I tried replacing my Zapier zaps with Gemini prompts — half of them went back

The experiment: take eight of my running Zaps and see if Gemini prompts could replace them. The zaps were mundane — form responses to a sheet, email attachments to Drive, a daily digest to Slack. The theory was that a capable model with a good prompt beats a rigid automation, and for about half of them, it did.

The half that moved to Gemini: anything where the task was actually 'read this and summarize it.' My daily digest zap just forwarded titles; the Gemini version reads the articles and writes two sentences on each, which is what I wanted the whole time. Same with triaging form responses into 'urgent, this week, ignore' buckets. Judgment tasks, model wins.

The half that went back: anything on a schedule or triggered by an event. Gemini doesn't know my form exists. I had to remember to prompt it, and I forgot, twice, and missed a submission that mattered. Zapier is dumb but punctual. It ran at 3am when the lead came in, which is the entire job.

There's a hybrid that actually stuck: Zapier catches the event and calls Gemini as a step. Form comes in, Zapier fires, Gemini drafts the reply, Zapier sends it. That's most of my old zaps now, and they're smarter than they were. But the trigger, the thing that makes it run without me — that stays in Zapier.

So the verdict: Gemini replaces the brain of an automation, not the pulse. If your task needs judgment, bring the model in. If it needs to happen at a specific time without anyone remembering, keep the pipe.