Tool Friday #14 — n8n: Is It Outdated Now That We Have AI?
A lot of people have been asking me lately: “is n8n outdated now that we have AI?”
The opposite. I’d argue n8n is more relevant in 2026 than it was in 2024.
I run n8n workflows that pull data, filter it, hand pieces to Claude or Gemini for analysis, and route the results to where they need to go. None of them is an AI agent doing everything end-to-end — and that’s the point.
Here’s why the question itself is the wrong question.
TL;DR: AI hallucinates. AI skims. AI breaks the rules. n8n builds repeatable flows with guardrails — same input, same output, no skipping the steps that matter. AI is the smart part. n8n is the boring part that makes the smart part actually run every day. You need both. AI alone is a chat window. n8n turns it into infrastructure.
What does n8n actually do?
n8n is a workflow automation tool that connects apps, APIs, and AI calls into deterministic pipelines triggered by a schedule, a webhook, or an event. Drop in nodes for “pull from RSS”, “send HTTP request”, “call Anthropic API”, “save to Notion”, “Slack alert if X”. Connect them with arrows. Set the trigger. The workflow runs.
The thing that matters: every node runs in a fixed order. Every step has explicit inputs, explicit outputs, explicit error handling. Nothing improvises. Nothing “decides” to skip a check because it thinks it doesn’t need it. That’s the whole product.
| Feature | What it actually does |
|---|---|
| Visual editor | Drag-and-drop nodes. Each node is one step. The flow is the diagram. |
| 400+ integrations | Native connectors to most APIs you’d plug into. HTTP Request for the rest. |
| Self-hosted or Cloud | Run it on your own machine, a $5 VPS, or n8n Cloud. Same workflow, same behavior. |
| AI nodes | Built-in Anthropic, OpenAI, Gemini, vector stores, LangChain. Or skip those and HTTP-Request the API directly. |
| Webhooks + schedules | Trigger on cron, on incoming HTTP, on event. Routine work runs itself. |
| Error workflows | When a flow fails, route to a recovery flow. Retry, log, alert, fall back. |
Why “AI replaces n8n” misses the point
This is the question I keep getting. The reasoning goes: AI agents can call tools, plan steps, write code, retry on failure — so why bother with rigid pipelines?
Because AI agents skip steps.
Hand an LLM a 12-step process and tell it “do this end-to-end”. On run 1 it does all 12. On run 50 it does 11 — quietly merging two steps because they “feel similar”. On run 200 it skips the deduplication step because the input “looked clean enough” — and now you have duplicate alerts firing into Slack at 3am.
You won’t see this happening. The agent’s reply will look fine. The output will land in the right place. The error will surface a week later, downstream, with no obvious cause.
n8n doesn’t have this failure mode. The deduplication node either runs or the workflow halts. Same input, same output, every time.
The right pattern isn’t “AI agent does everything”. It’s:
- n8n handles the deterministic parts — fetch, filter, dedupe, parse, route, save, retry
- AI gets called only on the parts that need judgment — classify this batch, summarize this article, draft this reply
Concrete example. A news pipeline pulls dozens of RSS feeds, filters by category rules, deduplicates, and lands a clean batch into Claude for analysis. The AI never sees the raw firehose. It sees the filtered handful, gets asked one specific question, returns one structured answer. That’s a one-shot AI call inside a multi-step n8n workflow.
Try the same thing with an “AI agent” running end-to-end. The agent fetches, “decides” filtering rules on the fly (different every run), forgets to deduplicate twice out of ten, and produces output that looks plausible but isn’t reproducible. You can’t audit it. You can’t trust it.
n8n + targeted AI calls > AI agent end-to-end. For anything that has to ship every day.
How I actually use it
Three workflows that show the pattern.
Competitive research. Drop a competitor URL into a folder, n8n triggers, fetches the homepage, the pricing page, the team page, scrapes LinkedIn for headcount, queries Crunchbase for funding. Every fetch is a deterministic node — same call every time, retried on failure. AI gets called twice at the end: once to synthesize a profile from the structured data, once to write a SWOT. Hours later, a Profile and a Brief appear as markdown files. No agent in charge.
News pipeline. RSS feeds get pulled on a schedule. Category filter is rule-based, not AI — keywords + source whitelisting + date freshness. Deduplication runs against a small history database. Only the cleaned batch reaches Claude, which writes a paragraph for my Inference newsletter draft. The AI step is one prompt; the rest is plumbing.
Content repurposing. A webhook listens for a published LinkedIn post. n8n fetches the text, sends it to Claude with a structured prompt, parses the response into a Twitter thread + a Reddit-shaped post, then pipes the result through a second AI call that scans for “AI tells” and invented data. If the gate fails, it retries with a stricter prompt. If it fails twice, it routes to my inbox flagged for manual review.
The pattern, every time: n8n carries the structure. AI shows up at one or two specific steps, never in charge of the whole thing.
The news that landed two days ago
Two days before this post went live, n8n shipped something worth pausing on. Their official MCP server can now build and edit workflows from a prompt. Point Claude (or ChatGPT, or Cursor) at your n8n instance, describe what you want in plain English, and the agent builds the workflow node by node, validates it, runs a test execution, and fixes itself if anything breaks.
Read that again, because this is the part that matters: the agent builds the workflow. n8n runs it.
That’s the division of labor I’ve been arguing for the whole post. Use the agent where judgment helps — designing a flow, writing the prompts, structuring the steps. Use n8n where determinism matters — actually executing those steps, every day, the same way, with retries and error handling and audit trails.
The MCP launch isn’t “AI replaces n8n”. It’s “AI builds n8n workflows for you”. The bottleneck for shipping a new automation is now “can I describe it in English”, not “do I have time to click through 30 nodes”. The deterministic substrate doesn’t change. The thing on top of it just got faster.
That’s the future of this stack. Claude builds. n8n runs.
What surprised me
The simplest pattern beats the agent loop. Modern n8n has an AI Agent node that loops with tool use. It’s powerful. The pattern that ships in most production workflows is the simplest one: HTTP Request → Anthropic API → JSON parse → next step. No agent loop, no tool use. Just “send this prompt, get this answer, route it onwards”. Boring, reliable, predictable. The agent loop is for cases where the next step genuinely depends on the AI’s reply — rarer than the demos suggest.
Self-hosting on commodity hardware is genuinely free. A Mac Mini in a corner of your office runs n8n with room to spare — see TF#13 for the setup. Marginal cost: zero. Cloud Starter would be €20/month. Over a year, the Mini pays itself back several times over.
The visual flow is documentation that doesn’t go stale. Automation I’ve written in Python lives in scripts I have to re-read to understand. n8n workflows live as diagrams. Open one cold six months later and you can follow the logic in 30 seconds.
My take
What stuck: the diagrams are documentation. The visual editor catches structural mistakes early. The HTTP Request node is the universal escape hatch — anything without a native connector, you can wire by hand in 5 minutes.
What changed in my work: I stopped writing custom Python scripts for “small” automations. The threshold used to be “if it’s more than 50 lines of Python, build it in n8n”. Now it’s “if there’s any chance I’ll want to change it later, build it in n8n”. Visual + versionable beats clever.
Real talk on the learning curve: the first five workflows take longer than the next fifty. That’s worth knowing before you start.
What are n8n’s limitations?
Learning curve is real. First time through, you’ll spend hours figuring out where to click. The visual editor is approachable; the mental model behind it (nodes, expressions, JSON pinning, error workflows) takes a week to internalize.
Workflow JSON isn’t git-diff friendly. Workflows export to JSON. Diffing two JSON exports by hand is painful. The Cloud Business plan ships native Git version control; on self-hosted you’re left with manual export-commit cycles.
Mobile editing is impossible. The web app technically loads on a phone. Building or fixing a workflow from one is a non-starter. If you need to deploy fixes from the road, you need a laptop.
Third-party node quality varies. The community node ecosystem is large; maybe 1 in 5 third-party nodes is well-maintained. Stick to nodes with active commits in the last six months.
No real-time collaboration on the same workflow. If two people edit the same flow at once, you’ll lose work. Single-player tool that happens to sync.
Self-hosting requires comfort with Docker. Not deeply, but enough. If “container” sounds foreign, start on Cloud and migrate later.
How much does n8n cost in 2026?
n8n is free if you self-host. Paid plans start at €20/month for Cloud Starter.
| Plan | Cost | What you get |
|---|---|---|
| Self-hosted (Community) | Free | Full app, unlimited workflows, your machine, your data |
| Cloud Starter | €20/mo | 1 shared project, 2.5k executions/mo, 50 AI credits |
| Cloud Pro | €50/mo | 3 shared projects, 10k executions/mo, 150 AI credits, admin roles |
| Cloud Business | €667/mo | 6 shared projects, 40k executions/mo, SSO/SAML, Git version control |
| Enterprise | Custom | Unlimited projects, external secret store, log streaming, SLA |
If you’re starting fresh and don’t have a home server, Cloud Starter at €20/mo is the right entry point. Migrating to self-hosted later is a one-time export-import.
Alternatives
| Tool | Best for | How it compares |
|---|---|---|
| Zapier | Marketers, simple triggers, fast setup | Better UX, fewer steps allowed, more expensive at scale. Limits hit fast. |
| Make (Integromat) | Visual builders, complex flows, mid-market | Closest competitor. Better visualization, cloud-only, pricing per “operation” gets weird at volume. |
| Pipedream | Developers, code-first, JS/Python in workflows | Code-native, generous free tier, audience overlap. Less visual. Closest to “n8n for engineers”. |
| Pabbly Connect | Solopreneurs, lifetime deal hunters | Lifetime pricing wins on cost; integration depth and node count are below n8n. |
| Custom code (Python + cron) | Control freaks, simple jobs, dev shops | Total control, maximum work. Worth it for 1-2 jobs. Doesn’t scale without writing your own n8n. |
If you want zero learning curve, Zapier. If you live in code, Pipedream. If you want the most powerful free option for builders running real workflows, n8n.
The verdict
Score: 8/10
n8n is more useful in 2026 than it was in 2024 — because AI workflows need orchestration more than rule-based workflows ever did. Every AI call has retries, fallbacks, structured output parsing, error routing. n8n was built for exactly that shape of work.
The 2 points off: learning curve is real, and the tooling around workflows (version control, mobile, multi-user) lags behind the runtime quality. Build with discipline; the first five workflows are an investment.
n8n isn’t outdated. It’s the layer most people are missing when their AI side projects don’t ship. AI agents end-to-end will skip a step that’s absolutely key. n8n won’t.
Discover n8n at n8n.io.
Tool Friday is a weekly series where I review one tool I actually use in my workflow. This week, the orchestration layer behind every AI workflow worth shipping.