In short
n8n and Make are good when the process is linear, integrations are clear, the cost of error is low, and the team needs a fast launch. Custom development is needed when the workflow requires permissions, complex business logic, non-standard 1C, documents, RAG, human approval, logs, evals, and production responsibility.
The question is not which tool is “better”. The question is where automation ends and a product system begins.
If the job is “form arrives, send Telegram message, create row, draft reply”, low-code is a sensible start. If the agent reads WhatsApp, checks CRM and 1C, understands documents, shows sources, asks for approval, and logs the decision, it needs more than a chain of blocks.
When low-code fits
Low-code fits fast, bounded workflows: new lead notifications, moving data between form, spreadsheet, and CRM, drafting an email, simple request classification, daily Telegram reports, lead enrichment, reminders, status sync, and AI prototypes before custom development.
The strength is speed. You can build a working flow, show the team, test value, and avoid writing a backend too early.
Where it starts to crack
Problems start when the process stops being linear: many exceptions, user roles, source citations, personal or financial data, CRM and 1C conflicts, PDF/photo/Excel documents, RAG, detailed logs, error review UI, decision-making rather than data movement, and weekly scenario changes.
Low-code can be stretched far. At some point, though, it becomes fragile: many blocks, unclear failures, hard testing, risky changes, and blurred responsibility.
Evals are a separate boundary. You can build an AI-flow prototype in n8n or Make: collect 20-30 examples, run them manually, and see whether the agent behaves roughly as expected. But serious evals usually do not belong inside low-code. Production systems need versioned case sets, expected behavior, repeatable runs, model and prompt comparisons, regression reports, release quality gates, and error history. That is an engineering layer, not just an automation chain.
If an agent answers customers, reads documents, checks 1C, or affects CRM, evals become part of the system. Without them, the team cannot tell whether a prompt change improved quality, broke an old scenario, or is safe to release. The method is covered in why AI projects need evals.
When custom development is needed
Custom development is needed when AI becomes part of operations: WhatsApp lead qualification, internal assistants with access control, invoice and 1C checks, support answers with sources, CRM hygiene monitoring, multi-document RAG, or quality dashboards.
Here the hard parts are product concerns: roles, interface, action history, approvals, errors, monitoring, evals, knowledge updates, and security.
Hybrid approach
The best answer is often hybrid. Use low-code for prototypes, internal notifications, sync tasks, and quick connectors. Use a custom backend for critical logic, permissions, RAG, evals, logs, and interfaces.
For example, n8n receives a CRM event and sends it to a custom service. The custom service runs AI logic, checks sources, stores logs, and returns a result. Low-code then notifies the manager. That can work well if boundaries are explicit.
Decision test
Ask five questions: what happens if the automation is wrong, whether there is one source of truth, whether human approval is required, how quality will be tested, and who will maintain the workflow in six months.
If the error affects money, contracts, customers, personal data, or reputation, use a stricter system. If you cannot replay cases and see what broke, production will be stressful.
Practical path
Start with low-code when the task is simple and value needs to be tested quickly. Define the signals that trigger custom work: critical workflow, personal or financial data, 1C or documents, decision history, growing exceptions, roles, SLA, or too many workarounds.
If those signals are present on day one, build an engineering pilot and define production boundaries from the start. The AI pilot in 30 days format works well for that.
FAQ
Can an AI agent be built in n8n?
Yes, as a prototype or simple flow. For permissions, sources, documents, CRM, 1C, and handoff, add a custom layer.
Is Make worse than n8n?
No. The choice depends on team skill, integrations, self-hosting, budget, control needs, and support.
Can we start low-code and rewrite later?
Yes, if the prototype is not treated as production. Keep clear boundaries and make the future data model visible.