Why I stopped building a chatbot and built a secretary

I started out building a chatbot.

You know the type — ask it a question, get an answer. Clever responses, webhook integrations, the whole "AI assistant" thing. It was technically impressive. People said "cool demo."

But nobody actually used it.

The Problem with Chatbots

Here's what I realized: People don't want to chat with AI.

Chat is a medium, not a value proposition. When you open ChatGPT, you're not there because you love the chat interface — you're there because you need something done and this is the best tool available.

But chat has friction:

• You have to remember to open the app
• You have to formulate the right prompt
• You have to engage in back-and-forth to get what you need
• And when you're done, you have to do something with that information

Most chatbots are just search engines with better UI. They answer questions, they don't solve problems.

The Delegation Gap

What people actually want is someone — or something — they can delegate to.

Think about how you actually use an assistant (if you're lucky enough to have one):

• "Find me the cheapest flight to SF next week"
• "Reschedule my Thursday meetings"
• "Remind me to follow up with Sarah if she hasn't replied"

You don't want a conversation. You want the task done.

This is the gap most AI tools miss. They're built by engineers who love the technology (guilty) rather than by people solving real workflow problems.

The Shift to "Jobs"

So I rebuilt from scratch. Instead of a chatbot, I built an AI secretary that lives in WhatsApp.

The core concept: Jobs — recurring delegations that run autonomously.

Instead of asking "what's on my calendar today?" every morning, you set up a Job once: "Send me a daily brief every morning at 8am with my schedule, priorities, and any urgent emails."

Instead of manually triaging your inbox, you delegate: "Label and prioritize my emails. Surface only what needs my attention."

The difference? Set once, runs forever. No daily engagement required. No willpower drain. Just results.

Why WhatsApp?

Two reasons:

1. Zero friction
No app download. No new login. No onboarding flow. If you can text, you can use it. My mom can use it. My tech-averse uncle can use it.

2. Context switching is the enemy
People check WhatsApp 50+ times a day. Meet them where they already are. The moat isn't the LLM (everyone has that now). It's the workflow integration.

Technical Architecture (Briefly)

• WhatsApp Business API for delivery
• FastAPI backend
• Postgres + pgvector for memory
• GPT-4o for reasoning
• Structured "Job" primitives instead of open-ended chat context

The key technical insight: scoped context windows. Each Job has its own isolated context. No quadratic token growth. No "infinite chat history" problems. Predictable costs, predictable behavior.

The Psychology of Delegation

There's an interesting psychological barrier to delegation — people feel guilty asking others to do things they "should" do themselves. This is why people struggle with virtual assistants too.

But with AI? No social cost. You can delegate the smallest, most repetitive tasks without feeling like a burden. "Remind me to call mom every Sunday at 3pm" — you'd never hire a human for that. But it's exactly the kind of thing that slips through the cracks.

Results So Far

Early users who set up 5+ Jobs in their first week stick. Under that, they churn. It's the "habit replacement threshold" — once the AI is handling enough of your mental load that you'd miss it, you're locked in.

Retention curves look very different from chat-based AI tools. We're not fighting the "novelty wears off" problem because we never relied on novelty. We rely on utility.

Lessons for Builders

If you're building with AI:

1. Start with the outcome, not the interface. Chat is a means, not an end.
2. Reduce cognitive load, don't add to it. Every interaction should save more mental energy than it costs.
3. Meet users in existing workflows. The best UI is the one they already know.
4. Scope narrowly. Broadly capable agents are expensive and unpredictable. Narrow ones are profitable and reliable.

What's Next

Currently working on the killer feature: autonomous meeting booking. "Find time with John next week and send the invite." That's the delegation holy grail — when the AI doesn't just remind you, it actually does the thing.

---

Building irel.ai — an AI secretary on WhatsApp. Bootstrapped, solo founder, figuring it out in public.

What would you delegate if you had an AI that actually did things instead of just chatting?