The guide · playbook
The AI automation agency playbook.
The automation-agency model that blew up in 2023–2024 was built on one idea: connect apps with n8n, Make or Zapier and charge $2–5k a month. That window is closing fast — clients watch the same tutorials, and templates make 80% of workflows copy-paste. The agencies that survive won't be workflow builders. They'll be AI infrastructure partners. Here's the exact stack, the cost math, and the secret sauce.
The full playbook
🔒 Drop your email to reveal the full playbook
Get the full playbook, free.
Drop your email and the playbook unlocks instantly, plus every other guide and resource in the library. One email a week, unsubscribe anytime.
No fluff, no pitch. This is exactly how a modern agency runs right now, and why the old model is dying.
The problem nobody wants to admit
The model that blew up in 2023–2024 was one idea: "I'll connect apps together for clients using n8n / Make / Zapier and charge $2–5k/month." That worked when clients didn't understand automation. That window is closing fast:
- Clients are watching the same YouTube tutorials you learned from.
- ChatGPT and Claude can explain any n8n workflow in 30 seconds.
- Templates and community nodes make 80% of workflows copy-paste.
- The perceived value of "I connected your CRM to your email tool" is collapsing.
The agencies that survive will not be workflow builders. They'll be AI infrastructure partners.
The old stack vs the new stack
What 90% of agencies are still selling: Client request → n8n / Make / Zapier → trigger → node → node → node → output → pray nothing breaks → client pays $3k/mo for duct tape.
The reality of that model:
| Problem | What actually happens |
|---|---|
| Fragile by design | One API update breaks the whole chain. |
| Maintenance trap | 30–40% of your time is fixing existing workflows, not building new ones. |
| No intelligence | Workflows don't understand context. They just follow instructions blindly. |
| Race to the bottom | Every new agency owner undercuts you with the same templates. |
| Client churn | Clients eventually realise they can build the same thing themselves. |
What the top 10% are moving to: Client request → Claude Code + MCP servers → the agent understands the goal, figures out the steps, and adapts when things change → a custom dashboard for the client to see everything → client pays $5–15k/mo for an intelligent system they can't replicate.
The new stack explained
1. Claude Code (the brain)
Claude Code is not a chatbot. It's an autonomous coding agent that runs in your terminal. It can read entire codebases and project contexts, plan and execute multi-step tasks, write, test and debug code, connect to any external tool through MCP, and self-correct when something fails.
An n8n workflow is a static set of instructions. Claude Code is a thinking operator. When an API changes, an n8n workflow breaks at 3am. Claude Code reads the error, understands what changed, and adapts.
2. MCP servers (the connections)
Model Context Protocol is an open standard that lets Claude Code talk directly to external tools. No middleware. No webhook chains. You build or install an MCP server for each tool:
| Tool | What the MCP server does |
|---|---|
| Slack | Read/send messages, manage channels. |
| Notion | Create/update pages, query databases. |
| Google Sheets | Read/write/format spreadsheets. |
| CRM (HubSpot, etc.) | Create deals, update contacts, pull reports. |
| Stripe | Check payments, manage subscriptions. |
| GitHub | Create PRs, review code, manage issues. |
| Custom APIs | Anything with a REST or GraphQL endpoint. |
Claude Code calls these directly. No nodes. No visual spaghetti. Just tool use.
3. Custom dashboards (the client layer)
This is the moat most agencies are missing. n8n gives your client nothing to look at. Maybe you share a workflow screenshot. Maybe you send a Loom every month. That's not a product — that's freelancing.
Instead, build a lightweight custom dashboard for each client where they can see every automation that ran and what it did, view results, reports and outputs, trigger specific tasks manually, and monitor costs and usage.
Tech: Next.js or a simple React app. Claude Code builds 90% of it for you. Deploy on Vercel. Takes 1–2 days to set up per client. The client sees a real product, switching cost goes through the roof, you can charge 3–5x more, and you own the interface between the AI and the client.
The cost comparison
Old model (n8n agency):
| Item | Monthly cost |
|---|---|
| n8n cloud or self-hosted VPS | $100–300 |
| Your time maintaining workflows (30–40% of hours) | $3,000–6,000 |
| Debugging and emergency fixes | $500–1,500 |
| Total cost to you | $3,600–7,800 |
| What you charge the client | $2,000–5,000 |
| Actual margin | Thin or negative |
New model (Claude Code + dashboards):
| Item | Monthly cost |
|---|---|
| Claude API — Sonnet 4.6 for 95% of tasks | $200–600 |
| Claude API — Opus 4.5 for complex reasoning | $50–150 |
| Dashboard hosting (Vercel) | $0–20 |
| Your time building, not maintaining | $1,000–2,000 |
| Total cost to you | $1,250–2,770 |
| What you charge the client | $5,000–15,000 |
| Actual margin | 70–80% |
The credit-optimisation trick nobody talks about (secret sauce)
Most people run everything on Opus 4.5 ($15/$75 per million tokens) because they think bigger model = better results. Wrong. 95% of tasks run perfectly on Sonnet 4.6 ($3/$15 per million tokens) — a 5x price difference. File reads, API calls, status checks, basic code changes, data formatting — all Sonnet tasks. Opus is only for complex architecture decisions, multi-file refactors, and deep reasoning chains.
Add this one line to your CLAUDE.md file:
Set it once. Every session Claude will coach you on cost automatically. This alone saves hundreds per month per client.
If I was starting from zero in 2026
Week 1–2: Foundation
Install Claude Code and learn the basics. Set upCLAUDE.md with your working style and the cost-optimisation line above. Build or install MCP servers for the 5 most common tools your clients use. Practise rebuilding simple automations you'd normally build in n8n.Week 3–4: First client
Find one client who currently pays for n8n/Make/Zapier automations. Offer to rebuild their top 3 workflows as a pilot. Build a simple dashboard showing them the results. Price it at 2–3x what they currently pay, but sell the intelligence and reliability.Month 2–3: Productise
Create a repeatable template for the dashboard. Document your MCP server library. Build case studies from your first client. Start positioning as an AI infrastructure partner, not an automation agency.Month 3–6: Scale
Each new client takes less time because your MCP servers and dashboard templates are reusable. Claude Code builds most of the custom work for you. Your job becomes architecture and client strategy, not node wiring. Hire one technical person to manage deployments.What we still use n8n for (being honest)
We're not saying burn it all down. Three cases where n8n still earns its place:
- Dead-simple stable webhooks — "Stripe payment → Google Sheet row." If it hasn't changed in 6 months, don't touch it.
- Client-facing visibility — some clients want to see a visual workflow. n8n's editor works for demos.
- Legacy integrations — some older APIs have solid n8n community nodes that aren't worth rebuilding.
Everything else has moved to Claude Code.
Why 90% will fail
They'll fail because they'll keep selling workflows that clients can build themselves, competing on price with every new agency owner who watched a YouTube course, spending half their time maintaining brittle integrations, and offering no defensible product layer between them and the client.
The 10% will win because they'll sell intelligent systems that adapt and self-correct, build custom interfaces that create real switching costs, operate at 70–80% margins instead of razor thin, and position as AI infrastructure partners, not Zapier consultants. The shift isn't coming. It already happened — most people just haven't noticed yet.
Resources
- Claude Code → anthropic.com
- MCP Protocol → open standard by Anthropic for tool-use integrations
- Secure agent architecture example → productcompass.pm
- Multi-agent setups with Telegram → dan-malone.com
- Claude Code + n8n integration patterns → amplifyaiworkshop.blog
Original guide by Artem Novitckii. More from him on Instagram, in the Skool community, and across his resource library.