n8n vs Custom AI Agents: A Build vs Buy Decision Matrix

Should you use n8n or build a custom AI agent?

Use n8n when the work follows a stable sequence across existing applications and the people who maintain it benefit from a visual workflow. Build a custom service when the core decision, state model or reliability contract is specific to your business. A hybrid often makes sense: n8n owns triggers and routing; a small service owns the differentiated operation.

That is an architectural recommendation, not a benchmark claiming one platform is universally cheaper. The useful comparison is the cost of a correctly completed business task, including review and recovery. Start with the broader build vs buy AI framework if the business objective is still unclear.

A decision matrix for the workflow boundary

Do not count favorable cells and declare a winner. A single non-negotiable requirement can decide the boundary. For example, a financial record that must never be written twice needs an idempotency contract regardless of which box is selected.

Worked example: inbound lead qualification

Consider a hypothetical team handling 1,000 inbound leads each month. The task is to read a submission, retrieve account context, draft a qualification and route uncertain cases to a person. The CRM remains the source of truth.

A sensible first version keeps form intake, notifications and routing visible in n8n. A separate scoring service becomes justified if qualification depends on proprietary account data, complex versioned rules or repeated evaluation of competing approaches. Neither component should write to the CRM without a stable record identifier and a recoverable execution record.

Measure the following for each candidate implementation:

1. Accepted qualifications, judged against the same sample and rubric.
2. Operator minutes spent reviewing and repairing results.
3. Duplicate writes, lost records and unresolved runs.
4. Model, hosting, platform and maintenance cost.
5. Time needed to replay a failed run safely.

Suppose candidate A costs $120 in runtime and platform charges and needs eight hours of review at an assumed $40 per hour. It completes 900 accepted tasks. Its operating cost is ($120 + $320) / 900 = $0.49 per accepted task. Candidate B costs $250 and needs three hours of review, completing 950 accepted tasks: ($250 + $120) / 950 = $0.39. These illustrative figures exclude initial development and are not n8n or model vendor prices. Add implementation cost over an explicitly chosen payback period before making an investment decision.

What self-hosting changes

n8n documents a self-hosted deployment option. That gives you a deployment choice; it also leaves an operating responsibility with your team. Inventory backups, upgrades, credentials, incident handling and access management before counting a free edition as a zero-cost system. Product editions and commercial terms should be checked against current vendor documentation during procurement.

A deployment on your infrastructure also does not establish where every connected model or application processes data. Map the outbound calls separately. A local workflow can still send a document to a remote model.

Put approval at the tool boundary

n8n provides human review for selected AI tool calls. For your workflow, decide which actions need approval and who receives the request. A useful approval view shows the proposed destination, the exact fields to change and enough source context to judge the action. Add an expiry and a recovery path in your own design so a forgotten request does not leave an invisible queue.

A draft and an external send are different operations. Keep them separate in the workflow, and test refusal, duplicate approval and retry behavior. The implementation pattern is discussed further in idempotency for agentic writes.

What to request before committing

Ask for a runnable workflow, an exported configuration, a credential inventory without secrets, a small acceptance dataset and a failure-recovery demonstration. For custom code, request the same evidence plus the source and deployment instructions. For a hybrid, require one person to own the end-to-end result rather than letting each component report its own success.

The decisive test is simple: can the future owner change one rule, run the acceptance cases and recover a deliberately failed task? If the answer is no, you have a demo rather than an owned operating system.

Sources and scope

Official documentation checked October 24, 2026. The matrix, boundary choices and example arithmetic above are the author's analysis; no vendor performance comparison is implied.

• n8n: hosting options and self-hosted editions
• n8n: human review for tool calls
• Anthropic: workflows and agents as distinct architectural patterns

For a scoped review of your own workflow, send the process, constraints and acceptance criteria.