n8n: where a workflow needs another decision layer

n8n connects applications, databases and services into a workflow of triggers, nodes and conditions. A team can use n8n Cloud or run the platform on its own infrastructure. The right choice begins with the process and the people who will operate it.

The official n8n deployment guide describes Cloud as managed and self-hosting as an option for teams that need deployment control. The pricing page is the source for current limits, plan features and prices.

Map the case before the nodes

Write one case from trigger to accepted result:

• source event and required fields;
• records read and written;
• transformations and known branches;
• retry and timeout behaviour;
• human review and failure path;
• evidence retained after completion.

n8n fits well when each branch can be described in advance. Examples include synchronising records, sending a notification after a status change, moving a document through known steps or preparing a recurring report from defined sources.

Use a process owner who can explain the exceptions. A visual workflow becomes difficult to operate when its meaning lives only in node names and personal memory.

Pick Cloud or self-hosted operation

The n8n deployment documentation lists self-hosting for teams with technical resources and Cloud for teams that prefer managed infrastructure. Both choices need access roles, credential rotation, backups, monitoring and an owner.

Count the full operating cost

n8n Cloud counts workflow executions according to its current pricing page. Self-hosting avoids a hosted subscription for the Community edition but adds server, database, updates, security and incident work. AI nodes add model and provider usage.

Use a process estimate:

monthly workflow cost = platform or server cost
                        + model and provider calls
                        + storage and monitoring
                        + owner and review time

Include retries, scheduled polling, payload storage and manual cases. A free software licence still has an operating cost.

Find the point where fixed steps stop

A workflow can branch on fields and rules that were written ahead of time. Some cases arrive with varied language, documents or missing context. The process then needs a component that can interpret the case, choose among approved actions, explain the source and send an exception to a person.

Calculate the share of cases that need judgment:

judgment share = cases requiring owner review ÷ total cases

If the share is low and the review queue is manageable, n8n may remain the complete solution. If the share is high, compare a narrow agent that prepares a decision while n8n handles triggers, queues and approved writes.

Keep decisions and execution separate

When an agent works with n8n, define the handoff explicitly. The agent can return a structured proposal with source references, confidence and a review reason. n8n can store it, notify the owner and perform only the approved operations.

Use distinct credentials for reading, drafting and writing. Keep outbound messages behind a review step when the process carries commercial or personal data. Record the workflow version, agent version, source records and owner decision.

Use the current documentation as a control

n8n features and plan limits change. Before purchase, check the pricing page, the deployment choices and the self-hosting guide. Read the license if your company plans to provide n8n-based services to others.

Decide whether n8n is enough

n8n is a strong fit when the process is stable, the rules are explicit, the team can operate the chosen deployment and exceptions have a clear owner. Add an agent or another application when the case requires interpretation, varied evidence, approved action selection and an inspectable decision record.

Syntalith lists process automations from €3,500 net and apps or agents from €6,000 net. A free process scan can map one workflow, calculate its exception share and identify the smallest sensible implementation.

n8n readiness list

1. What starts one case, and what counts as complete?
2. Which branches are explicit and which need judgment?
3. Who maintains the deployment and credentials?
4. Where are source data, logs and backups stored?
5. How are retries and duplicate actions controlled?
6. Which actions require approval?
7. What does a complete case cost?
8. What evidence would justify adding an agent?