n8n, an AI agent or custom software: which approach fits?

Why distinguish these approaches?

A business can want to save time without needing an autonomous agent. It may need to interpret documents without rebuilding its entire system. Calling every situation “AI” makes proposals difficult to compare. Describe the task before choosing the solution.

“If the form is complete, create a record” is an explicit rule. “Understand the need expressed in this message” requires interpretation. “Apply our access rules and data model in a dedicated interface” is a product architecture requirement. A project can include all three, with different acceptance tests for each.

What does an n8n workflow do?

A workflow organises a trigger, processing steps and actions. It can receive an event, check fields, consult an application and pass on a result. The sequence becomes visible and some repetitive integration work is reduced. Rules still need design, testing and documentation.

Check connector capabilities and limitations. A connector does not guarantee access to every function or sufficient account permissions. Verify volumes, timing, connected-service limits and failure behaviour. A workflow screenshot is not evidence that incomplete inputs have been tested.

Maintenance covers credentials, versions and changes in connected applications. If a field disappears, the scenario should report it clearly. For repeated operations and data movement between tools, n8n automation is an approach worth evaluating.

What does an AI agent add?

An agent uses a model to interpret context and select steps from permitted capabilities. It may consult documentation, propose an answer or prepare an action. This helps where language varies and a simple explicit rule is insufficient.

Control begins with authorised tools. Reading and editing a record are different permissions. Drafting and sending a message are different responsibilities. Define prohibited actions, human checks and the response to missing information. An agent must be able to stop and request clarification.

Evaluate representative cases, including ambiguity and conflicting information. Assess output quality and adherence to boundaries. A business AI agent is not validated by one successful demonstration.

When is custom development justified?

Custom work becomes relevant when a specific interface, complex rules, access constraints or unusual integrations shape the product. The application controls data and journeys. It can still use workflows and models for bounded functions.

Ask what justifies custom code. A real operating constraint or missing external capability may justify it; the supplier’s preference alone does not. Separate the delivered product, reused components and separately billed services.

A custom AI integration also needs documentation, tests and recovery procedures. Code ownership does not provide independence if nobody else can install or operate it.

How do you compare the options?

Use the table to frame a discussion, not to assign tools automatically. Several implementations may solve the same operation. Compare running cost, observed quality and recovery, not only the initial quote.

What should a pilot demonstrate?

Choose a normal case, incomplete case, duplicate and connection failure. Specify the expected outcome and what must never happen. Use test data. Inspect traces that explain the steps without exposing secrets or unnecessary personal data.

Change one rule and find out who can update it. Check how to interrupt the process and continue manually. A system that works only under perfect conditions is not ready for everyday operations.

What does a hybrid solution look like?

A sensible design may combine a custom form, a workflow for data movement and a model proposing a classification. A person reviews uncertain cases, while the CRM remains the reference record. Each component has a bounded responsibility and a replacement path.

An automation review helps establish that separation. Describe your task and unacceptable failures in the project form, before deciding which tool must be used.

Technical reference checked on 20 September 2026: n8n AI documentation. The comparison is a scoping method, not a reliability or profitability guarantee.