StackEndBox

Automation buying decisions

n8n, Zapier, Make or custom code: choose the automation your team can run

By StackEndBox. Published .

Choose the least complex implementation that can complete your business workflow, expose its exceptions and leave a named person able to operate it. Start with the required actions, permissions and failure behaviour. A platform's connector catalogue or a successful demonstration cannot answer those questions alone.

This guide is for an operations, support or product team buying implementation. It compares a managed platform, self-hosted workflow software and custom engineering without declaring one universal winner. Vendor facts were checked against the linked official pages on 27 October 2026. Features, licence terms and plan limits can change, so verify the actual contract before committing.

Which option should a business shortlist?

Shortlist by the requirement, then prove it on one workflow
OptionWorth evaluating whenQuestion to settle before purchase
ZapierYour team wants a managed workflow with supported actions for its existing applications.Do the actual connector actions, permissions, billing rules and recovery path cover the whole task?
MakeA scenario-based platform fits the required transformations, branches and application connections.How do representative inputs, iteration, exceptional runs and AI steps consume credits on the proposed plan?
n8n CloudYou want n8n's workflow model with vendor-managed hosting.Do the selected plan's execution, concurrency, governance and support boundaries fit your operation?
Self-hosted n8nYou have a justified infrastructure requirement and an accountable team to run the service.Who owns updates, credentials, backups, monitoring, restoration and licence review after handover?
Custom code or a hybridThe required data contract, approval, application state or recovery behaviour cannot be met economically with the shortlisted platform.Which precise constraint needs code, and who will test, deploy and maintain that part?

These are evaluation paths, not claims that a named product cannot handle complex work. Test the requirement against its current capabilities. A hybrid can keep an ordinary trigger and supported connector on a platform while a small service enforces a specialised rule. Avoid replacing a working platform merely to move the diagram into code.

Compare the same business outcome in each proposal

Consider a request arriving in a shared business inbox. The intended outcome is a saved request assigned to the right team, with uncertain cases held for review and an operator able to see notification failures. This is a planning example, not a reported client result.

  1. Define the trigger and required fields. Decide how forwarded, repeated or incomplete requests are recognised.
  2. Use ordinary rules for known categories. Evaluate a model only if interpreting free text is necessary.
  3. Save the accepted request and its operation reference before treating a notification as proof of completion.
  4. Route uncertain classification to a person. Keep consequential customer messages behind the agreed approval.
  5. Show the operator the saved state after a failed dependency or lost response. Establish what may safely be retried.

Ask each provider to implement this same boundary in the proposed review environment. A platform demo that sends one message is a smaller deliverable than the complete intake, review and recovery journey. Compare the source system's access and event behaviour as well as the workflow editor.

Tasks, credits and executions are different cost units

Zapier's official task accounting distinguishes successful actions, replayed steps, exempt steps and products with different usage rules. Make's pricing explanation uses credits, with consumption depending on the actions and features used. n8n's pricing page describes monthly workflow executions. These units are not interchangeable business outcomes.

Write down expected business cases per month, the ordinary path, exception share, branches, repeated actions and any separate model usage. Run permitted examples and inspect the platform's actual usage records. Check plan features, limits, overages and what stops when a quota is reached. Do not infer the cheapest implementation from a headline monthly subscription.

A useful comparison separates setup, subscription or hosting, third-party usage, human review, maintenance and incident handling. Calculate cost per accepted useful case using a consistent definition. A process can consume fewer platform units while sending more cases to a person. Our worked automation economics example explains why review effort belongs in that calculation.

Self-hosting changes who operates the service

n8n supports deployment on your own infrastructure, as described in its hosting documentation. Choosing that route creates operating responsibilities. Agree who patches the service, manages its database and secrets, monitors capacity, tests backups and restores a failed environment. Distinguish those duties from editing an individual workflow.

A server in a chosen region does not by itself keep every input there. Review connector destinations, model providers, execution history, logs and support access. Confirm which governance features the selected edition includes. Internal automation, client-managed implementations and embedding a platform in a product can have different licence and contractual requirements, so review the intended use before promising resale or unrestricted portability.

Five acceptance checks before scaling

  1. Normal result: the operator finds the expected saved record after refresh, with the correct owner and status.
  2. Uncertain input: an incomplete or ambiguous case reaches the agreed review path instead of silently receiving a guessed value.
  3. Repeated event: replay a labelled test input and inspect both workflow and destination state. Do not assume a connector retry setting prevents duplicates everywhere.
  4. Unavailable dependency: fail a notification or downstream action in a safe review environment. The operator must distinguish saved work from incomplete delivery.
  5. Handover: the nominated operator uses the documented controls, sees relevant errors and follows the recovery procedure without the original developer doing it for them.

These are checks to commission, not tests we claim to have run on your accounts or every platform. Record actual failures and unsupported requirements before expanding. If the problem is a clean connection between systems, use the integration scoping checklist. If a model genuinely must choose its next tool, review the workflow-versus-agent decision separately.

What to send an implementation partner

Send one non-confidential workflow: its trigger, current applications, required saved outcome, approximate volume, exceptions, approval owner and nominated operator. Include the platform you already use and the specific limitation you have observed. Agree source access and representative samples later through a secure process.

Ask for a platform-fit recommendation, included deliverables, exclusions, an acceptance plan and separate running costs. Retain a working implementation when it meets the requirement. StackEndBox can scope workflow discovery, platform implementation or custom automation engineering. The proposal establishes stack fit, assigned people, fees and support. Our handover comparison sets out the operating questions.

Discuss your automation implementation. Describe the repeated task and the result your team needs. No credentials or private customer records are needed for the initial enquiry.

Buyer questions

Should we use n8n, Zapier or Make before commissioning custom code?

Evaluate the existing platform and supported actions against the actual workflow first. Use custom engineering for a demonstrated requirement or a justified total operating-cost advantage. A hybrid may need less new code than rebuilding every step.

Is self-hosted automation free to operate?

No. Separate any software licence entitlement from hosting, administration, backups, security maintenance, monitoring and support. Confirm the licence and edition for your intended use.

Does an AI workflow need an agent?

No. Prescribed steps can use ordinary rules and a bounded model step. An agent is appropriate only when choosing the next step or tool is part of the task and its permissions and evaluation can be defined.