I was building AI agents for clients and every one of them needed live LinkedIn data. The normal ways to get it are renting somebody’s LinkedIn account or running scrapers that break, and both quietly put the account at risk. I wanted that layer to be boring and legitimate instead, so I built it. ZOOQ is an API and MCP server that hands agents live profiles, companies, jobs and posts, billed by credit. The genuinely different part is that my customer is usually not a person, it is an agent that decides on its own what to fetch next. The product is built around that reality: when an upstream call fails, the credits refund automatically, because an agent cannot email support and argue about it. Eleven beta testers from around the world signed up within one day of the beta opening, which told me the problem was not just mine. I run the whole company alone, and agents I built do most of the marketing.

Almost everything comes from LinkedIn today. I post near daily, grew from 3,500 to 7,500 followers in about a year, and I run a community of about 300 senior marketing leaders that started at 15 members. My outbound is personalized by agents, and in my own A/B test the personalized campaign got 28% replies against a 19% baseline over about 200 sends. The experiment I keep coming back to is ads inside AI chat tools. I paid about 6 cents a click on US traffic there, and signups came in roughly half the cost of Facebook. That is where I wish more of it came from, and I think it will. My product ships as an MCP server, so the honest answer is I want agents finding it, not just people.

Handed over: cold outreach personalization. An agent reads the lead and writes the first line, and in my own data personalized first lines get 48% more replies than generic ones. I also handed over most of my content production and my community onboarding, which runs on a WhatsApp bot that costs about 34 shekels a month. What went back to being human work is the judgment on top. My reply rate fell from 30% to 21% and I did not notice for weeks, because targeting had broken quietly and the agent kept sending politely into the wrong audience. Nothing crashed, so nothing alerted me. A model upgrade fixed the targeting and the rate recovered to about 30%, but since then the sending stays automated and the noticing is my job again.

Some of it, yes. I sell to people building AI agents, so my buyers already live inside these tools. I tested it directly and bought ads inside AI chat tools: about 6 cents a click on US traffic, with signups roughly 50% cheaper than Facebook. Those are my own ad account numbers, not a study. My rough field estimate from a year of talking to marketers: about 10% of people know what an AI agent is, 5% pay for one, and half a percent build with them. The channel is early, which is exactly why the clicks are cheap. The future of advertising is showing up inside the chat.

Do: pick one process that already has a written runbook and hand it to one well-built agent. One agent does about 90% of what people try to do with twenty, in fewer tokens and usually better. Give it real autonomy, because if it still waits for you to open a chat window, it is a fancy calculator. Budget for the failure rate honestly too. Automation at scale means roughly 1% of it goes wrong, and you should decide in advance whether you can live with that. Ignore: anyone promising five figures a month while you sleep. I run my whole company on AI and I do not make money while I sleep. Most people talking about AI online are selling you magic tricks, and the ones with receipts will usually tell you it is harder than it looks.