Why I stopped using Zapier for client automations (and what I built instead)

Cross-posted from my Substack notes (einfachai Substack)

A short field-note for solo-founders and small agency owners who keep hitting Zapier /
Make pricing cliffs as their first paying client lands.

What the cliff looks like

If you crossed the 750-task-per-month threshold on a single workflow, you already
know the math: that one client now costs ~$50 / month just to move their data
around. Add a second client, and the same tooling step costs you both the tool
fee and a per-seat fee. The first $200 of recurring revenue you earn from that
client disappears into vendor rent.

For a solo consultancy serving 3-8 clients in the DACH region, this is the
moment where "low-code automation" stops being low-cost.

Three shifts that worked for me

I rebuilt the same client workflows on plain Python + a Postgres scheduler.
Yes, more code; but here is what changed:

1. Per-client cost dropped from ~$50 to < $2 / month. A Hetzner CX22 in FSN1 (€4.50/mo flat) runs all my client workflows. The first €15 I save per client is the difference between a profitable month and an unprofitable one in the first quarter.
2. Latency I can actually charge for. Native cron + a job queue gives me sub-second triggers; Zapier's "5 minute min" turned a perfectly synchronous workflow into a polling loop that visibly degraded UX.
3. One repo, real audit trail. Every workflow is a unit test. Zapier zaps were not — when a client asks "what happened on Tuesday at 14:03 CET?", the answer in vanilla infra is git log + a DB query, not a black-box re-run.

The catch is the up-front ~3 days you spend rebuilding the first 2-3 workflows.
After that the pattern is transferable; see the case-study writeup at
EinfachAI for the actual job-queue skeleton I use.

Where it breaks (be honest)

If your client needs an enterprise SSO round-trip in the trigger chain, don't
roll your own. Use the managed offering. The break-even only exists for
straight-through workflows where you own the whole data path.

If you want to see the stack details (Hetzner + n8n + a 60-line scheduler)
changelog'd weekly, I post the notes over on Substack
and on dev.to. Happy to compare notes if you are rebuilding something similar.

---

Builder practice-sharing from a solo AI consultancy in Zürich.