A while ago I posted that most people still prompt AI while the top one percent run it on loops. Boris Cherny, the person who built Claude Code, put it more bluntly. He said he does not prompt Claude anymore and that his job now is to write loops. That line stuck with me because it is exactly the gap I watch every week between the teams getting real work out of an AI sales agent and the teams getting a slick demo that dies the moment it touches a live pipeline. I run Growth Cab, a GTM advisory, and I have built and killed enough of these things to tell you where the line actually sits.
in
“Most people still prompt AI. The top 1% run it on loops. A prompt asks AI to do something once. A loop is made of four things: trigger, execution, verification, and memory.”
What an AI Sales Agent Actually Is
Most of what gets sold to you as an AI sales agent is a prompt on a schedule. Someone wraps a decent prompt in a cron job, points it at a list, and calls it autonomous. It runs, it sends, and the first time reality drifts from the template it keeps going anyway, confidently wrong. A prompt asks the model to do one thing once. An agent is supposed to own an outcome over time and adjust as the world changes underneath it. Those are two different jobs, and the distance between them is the whole game. If your agent cannot notice that a claim it is about to make is false, you do not have an agent. You have a faster way to be wrong at scale.
The Four Parts of a Loop That Make an AI Sales Agent Work
When Boris talks about writing loops instead of prompts, a loop is a specific thing. It is made of four parts, and every AI sales agent that survives contact with a real quarter has all four. Trigger, execution, verification, and memory. Skip any one of them and you are back to a prompt with extra steps.
Trigger is what starts the work. A weak agent triggers on time alone, so it fires the same motion at nine in the morning whether or not anything happened. A useful one triggers on a signal. A new account lands on the target list. A prospect opens the same page twice in a week. A decision-maker changes roles on LinkedIn. The trigger is the difference between reaching out because something is true and reaching out because the clock moved.
Execution is the actual work. Pull the account, read the last ninety days of public signal, draft the message, prep the task for the rep. This is the part the demos show off, and honestly it is the part the models are already good at. Execution was never where these things broke.
Verification is how the agent decides whether the work is good enough to act on. This is the step almost everyone skips because it is unglamorous and hard. Without it, the agent has no way to catch its own mistakes, so it compounds them. Verification is the single line between an assistant that helps and an agent you can trust to move on its own.
Memory is what it carries forward. Which message got a reply. Which account said not right now and meant it. What your best-performing angle looked like last month. An agent without memory relearns nothing and repeats everything, including the approaches that already failed.
Verification Is Where Most AI Sales Agents Quietly Fail
Here is a real one. I had an agent that researched accounts and drafted outreach, and in the demo it was beautiful. In production it decided a company had raised a funding round that never happened, wrote a warm opener around that fake round, and sent it. A prospect replied to correct us. Embarrassing, and entirely avoidable. That was a verification failure, and blaming the model misses the point. The model will always be willing to fill a gap with something fluent. My job was to build the check that stops the fluent guess from leaving the building.
So now every draft the agent produces passes one rule before it can send. Every factual claim has to trace back to a source the agent actually pulled in this run. If it cannot cite the funding round, the headcount, the recent launch, it does not get to mention it. That single verification rule killed almost all of the cringe, and it cost nothing but the discipline to write it down. An AI sales agent without a verification step is not autonomous. It is unsupervised, and those are very different words.
Where the AI Sales Agent Still Breaks
I am not going to pretend this is solved. An AI sales agent still breaks on judgment. It cannot read that a deal went quiet because the champion is on leave rather than because they lost interest, so it keeps nudging at exactly the wrong moment. It breaks when the data underneath is thin, because a message is only ever as specific as what you fed it, and a sparse profile produces confident filler. And it breaks on timing that needs a human read of the room.
There is also a failure mode people underrate. The more autonomy you hand an agent, the more a single bad loop scales the damage before anyone notices. A human sending one wrong email is a bad morning. An agent sending four hundred wrong emails is a brand problem. That asymmetry is why, for anything touching a live deal, I keep a person on the verification step. The agent proposes, a human approves, and only the low-risk slice runs on its own.
How to Build Your First AI Sales Agent This Week
Do not try to automate your whole funnel. Pick one narrow loop. One trigger, one job, one verification rule. At Growth Cab the first loop that actually earned its place was simple: when a new account hits the list, research it and draft a first-touch, then check every claim against the sources before anything reaches me. For the first week run it in draft mode where it proposes and you approve every output by hand. You will find the holes fast, and each hole becomes a verification rule. Once it stops surprising you, let it act on the safe slice and keep the human on anything with revenue attached.
The honest math is that our research-and-draft loop saves me a real chunk of prep time every week, but only because a person still owns the send on anything that matters. That is the whole trick. The teams winning with an AI sales agent are not the ones who removed the human. They are the ones who moved the human from typing the email to approving the loop.
If you want the working version of this, I break down one GTM loop a day in The Revenue AI Brief, my newsletter, with the trigger and the verification rule written out. And the original thread that kicked this off, with the four-part loop in full, is on my LinkedIn. Come argue with me there if you think I am wrong about the human staying on the verification step.
