Tjaša Kravanja ·
I Built AI Agents to Do My Job. They Can't.
Here's what broke, what didn't, and what I'd never let anywhere near automation .... and the answers are messier than anyone on LinkedIn and X will admit.
I'm not here to tell you AI is replacing marketers. I'm telling you what I actually built, what works, because I've been running this since OpenClaw came out and the answers are messier than anyone on LinkedIn and X will admit.
A bit of context
I run GTM, product marketing, and growth across Web3, iGaming, and AI products simultaneously. The kind of work where you're managing hundreds of creators, building launch campaigns, and trying to move users from signup to actually doing something, all at once.
What I actually automated
I built an internal tool (KOL Hub) that manages ~7,700 creators.
On top of that, I run two agents.
One handles sourcing - pulling profile data and keywords tagged to each creator, so I can find the right people through an AI prompt search instead of digging manually for hours.
The other handles outbound - drafts messages, sends them, tracks replies, and routes anything important back to me.
What used to take a team days now runs overnight.
I also use Lovable to build product prototypes and landing pages end-to-end. Full information architecture, UX, copy, everything. The dev team uses them as direct references instead of waiting on Figma handoffs.
That replaced work that used to take a team.
And it's a fraction of the cost of hiring people to do the same thing.
What actually broke
1) The research agent lied to me. A lot. It kept confidently telling me accounts were real that weren't. So I built another agent specifically to check the first agent's work. The results are still not great. You need a human QA layer regardless or you'll end up paying creators who don't exist.
2) The outreach agent stops being useful after message one
It writes a decent first message and falls apart after that. No memory of the relationship, no read on tone, no sense of when to back off. I also have a separate agent that reminds the outreach agent to follow up on emails, which tells you everything about how reliable it is on its own
3) Lovable prototypes are a communication tool, not a product.
They look great in a demo. They fall apart the moment you need real state management. I use them to align with the dev team, not to ship.
What I would never automate
Negotiation. Knowing when someone is lying about their numbers. Reading the room.
Deciding what the product actually is. Taste.
And more importantly, understanding human psychology, how people move through it, where they drop off, what makes them stay. That's not pattern recognition, it's judgement built over time
That part still needs a person.
The honest takeaway
The part of my job AI made easier is the part I didn't want to do anyway.
The part that actually matters, positioning, judgement, taste, the reason a campaign works instead of just running, is more valuable now, not less.
The growth operators who get replaced in the next two years are the ones whose job was moving data between tabs. The ones who get paid more are the ones who can tell the agents what's actually worth doing.
The tool changed. The thinking didn't.
I still decide what to build, who to target, and why anyone should care. I just don't manually scrape Twitter for KOLs at 2am anymore.
Read the original article on X ↗
Republished from the original X article, published 15 April 2026. Product examples and numerical claims reflect the original publication.