Jason M. Lemkin argues that AI support tools fail without a strong knowledge base and deep integration. A Zendesk agent couldn't recognize a chair was already delivered, showing why generic AI layers don't work. Resolution rates hit 60-70%+ only when the system is trained on your actual data.
Chapters
I agree. Not only that, my advice is don't overanalyze it. Close the customer.
Yeah.
Too much talk about pricing, not enough talk about great product. It'll figure itself out. And even if Elias is trying to figure out, it doesn't, he can change his model in a year or two. Yeah.
Who cares? And that's the starting point. So when we look at our customers who are getting like 60, 70% resolution, with the agent, it's because they had a really good knowledge base.
Yeah. Yeah.
Because you got to train it. And if you have an awesome knowledge base, it's well trained, it can answer it. But there's a lot to the craft. Like you've got to upload that knowledge base. You've got to— it's going to get smart on that. Then you've got to have it listening to every support call that's going on, be able to turn it up and down on your best support reps over time, and that'll get better.
Yeah, I really, we're still all learning. I'm not a HubSpot support agent in 2025 with all of the HubSpot data. I'm not convinced the best standalone AI agent can beat it. I am convinced they can build, they already can build better technology. And I went through this with a company where they graded AI agents from technology versus implementation. And okay, we're not going to catch the one with standalone, the technology. But when I'm implementing with all of our data and the training, it's like, it's better, right? That's the meta question, right? Can it, right?
And I just, I look at agency and what Elias is building for CSMs.
Every day I get an email from Zendesk, the Zendesk implementation. It's not Zendesk's fault, but it's the implementation. We're still trying to track down your chair. I'm just watching this to see how it gets resolved. And nominally they have AI, they actually build their own support. They have Zendesk and then they have their own thing they built on OpenAI, but this cobbled together thing still doesn't know the chair was delivered. It still doesn't. Every day I just got an email today and I don't mean to be mean. I don't want to get two chairs, but I want to see what happens because the data's there, right? But their AI plus Zendesk can't connect at all. They don't know that the chair was already delivered to the fixed address.
It's just not completely there. It's amazing compared to what, 6 months ago. I can't wait to see what it's like 6 months from now, but it's all just, it's early, super early.
