First up: a 9th-grade class. Entrepreneurship. The brief is clear: Keep it simple, make it real and relatable, don’t turn it into a science lecture.

Then Aarhus Business College: E-commerce managers. Same theme as last year (good customer service), but this time with a lot more about AI agents. What’s actually useful in a webshop and what’s just noise.

And finally, on 8 October I’m at E-handelskonferencen in Copenhagen. Same conversation, just a bigger room: Focus on outcome.

Different audiences. I already know which questions I’ll keep coming back to.

What problem are you solving?
What is the system allowed to see and do?
What happens when it’s wrong?

That’s the framework. Whether I’m talking to teenagers or people who run customer service for a living.

Everything else is really secondary and just “tech behind the scenes”.

A lot of people still mix up chatbots and AI agents.

A chatbot answers from a script (or something that feels like one). An AI agent is supposed to understand the question, look things up, maybe take an action, check itself, and then reply… Or hand it to a human when it shouldn’t guess.

That last part matters more than most demos admit.

If your only KPI is “how many tickets did the AI close?”, you’re optimising for deflection. Not for whether the customer actually got a correct answer. And customers can smell that from a mile away.

When I teach this to ecommerce people, I spend more time on guardrails and escalation than on “wow, look what it can write”. Because the hard part isn’t generating text. The hard part is knowing when not to.

Quick side note on our own setup, because it always comes up…

We left OpenAI and moved to self-hosted GPUs instead. The cost was getting stupid, GDPR-sensitive customers didn’t want them as a sub-processor, and I honestly just prefer running things ourselves. So we moved to open-weight models on GPUs in Denmark. Same product surface for customers. Different stack underneath.

I’ve written about that before, so I’m not going to rehash the whole thing here.

The only point that belongs in this post: “who runs the model” is part of question two: what is the system allowed to see and do. If you care about that for your customers, you should care about it for your vendors too.

Anyway. Back to teaching…

With the 9th graders, I won’t start with neural nets or parameter counts. I’ll start with: here’s a boring problem a real business has. How would you solve it? Where would you use a tool? Where do you still need a person?

Same energy with the college students, just with returns, Trustpilot, shipping delays, and angry emails instead of school-project metaphors.

If you can’t explain the use case without saying “AI” every second sentence, you probably don’t have a use case yet. You’ve got a buzzword looking for a job.

This also applies for professional e-commercers. “AI” isn’t a silver bullet that fixes everything. We need a first principles approach to solving these problems.

Which brings me to E-handelskonferencen.

I’ll be there on 8 October on stage with Emil Nissen from Blandselvfrø.dk. We’re talking about how they’ve grown into 10 markets with a Trustscore around 4.8, and basically one person in customer service.

The point of that session isn’t “look how many tickets AI can close by itself.”

We’re looking at good ecommerce from a broader angle: product, policies, website, shipping, communication - the whole journey. Customer service sits at the end of that chain. If something is broken earlier, it shows up as tickets.

A 90% AI resolution rate sounds impressive on a slide. Sometimes it is. Often it just means a big chunk of the questions were so simple that an agent could answer them without breaking a sweat… Tracking links, “where is my order”, size guides, return rules, the same FAQ on loop.

Sure, they come from time to time, but if that’s your reality, the interesting question isn’t “how do we push that from 90% to 95%?”
It’s: why the hell are customers asking these things in the first place?

Maybe the shipping page is unclear. Maybe the confirmation email is missing the one detail everyone needs. Maybe the return flow is confusing. Maybe the product page doesn’t answer the question that shows up in every second chat.

Good customer service isn’t only answering faster. It’s part of it, for sure, but it isn’t everything. It’s also removing the need to ask. AI can help with both, but if you only measure “tickets solved by the bot”, you celebrate the symptom and ignore the disease.

That’s the conversation I want to have on stage with Emil. Concrete shop. Real constraints. Not an AI theatre.

Jimmy and I will also be at the conference the whole day. We don’t have a booth. We’ll just be walking around in Herodesk merch (basically a standing (and walking) booth, lol). Should be hard to miss, so c ome say hi if you want to talk CS, AI agents, or why your “amazing” resolution rate might actually be a product problem in disguise.

If you’re not going, the three questions still work from your desk.

Start there. The hype can wait.

We’ve gotten a lot of feedback on the “Got Your Back”-Release. More than 10% of our users have already upgraded to Herodesk Pro. So most of our time has been spent in minor fixes, improvements and updates.

But, we did release one big, new thing: A new Uni-tel phone integration

If you’re offering phone support to your customers, you should really check out this new integration! It’s awesome.

And last week, we celebrated a big milestone for us: More than 1.000 active, paying Herodesk users 🎉

Below are some of my favourite newsletters. So if you like this one, you should consider subscribing to these, too.