n8n vs Zapier vs Make vs Claude Code: My Honest Playbook for Picking the Right Automation Tool

Every week somebody asks me: "Should I use Zapier or n8n for this?"

Wrong question. The real one is: what does the job actually look like, and which tool is least likely to embarrass me in six months when it breaks?

I've built automation systems on all four — n8n, Zapier, Make (formerly Integromat), and Claude Code running as an agent in the cloud. Each earns its place. Each gets picked for the wrong reasons more often than I'd like. This is my actual decision framework. No hedging, no "it depends" (well, a little), and no diplomatic takes.

When I pick n8n

Client signature: they have a technical person — even one — they care about owning their stack, or the volume is high enough that Zapier's per-task pricing becomes insulting. Self-hosting matters to them, either for compliance (healthcare, finance, EU data rules) or just that "we don't want to rent our ops" energy.

Also: when the workflow has actual logic. Branching, loops, custom code, error handling. Not "when X happens do Y" — more like "when X happens, check A, B, C, branch on the result, transform the payload, hit three systems, retry the flaky one."

And when I want the thing to outlive the engagement. A $20–$40/month VPS replacing a $1,500/month Zapier bill is a conversation that closes itself.

Not for: demos, MVPs, or non-technical clients who'll be alone with it in six months. A beautifully engineered n8n instance abandoned on a server nobody logs into is worse than a Zapier setup that anyone can poke.

When I pick Zapier (honest)

When the client is non-technical and will maintain it themselves. When speed beats cost. When they need a connector nobody else built — Zapier's 7,000+ integrations is a real moat for obscure SaaS.

One-step and two-step stuff. "Slack message → Notion task." "Typeform submission → Airtable row + Slack ping." The boring plumbing that makes small-business life work.

Founders who need something live today, not next week. I'd rather ship Zapier today than promise Make on Friday.

Zapier is the McDonald's of automation. Not fancy, not what I'd cook for someone I'm trying to impress, but at 2 AM in a strange city I'm glad it exists. Don't overthink it — if the problem is small and the client is non-technical, Zapier wins on time-to-value every time.

When I pick Make

Make is the sweet spot between "Zapier's too limited or too expensive" and "n8n's overkill or nobody here can maintain it."

Make's iterators and aggregators are genuinely better than Zapier's. If the job involves reshaping data — loop through an array, transform each item, bundle results back — Make does it cleanly, Zapier makes it miserable.

Semi-technical clients who think visually do well here. Ops managers who can't code but can read a flowchart. The scenario model maps to how the work happens in their heads, not to how a developer would write the same thing.

Medium volume where Zapier pricing is ugly but you don't want the infrastructure burden. That's the lane.

The catch: debugging at scale gets painful, and the learning curve is real. You can't vibe-code your way through it.

When I pick Claude Code (or a Claude agent on cloud)

This is the one most consultants are still waving their hands about.

I pick an agent when the deterministic tool is the wrong shape for the job. The tell isn't complexity — plenty of complex things are still deterministic. The tell is edge cases.

If the flowchart needs a dozen "but what if…" branches and you still haven't covered every real case, you don't need a better workflow tool. You need judgment. That's what agents are for.

Signals I'm in agent territory:

Inputs are unstructured — emails, support tickets, PDFs, messy client messages, voice transcripts
The rules would take a week to write and still miss half the cases
A human solves this by reading context, not following steps
You tried to build it in Zapier, Make, or n8n and it keeps breaking in new ways
The work involves writing, summarizing, classifying fuzzy categories, or reasoning about context

Real example: support ticket triage. In Zapier you build 40 filters and still mis-route 20% of messages. With Claude Code and the right prompt plus tool access, you hit 95% on day one and it keeps getting better as you refine.

When the job needs a thinker, not a checklist, reach for the agent.

Where's the line

My rule: if I can explain every branch of the logic to a junior dev in fifteen minutes, it's workflow-tool work. If I can't, it's agent work.

Simpler version: structured in + structured out = workflow tool. The moment unstructured data enters — especially natural language — you're in agent territory, or at least hybrid.

I also use a cost-of-failure test. If mis-routing one in fifty items is fine, ship rules. If a single miss costs real money or reputation, the "judgment" of an agent is worth the token cost.

Most real systems I build are hybrids anyway. The agent makes the fuzzy call, n8n or Make handles the plumbing around it. That's what AI orchestration actually looks like in production — not "agent does everything," but "agent does the thinking, workflow tools do the doing." If you're designing for digital operations at scale, that split is the point.

One war story per tool

n8n — Built a lead enrichment and routing pipeline for a client quoted $1,800/month on Zapier. Moved it to self-hosted n8n for $40/month in infra. Eighteen months later it's still humming. Right call. (How the same stack-reduction move cut The Hub's software cost 60%.)

Zapier — A founder needed Slack messages tagged "VIP" creating Notion tasks before her board meeting the next morning. Built it in twenty minutes. Would n8n have been "better"? Sure. Would she still be waiting? Also sure. Right call.

Make — E-commerce client pushing 50,000 orders a month through a Frankenstein Zapier setup that was costing four figures and silently dropping orders. Rebuilt it in Make over a weekend. Ops lead could actually maintain it without calling me. Right call.

Claude Code — Built an agent to auto-draft customer support replies. V1 over-automated — it was sending replies humans should've seen first. Wrong call, or at least wrong scope. Pulled it back to "draft and flag, don't send." Now it saves the team fifteen hours a week and nobody's mad at me. Lesson: agents should escalate by default, not ship by default.

How I actually price this

Most consultants quote the tool cost and then eat the overrun when things scale. Don't do that.

Zapier and Make — The tool subscription is the client's line item, not yours. You charge for the build and ongoing support. Be honest upfront that per-task pricing climbs with growth. Don't pretend it scales linearly when it doesn't.

n8n — Setup fee, infrastructure cost passed through or managed, then a maintenance retainer. The math sells itself. "$40/month replacing $2,000/month" is a conversation I've had twenty times and it closes every time. (That's the shape of most of my fractional digital operations work.)

Claude Code and agents — This is the one nobody's pricing right yet. Token spend is variable and scales with usage. Don't price per run. Price per outcome — per ticket triaged, per lead qualified, per document processed. Token cost becomes your margin problem, not the client's surprise invoice.

And here's the frame that actually closes deals: an agent isn't replacing a $30/month SaaS tool. It's replacing a slice of a $60,000-a-year employee. Anchor against salary, not against Zapier. Token spend looks like a rounding error against headcount. It looks expensive against SaaS. Pick your anchor carefully — it is the whole conversation.

The spicy opinion: most "AI agents" are expensive regex in a blazer

Here's what I believe that most consultants won't say out loud:

Ninety percent of what's being sold as "AI agents" right now is a workflow with a language model stapled into one node.

It's Zapier with a GPT step. Make with a Claude call. n8n with an OpenAI node. That's not an agent. That's a very expensive regex wearing a blazer.

A real agent has autonomy. It decides which tool to use, in what order, and when to stop. It course-corrects when something doesn't work. It reasons about the task instead of just executing a chain someone else designed.

Most companies don't need that. Most problems are genuinely workflow problems, and a deterministic tool is the right answer. But the industry has collectively decided "agent" is the word we use now — because it sounds expensive and impressive, and because nobody wants to tell the client the truth:

You don't have an AI problem. You have a process documentation problem.

The other quiet truth: the real skill here isn't picking the tool. It's mapping the process honestly enough to know which parts are actually deterministic and which parts require judgment. Tools are commodities. Thinking isn't.

That's why I sell the thinking first and the tool second. If you only know one tool, every problem looks like a job for that tool. If you know all four — and when not to use any of them — you build digital operations that actually work.

So what do you do with this

If your team is drowning in repetitive work, losing hours to edge cases a rule engine can't handle, or spending four figures a month on automation tools that still need babysitting — that's exactly what I build for a living. Book a 15-minute call and I'll tell you honestly which tool fits, which one you're wasting money on, and what an AI-augmented version of your operation could look like. No pitch deck. No "solutions architecture" theatre. Just the diagnosis.

The short version

Zapier — fast, simple, non-technical, expensive at scale
Make — visual, mid-complexity, great for data manipulation
n8n — self-hosted, complex, high-volume, cost-efficient
Claude Code / agents — unstructured inputs, edge cases, judgment required

The real operators use all four. They just know when to pick which one — and when the answer is "none of the above, we need to fix your process first."