I got my first recommendation on Contra, from Ryan at Haven Lifestyles.
His n8n workflows were taking down the n8n Cloud instance every day. Enterprise plan, and it still fell over after a few hours because too many runs were firing in parallel. Memory limits and workflows too complex. So I refactored the workflows, moved everything to a self-hosted instance, and put monitoring on it with Slack and email alerts.
The daily collapse is gone. Error rate went from up to 12% down to 0.1%, and self-hosting now saves them over $1,500 a month. Thanks Ryan, that was a fun one to dig into.
A lot of companies are looking at automation right now, and n8n is a good place to start. For some clients it also runs AI inside the workflow, handling checks a person used to do by hand on every case.
If you have a process like that eating time in your business, message me.
RECOMMENDED
“Marcel was incredible to work with. Our n8n workflows were continuously crashing the n8n Cloud server. Marcel pinpointed the problems, proposed a solution, and fully implemented it. Tremendous ROI working with Marcel on this.”
The workflow I refuse to automate — and why that makes my AI builds better.
Problem: Every founder I meet wants to automate everything. But automating a judgment call doesn't save time — it scales bad decisions. The worst agents I've seen weren't underbuilt. They were trusted too far, too fast.
Build: Before I build, I run every workflow through three questions: Is it reversible? Does it look the same every time? What does a mistake cost? If a step fails any of them, it stays human — and the agent is built around it, with a clear handoff and an escalation rule.
Result: The automations I ship in my 5-day n8n x OpenAI sprints survive the first month, because the boundaries were decided on Day 1, not discovered in week 3. The runbook lists what the agent never touches, so the founder keeps control where it matters.
Offer: Tell me the one workflow you're unsure about automating. I'll tell you honestly whether it belongs in a sprint — and what should stay human.
Nobody seeing a 7 PM form until morning is exactly how these leads disappear. What happens when the homeowner doesn't answer the two qualifying questions? A second short follow-up the next morning is often where a quiet lead finally replies, so I'm curious whether the system sends one or hands it straight to the estimator.
Problem: Most AI demos die in week 2. The agent works in the demo, then nobody owns it, nobody knows what to do when it breaks, and the founder goes back to doing the task by hand.
Build: Every 1-week n8n x OpenAI sprint I ship ends with a 1-page runbook — what the agent does, what it never touches, and the 3 checks that take 5 minutes a week — plus a 30-minute handover. I pick the workflow by frequency x time cost x error cost, so the one we automate is the one that earns the sprint.
Result: 1 of the 3 Capafy agents I published earned its first paying subscriber in production, not in a demo. The other 2 are still running on the same runbook.
Offer: If one manual workflow runs your week, tell me which one — I will tell you whether it is a 5-day sprint or not worth automating.