Verticals
BlogContactHiring
5 min read

25 People or Agents: Where AI Really Belongs in a Performance Business

One operator runs media buying with 25 people. We run the same volume with one person and agents. Same models, same market — the difference is where AI sits.

US auto insurance is one of the toughest advertising auctions in the world. By my estimate, lead generators in this niche alone hold about two million dollars a day in ad budgets. Insurers will pay well for a customer who stays for years, so hundreds of campaigns compete for the attention of every driver unhappy with a new bill.

Calma works inside this auction. We buy traffic, mostly on Meta, also on Google and other networks, and turn it into qualified leads for insurance, loans and home services in the US. Our budgets run to millions of dollars a year, and we have tens of thousands of tested creatives behind us. The economics are simple and merciless: the money for an impression leaves the account at once, and the price of a lead is known only later, after buyers have bid on it. On identical traffic a day can close up or down, depending on what the buyers bid. In March the market also went through a wave of tighter compliance, and angles that had made money for months had to come down.

In a business like this, the question of AI stopped being philosophical a long time ago. I saw that especially clearly on a recent call with one of the operators in our niche.

Two machines in one market

Competitors swapping experience is normal here: the auction is big enough, and the chaos in it is shared by everyone. So we compared structures openly.

Their machine is impressive. About 25 people in media buying, plus analysts and a technical team. Testers look for winners, senior buyers take the creative that was found and scale it aggressively. A creative team works alongside them, and there are separate tracks for different products and platforms. It is a mature, well-built organization, and I say that with respect.

Our machine solves the same problem differently. The same volume is run on our side by one person with agents. Bad ads are stopped by rules, around the clock, using criteria we developed ourselves and keep improving. Creatives are uploaded to the right campaigns automatically, and when the numbers cross a threshold, the system itself opens a task for new iterations. A person checks where the cost of an error is highest: the account, the pixel, the branch of the structure.

One against twenty-five, on the same volume. We confirmed one more thing straight away: in both companies, the number of creatives stopped meaning the number of winners long ago. Out of roughly a thousand active creatives, five of ours can carry two thirds of the budget. So we stopped growing test volume and invest in the quality of the ideas themselves, not in extra hands on buying.

The most interesting part came from them: their division of roles took shape when AI did not yet play this kind of part, and today they need fewer people for launches, tests and routine. They are not planning layoffs: their labor is inexpensive, and the structure still pays for them. A Western company does not have that option. We chose not to compete on headcount, and that is why AI became our main lever.

This comparison does not measure profit, of course. There is one judge in this game, the P&L, and we hold every decision to it. But the lever is visible to the naked eye. And it did not appear because we have a better model. The same models are available to everyone. It appeared because of where we put them.

Where the money is made inside a company

From the Adizes methodology I remember a simple division: there are green units that bring in money, and there are functions that keep them working. For us, the green units are media buying and creative production. Development, finance and analytics are critically important, and the company would not survive without them. But their value is measured by how much stronger they make the people who earn.

Read the profit end to end, from the first dollar spent to the payout for a lead, and you see where decisions change the result and where they only serve it. That map determines what each next agent will give the company.

Three levels of return

Level one: the agent saves. AI already works flawlessly here. Manual buyer reporting is gone for us, automation handles part of the checks before compliance, and routine no longer eats the hours of strong specialists. This is real money. But savings have a ceiling: they end where your costs end.

Level two: the agent strengthens the person who earns. There is almost no ceiling here. A strong buyer can keep a limited number of campaigns in their head, while their criteria, turned into rules, run across every account at once, including at three in the morning Kyiv time, when it is prime time in the US. One specialist with these levers holds a volume for which companies usually build a whole department. That is the difference between the two machines above.

Level three: the agent earns on its own. We are already testing an AI buyer on a separate track: its own decisions on launches and kills, its own P&L, a hard ceiling on losses. There is no verdict yet, and I will not promise one in advance. But this is already work inside an ad account, not a slide at a conference. The near future of our industry will be decided right here.

The condition is the same at all three levels: an agent goes only where the process is understood down to the numbers. Our kill rules appeared after thousands of decisions made by hand. Expertise first, scaling it second. Automating what you do not yet understand yourself only means making mistakes faster and on bigger sums.

An audit that won't hurt anyone

A simple exercise helped us: a short pause with no new tools, and one list. All the agents and AI services, with three columns next to them:

  • Where it sits: a unit that earns, or a supporting function.
  • Which level: saves, strengthens or earns.
  • Which P&L line it moves: and how you know.

It takes a few days, often a few hours, because the picture is usually transparent. It is just that few people stop to look at it. If the whole list lands on level one, that is already decent savings, and the next agent will probably give more next to the people who bring in the money.

And all you can lose during such a pause is the fuss. Everyone knows how to count how many people a company needs these days. I find another question more interesting: how much can each of them earn.

Ruslan Drozdov, founder of Calma. Also published on Medium and, in Ukrainian, on Speka.

More from Calma

Read articleRead article