27,000 AI Agents, 200 That Mattered

iFood built 27,000 AI agents in ten months. Only 200 of them truly mattered.

The goal seemed simple at the start: one agent per employee, priority on volume, errors treated as part of the process. They got there fast, with 12,000 agents active per month. One of them read twelve months of meeting notes and returned ready-to-use feedback in 20 minutes. Every AI training session began with the same rite of passage: build a snake game by voice command, no coding knowledge required.

But when someone separated the wheat from the chaff, the result was that 82% of agents saved fewer than 20 hours per month, the kind of personal assistant that helps one person. Less than 1% did the work of dozens. The company had an agent inventory and could not say which ones mattered.

The answer was to change the metric. The internal conversation shifted from "how many agents exist" to "what problem does each one actually solve," and changing that vocabulary in two months caused obvious discomfort: people who had just learned to build agents were told the effort would be measured differently.

That discomfort makes sense. Counting agents is easy and requires no judgment. It is the metric that fills board slides. Deciding which skill deserves to scale is painful, because it forces someone to admit the other 26,800 were not worth the effort they cost.

iFood's dev team came out of that shift with 16% more productivity: a team-wide gain, not a hero's stat. Metrics exist to force decisions, not to decorate presentations.

---

Sources: MIT Sloan Management Review Brasil · Mobile Time