n8n in 2026: A Review After Three Years of Production

Updated May 2026. Rewritten as a practitioner review based on three years of production n8n deployments, with current pricing, AI Agent node observations, and the comparisons people actually search for.

Most n8n reviews online are written by people who installed it last Tuesday. This one is written after three years of running n8n in production for clients across healthcare, recruitment, finance, e-commerce, and professional services. We have rolled it back exactly twice in that time, both for reasons that were not the platform’s fault. That is the framing.

We are Osher Digital, a Brisbane-based automation consultancy. n8n is the workhorse for most of the workflow automations we ship. We have also tried Make, Zapier, Power Automate, Activepieces, and a half-dozen smaller tools. n8n keeps winning, but not for the reasons the marketing copy emphasises.

This review covers what holds up, what bites, the AI Agent node (which is genuinely the biggest change of the year), and the boring operational details that decide whether n8n is right for your team. If you are looking at adopting n8n or comparing it against alternatives, this is the version we wish we had read in 2023.

The Thirty-Second Verdict

n8n is the right choice for organisations that want serious workflow automation, self-hosting as an option, low per-execution cost at scale, and the ability to drop into JavaScript or Python when the visual nodes run out. It is the wrong choice for organisations that need 1,500 pre-built integrations or zero technical capacity on the team.

If you have one engineer or one technical operations person, n8n will outperform Zapier and Make on every dimension that matters at scale. If you have no one technical at all, Zapier is still the better starting point.

What n8n Does Brilliantly

Self-hosting that actually works

Most “self-hostable” SaaS tools are theoretically self-hostable. n8n is operationally self-hostable. A Docker compose file with n8n, Postgres, and a Caddy reverse proxy is the minimum viable deployment, runs on a $15-$30 AUD per month Vultr or Hetzner instance, and supports tens of thousands of monthly executions before you need to think about queue mode.

For Australian clients with data residency concerns (healthcare, financial services, professional services handling client records), this matters a lot. Workflow data stays on a Sydney or Melbourne VPS. We wrote up the production deployment patterns in our Docker hosting guide.

The Code node is the killer feature

Zapier has Code by Zapier. Make has Custom JS functions. Neither is comparable to dropping into a full Node.js or Python execution environment with npm or pip available. When the integration you need does not have a pre-built node, the Code node solves it in 15 minutes with whatever library you reach for. Most workflows we ship include two or three Code nodes for the parts that nothing else handles cleanly.

The AI Agent node is the biggest change of 2026

n8n’s AI Agent node, which started shipping seriously in late 2024 and matured through 2025, has changed what we put into n8n. Workflows that used to be six branching nodes are now one AI Agent node with three tools and a system prompt. The agent decides which tool to call, parses the response, and continues the workflow.

It works with OpenAI, Anthropic, Google models, and any provider exposing an OpenAI-compatible API including self-hosted Ollama, vLLM, and LM Studio. Tool calling is solid. The Window Buffer Memory and Vector Store memory nodes handle conversation state and RAG respectively. We covered the patterns in detail in our AI workflow guide.

The honest take: this is the feature that has kept us from migrating clients onto pure-Python agent stacks. n8n’s combination of workflow orchestration and AI Agent capability is unique in its tier. It is what makes “build me an AI customer support workflow” a 4-day job instead of a 4-week one.

Cost at scale

Self-hosted n8n is functionally unlimited. The cost is your VPS bill. We have run client workflows over 200,000 executions per month on a $40 AUD instance. The same workload on Zapier would have run into thousands of AUD per month; on Make, hundreds.

n8n Cloud’s pricing tiers (Starter at $20 USD/month, Pro at $50 USD/month, Enterprise quote-only) make sense for organisations that do not want to manage hosting. The Pro tier handles around 10,000 active executions per month before you should move to self-hosted.

Where It Bites

The honest list. Things we have hit and that you should know about before you build anything serious.

Version control is still painful

n8n’s workflow JSON exports are stable enough to commit to git. Diffing them is unpleasant. Reviewing changes in a pull request requires either rendering the JSON visually or using the diff feature in n8n’s UI. Three years in, this is still the operational pain point that bites our team most often.

The 2025 environments feature improved the dev-staging-production story significantly, but the lack of a clean “promote this change from staging to production” command remains a missed opportunity. We work around it with our own deployment script.

Encryption key loss is catastrophic

The single biggest debugging session we have lost time on was an encryption-key migration that went wrong, taking out every stored credential across 40 workflows. Restored from backups, but the recovery took 6 hours. The encryption key lives in 1Password now, with a second copy in a sealed envelope in the office safe. Do not skip this.

Execution data growth is unforgiving

Default settings store every execution result indefinitely. A workflow running every 5 minutes will fill the database faster than you expect. Postgres bloats, disk fills, n8n starts crashing. The fix is the environment variables EXECUTIONS_DATA_PRUNE=true, EXECUTIONS_DATA_MAX_AGE=336 (hours), and EXECUTIONS_DATA_PRUNE_TIMEOUT=3600. Set them from day one. We covered the full config in our n8n configuration guide.

The UI still confuses non-technical users

For all the talk of low-code, n8n’s UI assumes a certain comfort with JSON, expressions, and data structure thinking. Junior procurement officers and marketing coordinators do not pick it up in a one-hour training. Compare to Zapier, where they can.

Our practical handover model is one technical owner per organisation who maintains workflows, with business users requesting changes through tickets. That works. The “anyone in the team can build automations” pitch does not match our experience.

Some integration nodes are under-maintained

The official Salesforce, HubSpot, Slack, and Google Workspace nodes are excellent. A handful of community nodes for less-common services are spotty. We have rewritten three nodes ourselves in the past year (one for an Australian e-commerce platform, one for a payroll API, one for an old CRM) because the available options either had bugs or had not been updated for the latest API version. Budget for this if your stack includes anything off the well-trodden path.

Webhook debugging can be frustrating

The test webhook URL behaves differently from the production one. Test mode catches the first payload then disconnects. Production webhooks need the workflow active. This is documented and still trips people up weekly. The fix is patience and a habit of using ngrok or n8n’s test-mode-with-listen for the iteration loop.

n8n vs the Competition

vs Zapier

Zapier has more pre-built integrations and a much smoother experience for non-technical users. n8n has self-hosting, lower per-execution cost at scale, more flexible logic, and far better handling of complex branching workflows. The 2026 verdict: Zapier for small teams without a technical owner; n8n once you cross a few thousand executions per month or need self-hosting.

vs Make (Integromat)

Make is the closest direct competitor. Similar visual workflow approach, similar pricing tier, similar integration count. Make handles iterators and array manipulation slightly more elegantly in the UI. n8n wins on self-hosting, Code node power, and the AI Agent node. We have migrated clients off Make onto n8n more often than the other way around, usually because they wanted self-hosting or hit Make’s operation-count limits.

vs Microsoft Power Automate

Power Automate wins inside a Microsoft 365 ecosystem (SharePoint, Outlook, Teams, Dynamics) by a wide margin. It loses outside that ecosystem. The non-Microsoft connectors are inconsistent, the licensing is opaque (per-user, per-flow, premium connectors charged separately), and the UI is dated. For a Microsoft-centric organisation, Power Automate is a no-brainer; for everyone else, n8n is more flexible at a lower total cost.

vs Activepieces

Activepieces is the open-source up-and-comer. We have piloted it twice. The UI is cleaner than n8n’s in some places, the AI agent story is still catching up, and the integration library is smaller. For organisations specifically wanting an MIT-licensed alternative to n8n’s fair-code license, Activepieces is worth a serious look. For most cases, n8n’s larger community and integration count still wins.

vs pure Python with Temporal or Prefect

For engineering teams building workflows as code, Temporal and Prefect are technically superior. Better testing, better version control, real type safety. We use them for client workloads where workflow logic is mission-critical and the team is engineering-heavy. n8n wins when the workflow benefits from a visual representation, when business users need to read (not write) the logic, or when the iteration speed of a visual editor outweighs the rigour of code.

Pricing: The Honest Breakdown

Three deployment options, three cost profiles.

Self-hosted, single instance. $15-$40 AUD per month for the VPS (Vultr, DigitalOcean, Hetzner; Sydney regions available). Unlimited executions, unlimited workflows, unlimited users. Realistic ceiling around 50,000-100,000 executions per month before you need queue mode and a second instance.

n8n Cloud, Pro tier. $50 USD per month (around $75 AUD). 10,000 active executions per month, 20 active workflows, 5 users. Good for small-to-mid teams without an ops person. We move clients off this tier when they cross 8,000 monthly executions; the self-hosted economics dominate.

Enterprise (Cloud or self-hosted Enterprise license). Quote-only. Adds SSO/SAML, advanced RBAC, audit logs, external secret stores, dedicated support. Typical entry point for organisations with compliance, security, or audit requirements. Australian regulated industries (healthcare, financial services) usually land here.

One trap to avoid on n8n Cloud: the “active workflow” count includes workflows that are switched on but rarely trigger. If you have 50 webhook-triggered workflows that fire occasionally, you are paying the Pro tier limit without using 10,000 executions. Move to self-hosted.

When n8n Is Not the Right Tool

The honest “do not use n8n for this” list.

Sub-second latency requirements. n8n has overhead per node. A workflow with 12 nodes and external API calls runs in seconds, not milliseconds. If you are building real-time inference for a customer-facing UI, build it as a Python service and use n8n for the upstream and downstream pipeline.

Streaming data pipelines. n8n is a workflow engine, not a stream processor. For Kafka topics with millions of messages per hour, use Flink, Beam, or a purpose-built stream processor. n8n handles the orchestration of side effects and the cold-path enrichment.

Highly regulated workflows where audit trails must include full execution history at the field level. n8n’s audit logging is good but not finance-industry-good. For a bank running automated trade reconciliation, the audit trail needs to be primary, not a side benefit. Use a purpose-built workflow engine.

Pure ETL with structured tabular data. Use Airbyte, Fivetran, or dbt. n8n handles ETL competently and is not the right tool when the entire job is moving 50 tables from Postgres to Snowflake nightly.

Outside those cases, n8n usually fits. For most business automation, AI agent workflows, document processing, CRM glue, and notification routing, it is the platform we reach for first.

Getting Started Realistically

If we were starting from scratch in 2026, the order of operations:

Spin up n8n Cloud Starter for $20 USD/month. Build one workflow end-to-end — pick something that connects two systems you already use and saves a few hours a week. Learn the data model, the expression syntax, and the error handling. Three or four weeks of this.

Migrate to self-hosted Docker on a $20 AUD VPS in your nearest data region (ap-southeast-2 for Australian businesses). Move the workflow. Set the data pruning environment variables. Back up the encryption key. Set up Caddy for HTTPS.

Add the second workflow. Use the AI Agent node if your task suits it. Build the next three from real business problems, not from “what could we automate” exercises.

After about six months, do an honest review. Are workflows mostly stable? Are exceptions being investigated? Is one person able to maintain everything in a manageable amount of time? If yes, expand. If no, fix the operational discipline before adding more.

Frequently Asked Questions

What is n8n in plain English?

n8n is a workflow automation platform that connects apps and services through a visual node-based editor. You build flows by dragging nodes onto a canvas, configuring each step, and connecting them in sequence. Triggers fire workflows in response to events (webhooks, schedules, new emails, new database rows), and the workflow does the work — calling APIs, transforming data, writing to databases, sending notifications. It is open-source-ish (fair-code licensed), self-hostable, and competes with Zapier and Make.

Is n8n free?

Self-hosted n8n is free to download, run, and use for any purpose under the Sustainable Use License. You pay only for the infrastructure (a VPS, a Postgres database, a domain name for HTTPS). The total cost runs $15-$40 AUD per month for most small-to-mid deployments. n8n Cloud, the managed hosted version, starts at $20 USD per month for the Starter tier.

n8n vs Zapier — which is better?

Zapier is better if you have no technical team and want maximum pre-built integrations with minimum setup. n8n is better if you have at least one technical person, value self-hosting or data control, run anything over a few thousand executions per month, or need flexible logic and code. For technical teams, n8n wins on every dimension that matters at scale. For non-technical teams, Zapier is more forgiving.

Can I build AI agents in n8n?

Yes, and this is the biggest change of 2026. The AI Agent node supports OpenAI, Anthropic, Google, and any OpenAI-compatible API including self-hosted models. You define tools as workflow nodes, the agent decides which to call, and the orchestration around it is standard n8n. We have shipped several customer support, document classification, and research-assistant agents this way. Production-ready with appropriate guardrails.

Is n8n RPA?

Not in the classical sense. n8n is API-based workflow automation, not screen-scraping or UI-based bot automation. If you need to click through a legacy desktop application that has no API, that is classical RPA territory (UiPath, Automation Anywhere) and n8n is not the right tool. For everything that can talk over an API, n8n is what classical RPA pretended to be: fast, cheap, and reliable. We covered the comparison in our n8n for RPA guide.

How much does self-hosted n8n actually cost to run?

A small Sydney-region VPS at $15-$20 AUD per month runs n8n with Postgres for the average small-to-mid client workload. Add a domain and SSL (free with Let’s Encrypt), and the total cost of ownership for a single-instance deployment is well under $300 AUD per year. Add backups, monitoring, and a staging instance and you are still under $1,200 AUD per year. We have run client workloads with hundreds of thousands of monthly executions on infrastructure at this scale.

Should I self-host or use n8n Cloud?

Use n8n Cloud if you do not have technical capacity to maintain a server, if you process fewer than 5,000 executions per month, or if your workflows handle non-sensitive data and the simplicity is worth the price. Self-host if any of those conditions reverse: technical capacity available, higher volumes, sensitive data, or specific compliance requirements. The crossover point for most Australian SMBs sits around 5,000-10,000 monthly executions.

What is the biggest mistake you see teams make with n8n?

Building 50 workflows in three months with no operational discipline. Workflows that nobody owns, no error notifications, no execution monitoring, no documentation, no version control. By month four the team cannot tell which workflows are still running, which are critical, and which would just fail silently. Set the operational practices before scaling: one owner per workflow, error notifications via Slack, weekly review of failed executions, a documented purpose for each workflow. The teams that do this scale to hundreds of workflows comfortably; the teams that do not stall at fifty.

n8n in 2026 is the most capable workflow automation platform we use for client work, and it has earned that position through three years of production-grade reliability and the genuine leap of the AI Agent node. It is not perfect; the things that bite are listed above. If you want a second opinion on whether n8n is the right fit for your business, get in touch. We are based in Brisbane and have shipped n8n automation for businesses across Australia.