n8n Review: AI Agents, Pricing, and Self-Hosting Tradeoffs

This n8n review examines whether its visual automation builder is a useful foundation for AI agents and recurring business workflows. The appeal is control: connect applications, define processing steps, add a model where judgment helps, and choose managed Cloud or self-hosting.

The tradeoff is ownership. Someone still needs to understand the data, permissions, errors, and costs behind the canvas. This is a documentation-based assessment using official information checked on October 15, 2026, rather than a claimed hands-on performance test.

n8n Review: Quick Verdict

n8n is worth evaluating when a technical owner wants to connect several systems and inspect what happens between them. Its visual builder, custom code, and deployment choices offer room to adapt a workflow as requirements change.

It is less compelling when nobody can maintain the integrations. A simple automation can become a small software system once it handles customer records, exceptions, and model-generated decisions. The builder makes that system visible; it does not remove the maintenance work.

Define the job before choosing the tool. If you want to delegate broad tasks to a general assistant, SD's OpenClaw review explores an adjacent approach. n8n is most interesting when you can describe the process, its inputs, and the actions it may take.

What n8n Does Well

Combine visual steps with custom logic

n8n connects application integrations and API requests in a visual workflow. You can inspect step inputs and outputs and combine ordinary processing with AI. For example, a proposed support workflow could retrieve a ticket, classify its topic, look up an account, and prepare a response for review.

The Code node supports JavaScript and Python for transformations or logic that built-in nodes do not cover. That flexibility has boundaries: current documentation says Cloud cannot import external npm modules, while self-hosted installations can enable them. Python package access also depends on the deployment and runner configuration.

Check those details early if an existing library is essential. A proof of concept that works with a few hardcoded fields may need substantial changes when real data contains missing values, inconsistent formats, or multiple records.

Add AI where the task needs it

The AI Agent node connects a chat model with tools and lets the model decide which tool to call. At least one tool must be connected. That suits tasks involving variable requests or information gathering, provided the available actions are clearly bounded.

A fixed routing rule does not automatically need an agent. If a known field determines the destination, ordinary logic can be easier to explain and test. Reserve model judgment for uncertainty that rules cannot handle adequately, then validate its result before later steps depend on it.

Keep stable record identifiers through each step. A generated customer name is a poor basis for selecting which account to update. Where possible, retrieve the record through a known identifier, validate the proposed fields, and show the exact target in the review.

n8n Pricing: What You Actually Pay For

The official pricing page lists these entry allowances on annual billing:

• Starter: €20 per month, 2,500 monthly executions, hosted by n8n.
• Pro: €50 per month, 10,000 monthly executions, hosted by n8n.
• Business: €667 per month, 40,000 monthly executions, currently self-hosted only.
• Enterprise: custom pricing, with Cloud or self-hosted deployment.

An execution is a workflow run, rather than a separate charge for every step. Treat these as listed plan prices checked on the updated date; confirm the billing term, taxes, and selected usage tier at checkout.

Budget separately for model usage, connected services, infrastructure, and maintenance. Execution allowances and Assistant credits are different units; neither should be read as unlimited model consumption.

Estimate demand from actual events. As a simple planning example, 120 incoming requests per day over a 30-day month produce 3,600 starting events before retries or additional runs. A small-looking chatbot may generate considerable activity when each conversation contains several messages.

Compare total cost per successfully completed task. A cheaper subscription can lose its advantage if corrections, outages, or oversized model calls consume the savings.

Human Approval and Workflow Reliability

n8n supports human review before selected AI Agent tools execute. The workflow pauses, shows the requested tool and parameters through a configured channel, and waits for approval or denial. This allows review at the action boundary, such as before sending a response or changing a record.

Make the review useful. Show the recipient, proposed changes, supporting context, and consequences. A person cannot meaningfully approve an action from an unexplained “continue” request. Decide who handles requests that remain unanswered and how the workflow should recover.

The broader principles in SD's AI agent security guide apply here: limit credentials to the task, treat retrieved content as untrusted input, and preserve enough evidence to investigate mistakes. An approval step complements those controls; its presence alone does not establish safety.

For failures, n8n documents error workflows that start with an Error Trigger and can notify an operator. Build a clear failure path before expanding a pilot. Also distinguish failed execution from incorrect output: an API call can succeed while updating the wrong record.

Cloud vs Self-Hosted n8n

Cloud is the simpler starting point when you want the vendor to operate the platform. Self-hosting provides infrastructure choice, including your own servers or private cloud, while giving your team responsibility for keeping the installation running.

Community Edition is free, but it does not include every collaboration or governance feature. The current comparison excludes projects, workflow and credential sharing, SSO, Git version control, and external secrets. Queue mode is included; it should not be confused with multi-main mode, which is excluded.

Those differences matter before a solo experiment becomes a shared production service. Write down who needs access, which credentials they may use, and how changes should be reviewed. Then check the edition against that operating model.

Self-hosting also does not make every data transfer local. A workflow calling an external model or application still sends the selected data to that service. Map those destinations and decide which records can enter each step.

For operations, budget for updates, backups, monitoring, and recovery drills. n8n's security documentation covers measures such as secure connections, node restrictions, credential-key rotation, and execution-data controls. Verify which controls are available in your edition and configure them deliberately.

How to Evaluate n8n Against Alternatives

Use one representative process for a short pilot. Pick something with a clear expected outcome and a person who currently understands the exceptions. A narrow ticket-triage or internal-reporting task is easier to assess than an agent expected to “run operations.”

Create test inputs for a normal case, missing data, a duplicate event, an unavailable integration, and a request that must be rejected. Record the expected behavior before running them. Check whether retries can duplicate writes and whether a model's uncertainty reaches the reviewer.

n8n's evaluation tooling supports comparing outputs across test cases, including light evaluation during development and metric-based evaluation as the dataset grows. Choose measures tied to the job: correct routing, valid fields, unsupported claims, review time, and completed-task cost.

If your shortlist includes an organization-wide agent platform, SD's Microsoft Copilot Studio review provides another workflow and governance perspective. Compare the same process in each candidate rather than deciding from a feature checklist alone.

Include recovery in the comparison. Pause a test integration, restore it, and check whether the workflow resumes without losing or repeating work. Record how much investigation the operator needs. A process that succeeds on a clean demonstration can still be expensive to support.

Finally, establish ownership. A useful pilot ends with someone responsible for credentials, integration changes, failed runs, and the test set. If those responsibilities cannot be assigned, keep the workflow as a supervised experiment.

Conclusion

This n8n review finds its strongest fit in teams that want visible, adaptable workflows and can assign a technical owner. The combination of integrations, code, AI tools, and deployment choices is useful when control matters throughout the process.

Start with a bounded task, verify the relevant edition and total costs, and measure results with realistic inputs. Expand only when the workflow handles exceptions predictably and someone can maintain it.