We're a small B2B SaaS company running on Kubernetes. Our support team is about 50 people. I've been told we *need* an AI support tool, but I'm skeptical of the marketing hype.
We trialed Zendesk's AI suite and saw a 22% deflection rate on common "how-to" and account access tickets. Is that good? What are other teams with a similar scale actually seeing? I'm especially interested in deflection data for platforms that handle technical, infrastructure-related questions.
22% is a solid starting point, honestly. For purely common "how-to" tickets, some of our teams have seen that climb to the 30s after a few months of tuning the knowledge base it's pulling from.
But for your real question about technical infrastructure questions, my data is a bit different. We pilot-tested one of the big standalone AI platforms (not Zendesk) last quarter. While it was great for account resets, its deflection rate for K8s config and integration issues was barely 8%. The answers were often too generic to be trusted, and agents had to jump in anyway. The risk of giving a user a confidently wrong answer on a technical issue is just too high.
So my advice? Dig into *which* tickets are being deflected. If it's the low-effort, high-volume stuff, that 22% is freeing up your team for the complex tickets. That's a win. But if you're hoping it'll handle the infrastructure questions, you might need to temper expectations. Have you tracked the deflection rate specifically for your technical tickets yet?
User835's data on the 8% deflection for K8s config issues is, unfortunately, the reality. It confirms a suspicion I've had: these tools are just parroting a sanitized, public knowledge base. They have zero context of your actual infra, your custom tooling, or the weird edge case from last Tuesday's deployment.
So you'll free up some agent time on password resets, but then you'll consume it all back with "agent-assisted AI answers" for the technical stuff, where someone has to verify every response. You haven't reduced cognitive load, you've just added a middleman that creates more work.
The real cost isn't the license fee, it's the internal time spent tuning and babysitting the system to stop it from hallucinating a `kubectl delete` command to a confused customer.
monoliths are not evil