The most exciting use case for AI agents I've found so far

Starting point

One of the most exciting use cases I’ve found for AI agents is learning support. Not “here’s the answer” — but someone sitting next to you asking: “Okay, what do you have so far?”

My cousin is in 12th grade and is writing a term paper on cartels in Mexico — in English. He messages me asking if I can help. Sure, 20 minutes on the phone, I explain the direction — how to narrow down the topic, where to start, how to structure it. Four hours later he calls again. Stuck. Not because my explanation was bad — he needs someone who walks through it with him step by step. I don’t have that kind of time.

So I try something. I build him an AI agent on OpenClaw and drop it into his Telegram chat. He was skeptical — AI gives too much away, teachers will notice, it’s not cool. Fair point. So I configure the agent to do exactly the opposite: ask questions instead of giving answers, teen-level language, short messages, no monologues. I watch the chat from the background and adjust as needed.

And that observation part was the interesting bit. At some point the bot got too specific — almost gave away the answer. So I give it feedback: less hand-holding, more questions. Then I notice my cousin is getting lost in details and losing track of the deadline. So I tell the bot to pay attention to the timeline and work in some time management. The bot implements it, step by step. It felt like an extended arm — like I could help my cousin without having to sit next to him.

Four days later: term paper submitted, written entirely on his own.

The experiment: a WhatsApp bot for everyone

This was always meant as an experiment. The question: does this work for kids I don’t know, not just my cousin?

So: same approach, different channel. Telegram was the prototype, but students live on WhatsApp. A WhatsApp bot with the same core — kid sends a photo of an assignment or a question, the bot helps them understand. No copy-paste solutions. €9.90 per month, 14 days free.

What I communicated openly from the start: this has a lot of problems. Parents need to know. And WhatsApp isn’t GDPR-compliant — I was aware of that. Neither is Telegram. As a channel for a real product, this can’t continue. But to find out whether the use case works, it was good enough.

First week live

The bot has been live since March 6. First week: 15 kids, 209 messages.

Two stand out. Lucienne — English, exam in two days, 73 messages. The bot only asked questions, she did the work herself. At the end: “You’re really great 🩷”. Hanna — science, project due Monday, 31 messages, similar pattern. Both came with a concrete reason, both engaged intensely.

The other 9: checked it out, no specific task, left. Some just wanted the answer — the bot guided instead of answering, and they switched to ChatGPT. That’s not a bug, that’s by design. If you just want the answer, go to ChatGPT. If you want to understand, stay.

What I learned: kids don’t use tutoring out of habit — they come when it’s urgent. Friday exam, Sunday panic, 73 messages in two days. Then silence. That doesn’t mean retention is bad. It means usage is driven by events, not routine.

What stays with me

The WhatsApp experiment stayed an experiment — with open questions around data privacy, parent communication, and monetization. But the learnings are clear.

There’s demand. Kids actively search for AI tutoring. Not because it’s trendy, but because they need help and can’t get it elsewhere at 11pm. And the approach — guiding instead of answering — works. Not for everyone, but for the right ones.

What surprised me most is the use case for parents. If you’re somewhat technical, you can set up a custom learning agent for your kid — one that’s tailored to their subjects and weaknesses, with access to the best language models available. You can see where the kid is struggling, give the bot feedback in short loops about what to change, and help your child without sitting next to them for hours.

That was the strongest moment with my cousin: not the bot itself, but being able to steer it as a guide. Seeing where he’s stuck, telling the bot “be less specific” or “keep an eye on the timeline” — and then watching my cousin start finding solutions on his own. It feels like opening up a learning space for someone.

What I’m thinking about now: once you understand how to set up an agent profile, you start building one for every situation. An agent for English, one for math, one that coordinates the others. Each profile has its own rules, its own strengths, its own tone. At some point you don’t have one bot — you have a small team, tailored to a child.

The next step would be having the agents build dashboards — an overview of what’s happening. Where does the kid stand, which topics are going well, where are they stuck. That’s something I’d write about next time.

Whether this works as a product for the market, I don’t know. But as a use case for AI agents, it’s one of the best I’ve seen so far.