From Rules to Reasoning

Most of the Zapier flows I built three years ago are still running. They pull data, send Slacks, update spreadsheets. They work. So why am I in the middle of rebuilding all of them?

The honest answer: they work, but they don't think

Old-school automation is deterministic. If X happens, do Y. That's great until the real world does something slightly off-script — a field comes in blank, a vendor changes their API response format, a new hire spells their name differently than the CRM expects. Then the flow either breaks silently or produces garbage downstream.

I spent years patching these edge cases manually. Add a filter step here. Add a formatter there. The Zap goes from 4 steps to 14 steps and nobody except me understands what it does.

What AI-native tooling changes

When I started wiring Claude into n8n flows, something shifted. Instead of "if the status field equals 'approved', route to step 3," I could write "read this document and extract the key terms, even if the format is weird." The automation became robust to variation instead of fragile to it.

That's not a small improvement. That's a fundamentally different failure mode.

The flows I'm building now can:

• Parse unstructured text and produce structured output
• Make judgment calls on ambiguous data (with human review checkpoints where it matters)
• Draft communications that sound like a person wrote them, because they did — I wrote the prompt
• Adapt to format changes without me touching the flow

What I'm actually replacing

Not everything. Some Zapier flows are still the right tool — simple, high-volume, no ambiguity, cost-sensitive. I'm not rebuilding those.

What I'm replacing are the flows that were doing fake-smart things with real-dumb tools. The ones where I'd added a bunch of conditional logic to simulate judgment. Those are the ones where Claude earns its keep.

The thing nobody talks about

The hard part isn't the AI integration. It's knowing which problems deserve AI and which ones deserve a lookup table.

I've seen people pump LLMs into places where a regex would have been faster, cheaper, and more reliable. The opposite failure is using brittle conditional logic for genuinely ambiguous tasks where a model would handle it cleanly.

Getting that judgment right is the actual skill. The tooling is just tooling.