What a cold outreach campaign built and run entirely by AI agents gets right and wrong

I let a chain of agents write, send, and follow up on a cold outreach campaign for two weeks with almost no human editing. I wanted to know what breaks first when you remove the person from the process. Some of it worked better than I expected. Some of it embarrassed me.

What the agents got right

Volume and consistency are where agents win, no argument. They wrote personalized opening lines pulled from a prospect's website, sent at the right time zone, and never forgot to follow up. A human doing this manually gets tired by message forty and starts copying and pasting. The agent never got tired. Every single lead got the same quality of attention as the first one.

The research step was genuinely strong. Before writing anything, the agent pulled recent news, funding rounds, and product launches for each company and used that as the hook. That is the part that used to take me an hour per prospect. Now it takes seconds and reads almost as well as something I would write myself.

What went wrong

The agent could not tell the difference between a company that looked like a good fit on paper and one that actually was. It targeted a logistics company that had just laid off its entire ops team, technically a match for our automation pitch, practically the worst possible time to pitch them anything. A human would have paused. The agent did not pause, because nothing told it to.

It also got repetitive in a way that was hard to catch. Individual messages read fine, but read fifty of them in a row and you notice the same three sentence structures over and over. Prospects on the same email thread as their coworkers noticed too. Sameness at scale is its own kind of tell.

And it was slow to recognize a soft no. Someone would reply "not right now, check back later" and the agent treated that as a fresh lead rather than a signal to back off for a while.

How we run this at Esipick

We use agents for outreach ourselves, but never unsupervised end to end. Research and drafting are fully agent driven now, that part earned its trust. Anything that touches judgment, whether a prospect is actually a good fit, whether the timing is right, whether a reply means yes or a polite no, still gets a human check before it goes out. That is not caution for its own sake. It is the actual line we have found between where agents save real time and where they create real risk.

The honest takeaway is that agents are excellent at doing the same good thing a thousand times and bad at noticing when this one time is different. Build your outreach stack around that, not around the idea that autonomy means better.