Zapier and n8n both shared how they run automation internally this week. The shared rule: AI only on the step that needs judgment

Two automation companies opened up their own internal playbooks on the same day, and they agree on more than you would expect.

What happened

Zapier's COO described how the company audits AI spend. Usage data from Anthropic, Cursor, OpenAI and its own agent harness is pulled daily. Each person has a monthly allocation set above historical usage. At 75%, a Slack bot asks what they are working on. They reply in one sentence, a Zap with an AI step reads the reply, more budget is added automatically, and the manager is notified after two requests in a month. The post says AI handles only the step that takes judgment, reading a reply typed in someone's own words, and fixed rules decide what happens next.

At n8n, Romina da Costa, head of internal AI and automation, wrote that the company runs more than 12,000 internal workflows. One AI agent produced six figures of pipeline in a month, and her question was: compared to what? Her team now requires a baseline before any build. She pushes to buy proven tools over building, warns against agents for tasks like "forward this email to James" that a plain filter handles, and says every workflow needs an owner, documentation and error handling.

My take

These are large teams, but the rules shrink down well for a 10 person business:

1. Write the baseline first. How often does the task happen, how long does it take, what does good look like. A rough guess beats nothing.
2. Rules by default, AI by exception. Routing, formatting and moving data should cost zero tokens. Use a model where a human typed something messy.
3. Meter from day one. A monthly spend report and an alert before the limit catch a stuck agent loop early.
4. Give every workflow an owner and an error path. An automation nobody understands in six months is a liability.

Rule 1 is the cheapest to follow and the easiest to skip.