to sit with, because it reframes the whole question. The point of AI in support isn’t to shrink your team. It’s to change what your team spends its day on.

---

Thinking about where AI fits in your own support operation? We’ve packaged everything we’ve learned across 50+ Zendesk deployments into the Gravity CX AI Playbook — the two-play strategy, the four-phase rollout, and the checklists our own team uses. Download the AI Playbook →

---

Why most AI support projects underperform

If AI works this well, why do so many rollouts disappoint?

In my experience, the number one reason projects underperform is not the technology, the model, or the vendor. It’s that the team has nobody who owns the AI. Most companies deploy AI as a project, switch it on, and move on. What they actually need is a dedicated AI architect — someone whose job is to deploy it, monitor it, and optimise it constantly.

AI in support is not “set and forget.” Every “I couldn’t answer that” is a knowledge gap to fix. Every rejected draft is a signal about what the AI is getting wrong. Every new product or policy is something the AI needs to learn. Without someone owning that loop, quality quietly degrades, gaps never get closed, and six months later everyone concludes “AI didn’t work for us” — when really, nobody was steering it.

If you take one thing from this article: budget for the person, not just the platform.

The mistake that’s genuinely hard to recover from

The other failure mode is more dangerous, because it’s hard to undo.

AI should be deployed in phases, and each optimisation needs to be monitored before you move on to the next. This sounds like cautious best-practice advice, but there’s a specific, painful reason for it. If you turn on too much at once and something goes wrong, you can compound a poor customer experience at scale before you notice — and then you’re stuck trying to work out which change caused it. Was it the new knowledge source? The autonomy you granted on that intent? The workflow you connected? When everything went live together, you can’t isolate the root cause, and rolling back cleanly becomes very hard.

Phasing isn’t about being timid. It’s about always knowing exactly what changed, so that when something misbehaves you can fix that one thing instead of ripping the whole deployment out. The rollouts that fail most spectacularly are almost always the ones that skipped the supervised middle and jumped straight to “AI replies on its own.”

The right sequence looks like this:

• Phase 1 — Productivity. Summaries, intelligent triage, drafted replies. Low risk, immediate gains. Your team stays in control and gets comfortable supervising the AI.
• Phase 2 — Supervision. AI proposes responses and backend updates; a human approves before anything goes out. This is where you learn what your AI is genuinely good at — and it’s the phase everyone is tempted to skip.
• Phase 3 — Deflection. Autonomous resolution on your top intents and highest-volume channels, once you have the data to trust it. This is where The Bart Group lives.
• Phase 4 — Complexity. Cross-system workflows that reach beyond support into sales and ops.

You earn autonomy by proving reliability under supervision first. Never the other way around.

What has to be true before any of this works

AI compounds whatever foundation you put it on — good or bad. Before you scale, three things have to be in place:

Structured, current knowledge. This is the single biggest determinant of AI quality and the most underestimated. AI can read a PDF, but accuracy is meaningfully better with proper, text-based help articles — especially where every word matters, like pricing and policy. Your AI’s “I don’t know” reports are your knowledge backlog. The most efficient way to improve AI accuracy is almost never a model change; it’s a knowledge fix.

Clear governance. For each intent, decide up front whether the AI is assist-only or allowed to resolve autonomously, and write down why. Gate permissions by action, not just intent — an AI can issue a refund under a dollar threshold automatically, but anything above it goes to a human. Keep audit logs so you can answer “why did the AI do that?” months later, and test your kill switch before you need it.

A unified platform. If voice lives in one tool, email in another, and chat in a third, your AI is working with fragments — and so are your agents. Cross-channel context is what lets AI handle a customer who started in chat and followed up by email. This is why we build on Zendesk, with Zendesk Copilot and the Zendesk AI Agent: one platform, one view of the conversation, which is non-negotiable for the deflection play.

The objections I hear most (and my honest answers)

“AI will hallucinate and give customers wrong answers.” It can — if you point it at messy knowledge and give it autonomy it hasn’t earned. That’s exactly what governance, confidence thresholds, and the supervised Phase 2 are for. You watch it draft under human approval until you have hard evidence it’s reliable on a given intent, then you let it run. Trust is granted with data, not hope.

“My customers hate bots.” Customers hate bad bots — the ones that trap them in a loop and can’t answer the question. What they actually want is a fast, correct answer at 11pm without waiting in a queue. A well-deployed AI agent, resolving real questions accurately on the intents it’s good at, consistently matches or beats human CSAT in our deployments. The bot isn’t the problem. The lazy bot is.

“It’s going to replace my team.” This is the one I care most about correcting. Look at The Bart Group again: nobody lost a job. Three people moved up, into revenue-generating roles. Your frontline agents’ work shifts from “answer the ticket” to “supervise the AI that drafted the answer” — a genuine skill upgrade — and the capacity you free up gets reinvested in the work that actually needs a human.

So — can AI really handle your tickets?

Yes. AI can absolutely handle a meaningful share of your customer support tickets, and for some teams that share is large enough to change the shape of the whole operation. But it will only do that if you treat it as a discipline rather than a switch: structured knowledge, clear governance, one unified platform, a phased rollout you monitor at every step, and a dedicated person who owns the AI’s performance over time.

Get those right and the question stops being “can AI handle my tickets?” It becomes “what should my team be doing now that it doesn’t have to?”

That’s a much better question to be asking — and it’s the one our best clients are living in today.

---

Ready to plan your own rollout?

Everything above — the two complementary plays, the four-phase model, the foundations, and the exact checklists our team works from — is laid out step by step in the Gravity CX AI Playbook. It’s built from real deployments across 50+ ANZ Zendesk teams, and it’s designed to be copied straight into your project tracker.

Download the AI Playbook →

Gravity CX helps mid-market and enterprise CX teams roll out AI on Zendesk without breaking customer trust. Want to talk through your rollout? Get in touch.

Get your AI playbook