Apollo announced something interesting this month. Three AI agents, live inside Slack, quietly running your outbound in the background. One finds leads. One builds sequences. One watches your live campaigns and tweaks them when performance dips. You don’t configure them. You don’t really manage them. You just approve what they surface, and over time, the plan is you won’t even need to do that.
The pitch, in their own words, is borrowed from smart-home tech: nobody wants to configure an IP address, so why should anyone configure a sales sequence? Set it up once, trust the system, let it get smarter without you.
It’s a genuinely good idea, and I don’t say that about most vendor announcements. My worry isn’t the automation itself. It’s what happens when you point a self-improving system at a strategy nobody’s actually agreed on.
Optimising the wrong thing, faster
Here’s the bit that gets glossed over in a lot of these launches: an agent that learns from performance can only learn what you define as good performance. Give it “more replies” as the goal, and it will cheerfully learn to write subject lines that get replies for the wrong reasons: vague enough to intrigue, generic enough to work on anyone, memorable to nobody. Reply rate goes up. Pipeline quality goes nowhere. And because the system is proud of its own improvement, nobody questions it, because the graph is going the right way.
I’ve watched this happen with human sales reps long before AI agents existed. Give a rep a reply-rate target with no brief on who you’re actually for, and they’ll hit the number by sanding off everything specific about the offer. An agent does the same thing, just faster, at scale, and without getting bored of it.
What the agent actually needs from you
The agents Apollo describes (lead discovery, sequence creation, sequence optimisation) are all downstream decisions. They answer who should we talk to and what should we say and is it working. None of them answer the question that has to come first: what do we actually believe, and who is it genuinely for?
That’s not a data problem, and it’s not something an agent can infer from your CRM history, however clean it is. It’s a strategic call somebody senior has to make and write down: the actual point of view your marketing is built on, not just a list of features and a target headcount range. Once that exists, an autonomous agent is brilliant, because it’s now optimising against something worth optimising against. Before that exists, it’s just a very fast way to find out what happens when you let a machine guess.
The trust question is the wrong one
Apollo frames the rollout carefully. Every agent needs a human approval click today, “as we earn your trust before we take the human out of the process.” I understand why they frame it that way, but I think it’s solving for the wrong risk. The danger with these systems was never that the agent will do something obviously wrong and you won’t be there to stop it. The real risk is subtler: the agent doing something confidently fine, on-brand tone, plausible personalisation, decent open rates, while quietly drifting away from what actually makes your business worth choosing. You’ll approve it, because on the surface, there’s nothing to disapprove of.
Where this leaves founders and marketing teams
If you’re weighing up tools like this, the question isn’t “can we trust the AI yet.” It’s whether you’ve got a point of view clear enough that an AI, or a junior hire for that matter, could execute it accurately without you in the room. If the honest answer is no, that’s the gap to close first. Buy the agent afterwards. It’ll be a much better investment once there’s something real for it to be consistent about.
Automation doesn’t replace strategy. It just reveals, very quickly, whether you had one.
