n8n is a self-hostable workflow automation tool that connects 400+ apps with native AI nodes — and the only no-code automation platform I’d trust to run a solo founder’s daily operations in 2026. Yesterday I shipped a new affiliate to my Beehiiv list. The workflow that fired: pulled their contact info from a Tally form, enriched against Clearbit, dropped them into a Notion CRM, posted “new partner signed” to Slack, and triggered the welcome sequence in Beehiiv. Total time on my end: 3 minutes (typing the Tally response). Total automation time after that: 4 seconds.
I’ve been running n8n for 14 months. 12 active workflows currently. The 3 that broke taught me more than the 9 that worked. This article is the full list, with prompts, costs, and breakage stories.
If you’ve come from Zapier or Make and you’re wondering whether the migration is worth it, short answer: yes, if you’re past $2K/mo in MRR or running 1,000+ tasks a month. Below that, stay where you are.
Why n8n in 2026 (not Zapier, not Make)
The automation market in 2026 looks different than 2023. The shift: AI-native workflows became the default, not the exception. Zapier responded with AI features but kept the per-task pricing model that made it untenable for solo founders running real volume. Make stayed flexible but couldn’t match n8n’s open-source community velocity. n8n won the middle.
The honest math at the spend tiers I run:
| Tasks/month | Zapier | Make | n8n Cloud | n8n self-hosted |
|---|---|---|---|---|
| 1,000 | $19/mo | $9/mo | $20/mo | $5/mo |
| 5,000 | $49/mo | $16/mo | $20/mo | $5/mo |
| 25,000 | $103/mo | $29/mo | $50/mo | $5/mo |
| 100,000 | $419/mo+ | $79/mo+ | $200/mo | $10/mo (bigger VPS) |
That table is the whole argument. n8n’s pricing is execution-flat once you self-host. Zapier’s scales linearly per task. The break-even where self-hosted n8n pays for the 60 minutes of setup: roughly 2,500 executions/month, which I crossed in week 3.
Beyond price, the AI integration matters. n8n’s OpenAI, Anthropic, and LangChain nodes are first-class. You can run a Claude prompt as one node in a 20-node chain, pass structured JSON into the next node, branch on the response. Make and Zapier both treat AI as an add-on; n8n treats it as core. In 2026, that’s the difference.
The 12 workflows I actually run
I’ll go in order of value-per-month. The first three are the ones I’d ship on day one. The last three are nice-to-have. The middle six are where most of the leverage lives.
Workflow 1 — Daily revenue digest (15 min/day saved)
Every morning at 7am UTC, this fires:
- Pull Stripe payments from the last 24h
- Pull Beehiiv new subscribers from the last 24h
- Pull Plausible visits + top pages from the last 24h
- Format as a markdown digest
- Post to Slack #morning-numbers channel
11 nodes. AI usage: zero (data-only). Cost: ~$0.10/month in API calls (mostly Stripe rate limits).
I haven’t opened the Stripe dashboard for daily revenue checks in 6 months. The digest is in Slack when I wake up. The compound effect: 15 minutes a day × 365 days = 91 hours/year. That’s two work weeks.
Workflow 2 — Content factory trigger (90 min/article saved)
When a new row hits my “content briefs” Airtable, the workflow:
- Fires a webhook to a self-hosted Claude Code endpoint
- Claude Code reads the brief, generates the article draft
- Runs a quality auditor pass (separate Claude Sonnet call)
- If score ≥85, commits to GitHub
- Triggers Vercel/Cloudflare Pages preview deploy
- Posts the preview URL to Slack for my final review
22 nodes (it’s the most complex one). AI usage: 2 Claude Sonnet calls per article (~$0.40). Cost: ~$30/mo for ~75 articles.
This is the workflow that took 3 weeks to get stable. The first version generated articles in Maxime’s voice ~60% of the time. After 4 prompt iterations and a voice bible rewrite, it’s at ~85%. Still not perfect. Still saves me 90 minutes per piece.
Workflow 3 — Lead enrichment pipeline (10 min/lead saved)
When a Tally form is submitted (newsletter, lead magnet, contact):
- Pull form fields
- If email exists, call Clearbit Reveal (or fallback to Apollo)
- Enrich with company size, industry, LinkedIn URL, estimated revenue
- Drop enriched row into Notion CRM with appropriate tags
- If company size >50 or industry matches “agency”, post to Slack #high-intent channel
- Trigger appropriate Beehiiv welcome flow based on persona tag
14 nodes. AI usage: optional Claude call for lead scoring (~$0.05 per lead). Cost: $59/mo Apollo plan + $5/mo n8n hosting.
This one earned its keep in the first month. Two of the high-intent Slack pings turned into newsletter sponsorship conversations.
Workflow 4 — Cold email reply detector
Hermes (my outreach agent) sends ~50 cold emails a week. When a reply hits the Gmail inbox:
- Fetch the email content
- Pass to Claude with a classifier prompt: positive / negative / question / unsubscribe / spam
- If positive → tag in Apollo, draft response, drop in Slack for my review
- If question → similar
- If negative → tag, suppress from future sequences
- If spam → archive
9 nodes. AI usage: 1 Claude call per reply (~$0.02). Cost: pennies/month.
Save: I used to spend 2-3 hours per week triaging Hermes’ inbox. Now I review 10-15 high-signal drafts in 20 minutes.
Workflow 5 — Newsletter draft assembly
Friday at 5am UTC:
- Pull all articles published this week from the 500k.io GitHub repo
- Pull this week’s top X-trends from a custom news scraper (Workflow 8)
- Pull my Notion notes tagged “newsletter-saturday”
- Pass everything to Claude with a Beehiiv-format prompt
- Output draft to a Notion page
- Post Slack notification
13 nodes. AI usage: 1 Claude Opus call for draft (~$1). Cost: ~$4/mo.
I open the Notion page, edit for 20-30 minutes, paste into Beehiiv. From idea to sent newsletter: 45 minutes total on Friday morning.
Workflow 6 — Article internal-link suggester
When a new article hits the GitHub repo:
- Read the article body
- Pass to Claude with “find 6+ relevant existing articles from this list and suggest internal links” prompt
- Get a list of suggestions
- Open a GitHub PR with proposed link edits
- Slack me to review
8 nodes. AI usage: 1 Claude Sonnet call per article (~$0.20). Cost: ~$15/mo at 75 articles.
This was my first AI-in-the-loop workflow and the one that taught me to never auto-merge AI suggestions. The first version pushed direct commits. By article five it had hallucinated 3 broken links. Now everything routes through PR review.
Workflow 7 — Citation tracker (daily AI probe)
Daily at 11am UTC:
- Loop through 30 priority queries
- For each query, hit ChatGPT, Perplexity, Claude, and Google AIO programmatically
- Detect any mention of “500k.io” in the response
- Log to a Notion database with timestamp + query + platform + snippet
- Weekly digest fires to Slack on Mondays
18 nodes. AI usage: heavy (~120 calls/day, mostly cheap Perplexity API). Cost: ~$25/mo.
This is the workflow that makes the dashboard’s citation counter real. Without it, citation tracking is anecdotal. With it, I have weekly data. See how I track AI citations for the full methodology.
Workflow 8 — News & trends scraper
3x daily:
- Hit 18 RSS feeds (Stratechery, Latent Space, Anthropic blog, OpenAI blog, etc.)
- Filter for items published in the last 8 hours
- Pass each headline to Claude with a relevance classifier (“relevant to solopreneur + AI” yes/no)
- For relevant items, summarize in 2 sentences
- Append to a “today’s news” Notion database
11 nodes. AI usage: ~30 Claude calls/day at $0.005 each. Cost: ~$5/mo.
Powers the newsletter draft (Workflow 5) and my morning skim. Replaces ~45 minutes/day of feed-reading.
Workflow 9 — Synapse Circle daily check-in
Daily at 9am UTC:
- Fetch Synapse Skool community posts from last 24h
- Identify unanswered questions (no reply, no “answered” tag)
- Pass each to Claude with “draft a helpful, on-brand reply” prompt
- Save drafts to a Notion page for my review
- Slack me with the count
10 nodes. AI usage: variable (5-20 Claude calls/day). Cost: ~$8/mo.
I open Notion, review the drafts, edit/approve, paste into Skool. From “spend 90 minutes catching up on the community” to “spend 25 minutes.”
Workflow 10 — Subscription audit (monthly)
1st of each month at 6am UTC:
- Pull all my Stripe subscriptions (as customer) via Mercury bank transactions
- Cross-reference against a Notion “active tools” database
- Flag any tool I haven’t logged usage on in 30 days
- Generate a “review or cancel” Slack message
- Drop a kill-or-keep card in my “weekly review” Notion
8 nodes. AI usage: 1 Claude call to summarize ($0.05). Cost: pennies.
This is how I keep the stack page honest. Tools that don’t earn their $5-50/month get killed every month. Without this workflow I’d be paying for 6-8 tools I forgot about.
Workflow 11 — LinkedIn post scheduler
When I publish a new article:
- Webhook fires from the 500k.io build pipeline
- Pass article to Claude with “draft a LinkedIn post in Maxime’s voice highlighting the key insight” prompt
- Save to a Notion “social drafts” database
- Slack me to review/edit
- On approval (Notion checkbox), post via LinkedIn API at the next 9am/12pm/3pm UTC slot
12 nodes. AI usage: 1 Claude call per article ($0.15). Cost: ~$11/mo.
Save: I used to forget to share new articles on LinkedIn 60% of the time. Now it’s automated up to the approval step.
Workflow 12 — Dead-link monitor
Weekly Monday 4am UTC:
- Crawl 500k.io’s sitemap (~720 URLs as of May 2026)
- For each external link, HTTP HEAD check
- Compile broken links into a markdown report
- PR opened in GitHub with the broken-link list
- Slack me
6 nodes. AI usage: 1 Claude call to format the report ($0.02). Cost: pennies.
Boring but critical for SEO. Broken external links hurt your trust signal. This one runs in the background and I never think about it.
The 3 workflows that broke (and what I learned)
Break 1 — Auto-replying to leads
I tried a version of Workflow 4 that auto-replied to “positive” cold email responses without my approval. By day 4, the model had replied to a prospect with a confused message that didn’t match what they’d asked. The prospect ghosted. Lesson: AI drafts replies, humans send them. No exceptions on outreach.
Break 2 — Direct-merge content suggestions
The first version of Workflow 6 (internal links) auto-committed to main. On article 3, it added a link to a page that didn’t exist (hallucinated URL). Lighthouse caught the 404 a day later. Lesson: AI-generated commits always go through PR review, even when the change is “small.”
Break 3 — n8n on a $3 VPS
I tried running n8n on a $3/mo Hetzner CX11 to save the $2. It worked for 11 days. Day 12, a Claude API call timed out, n8n’s queue backed up, the VPS ran out of RAM, the process crashed, and I lost 2 days of executions before noticing. Lesson: $5/mo CX21 minimum. The $2 savings isn’t worth the risk.
The 3 workflows I’d build first
If you’re new to n8n and want maximum leverage in 30 days:
- Daily revenue digest (Workflow 1). 20 minutes to build. Saves 91 hours/year.
- Lead enrichment pipeline (Workflow 3). 90 minutes to build. Pays for itself the first month.
- Newsletter draft assembly (Workflow 5). 2 hours to build. Saves 1-2 hours per newsletter.
Build them in that order. Each one teaches you the n8n patterns you’ll need for the next. By workflow 4, you’ll be wiring nodes by reflex.
What I’d never automate (the line)
Three things I keep manual on purpose:
| Task | Why manual |
|---|---|
| Customer replies (premium tier, sponsorship inquiries) | Tone judgment + relationship signal — AI is 85% there, not 100% |
| Hiring or contractor decisions | The 15% of judgment that matters is what makes the hire work |
| Anything legal (contracts, terms updates, privacy policy) | Lawyer review non-negotiable, AI suggests, lawyer approves |
The line is whether the task has a deterministic “done” state. Pulling Stripe data is deterministic. Drafting a customer reply for a specific situation is not. Don’t automate across the line.
The honest stack cost
For all 12 workflows running on 500k.io as of May 2026:
| Cost | Monthly |
|---|---|
| Hetzner CX21 VPS (self-hosted n8n) | $5 |
| Claude API (Sonnet + Opus mix) | $40-55 |
| OpenAI API (mostly embedding for one workflow) | $3-5 |
| Apollo (Workflow 3 enrichment) | $59 |
| Total | ~$110-125/mo |
Replaces approximately 60 hours/month of manual work. At a notional $50/hr rate, that’s $3,000/mo of equivalent labor for $115/mo of automation. The math doesn’t survive on its own — I had to learn each workflow once. But the compounding kicks in by month two.
For the wider stack that pairs with n8n, see my live tool stack and the AI sales automation playbook. For the philosophy behind delegating to agents, the AI agency revenue ladder lays out where this fits in a solo business model.
FAQ
Why n8n and not Make or Zapier?
Three reasons. One, self-hosted. I run n8n on a $5/mo Hetzner VPS — Zapier's equivalent task volume would cost $69/mo minimum. Two, the workflow editor handles branching, loops, and conditional logic without forcing you into upgrade tiers. Three, the AI nodes are first-class — OpenAI, Anthropic, and the LangChain integration ship in core, not as paid add-ons. Make is a fair second place. Zapier is dead for anyone running real workflows in 2026.
Do I need to self-host or can I use the cloud version?
Cloud is fine to start. n8n Cloud is $20/mo for the Starter plan (2,500 executions). Self-hosted is free if you're willing to manage a VPS. I run self-hosted on Hetzner at $5/mo with unlimited executions. The break-even point is around 2,500 executions/mo — below that, cloud is easier. Above, self-host.
How long does each workflow take to build?
Simple workflows (3-5 nodes): 20-40 minutes. Complex with AI + branching (10-15 nodes): 1-3 hours. The first workflow always takes 2x longer than expected. By workflow five you'll be 3x faster. The investment compounds.
What's the most common reason a workflow breaks?
API changes. Notion, Airtable, OpenAI, and Anthropic all ship breaking changes 2-4 times per year. Setting up a Slack notification on workflow failure is non-optional. I learned this when a Notion update broke 4 of my workflows overnight and I didn't notice for 6 days.
Can n8n really replace a virtual assistant?
For repeatable workflows, yes. For judgment calls, no. The line is whether the task has a deterministic 'done' state. Pulling Stripe payments, formatting them, posting to Slack — that's deterministic. Replying to customer emails with the right tone for a specific situation — not yet, even with AI in the loop. Mix both: n8n handles the structure, humans handle the judgment.