Openclaw — Agentic AI

Blog · Workflow platforms

n8n vs Make.com vs Custom Code: A 2026 Decision Tree for Building AI Workflows

By Published on May 6, 2026 11 min read

After 47 production AI workflow deployments at Openclaw clients (2024–2026), the n8n vs Make.com vs Custom decision is no longer a religious debate — it’s a 4-question decision tree with predictable answers. n8n wins when you have <50 daily executions and need a visual builder. Make.com wins when you have a non-technical operator owning the workflow. Custom code wins everywhere else — and that “everywhere else” is 70% of real client cases. Here’s the framework, the hidden costs of each, and 2 anonymized case studies showing why.

Why is the n8n vs Make.com vs Custom debate even a thing in 2026?

The debate exists because the marketing of low-code platforms systematically hides the operational tax — server overhead, execution caps, debugging time, and the 6–12 month “rebuild it in code anyway” inflection point. The default answer most agencies give (“use Zapier / use Make / use n8n”) is correct ~30% of the time and catastrophically wrong the other 70%. The right answer depends on 4 variables that nobody bothers to ask the client.

We’ve shipped workflows on all three (n8n self-hosted, n8n cloud, Make.com Pro tier, Make Enterprise, custom Python/TypeScript on Railway). We’ve also rebuilt 6 client workflows from n8n/Make to custom because the original choice cracked under scale. Below is the framework we now use to avoid the rebuild.

What’s the 4-question decision tree?

The tree is dichotomous and brutally simple. Each question eliminates one option. No middle ground, no “it depends” cop-out.

Q1 — Will the workflow execute >50 times/day in steady state?
  └─ YES → Skip n8n/Make. Go to Q3 (Custom).
  └─ NO  → n8n or Make are viable. Go to Q2.

Q2 — Will a non-technical operator own the workflow long-term?
  └─ YES → Make.com (better UX, easier handover).
  └─ NO  → n8n self-hosted (cheaper, more flexible, devs can fork it).

Q3 — Does the workflow require >3 LLM calls in sequence with branching logic?
  └─ YES → Custom code, period. n8n/Make collapse on agentic flows.
  └─ NO  → Go to Q4.

Q4 — Are there strict latency requirements (<2s end-to-end response)?
  └─ YES → Custom code (n8n/Make add 800ms–2s overhead per node).
  └─ NO  → n8n self-hosted is workable for high-volume non-agentic.

Decoded: about 70% of our client workflows hit Q3 = YES (>3 LLM calls with branching) → custom code wins. Of the remaining 30%, half hit Q2 = YES (non-tech owner) → Make.com. The remaining 15% are split between n8n self-hosted (devs in the loop, low volume) and n8n cloud (rare — usually a stop-gap).

What does each option actually do well?

Each platform has a clean, defensible sweet spot. Below is what they’re genuinely best at when used in their lane.

Platform Sweet spot Strength
n8n self-hosted Internal dev tooling, ETL, glue between APIs Free, self-hosted, 400+ nodes, version control via JSON export
n8n cloud Quick prototypes, demos Hosted, no DevOps, fast time-to-first-workflow
Make.com Operations team owns workflow, visual debugging Best UX of the three, error handling visual, scenario library
Custom code Production agents, tool-use loops, anything >50/day Full control, lowest cost-at-scale, debuggable by devs

The mistake we see most often: clients pick n8n cloud or Make for a workflow that will hit 10 000 executions/month within 6 months, then discover the platform’s monthly operations cap (Make: 10K Pro / 40K Teams) and have to upgrade 3 tiers or rebuild. That’s a 6–9 month detour costing $4K–$12K in wasted effort.

What are the hidden costs nobody mentions?

Each platform has a “tax” that isn’t on the pricing page. After 47 deployments, the pattern is consistent.

n8n self-hosted hidden costs

Real cost on a serious deployment: ~$200/month all-in (infra + maintenance time). Not the $0 the marketing implies.

Make.com hidden costs

Real cost on a serious deployment: $80–$300/month + the rebuild risk if you outgrow Enterprise.

Custom code hidden costs

Real cost: $80–$250/month at run time + 1–2 hours/week of dev attention + the upfront build investment. Linear-cheap at scale, but never zero.

When did we choose n8n at a real client?

A French B2B agency (anonymized) needed an internal workflow for their sales team: enrich every new HubSpot lead with company data from 4 sources (Apollo, Clearbit, LinkedIn scrape, a custom database), score it against 6 criteria, push the result back to HubSpot, and notify Slack if score >70. Volume: 30–50 leads/day in steady state. Owner: their CTO (technical, comfortable with JSON).

We built it on n8n self-hosted (Railway + managed Postgres). Why:

Reason Detail
Volume <50/day Far below the inflection where n8n breaks down
CTO owns it Will tweak nodes and read JSON without us
4 API integrations All have native n8n nodes; zero custom code needed
1 LLM call (scoring) Single node, no branching agentic logic
Budget tight Build was $1 800 vs $5 500 for custom equivalent

Result after 8 months: runs flawlessly, ~36 leads/day, $30/mo Railway infra. CTO has tweaked the scoring rubric 4 times without our involvement. This is the n8n win pattern.

When did we choose Custom Code at a real client?

A LATAM dropshipper (anonymized) needed: ingest ad spend from Meta + TikTok + Google, scrape Dropi remittances + 99minutos delivery status (no API), reconcile against Shopify orders, run a daily P&L per SKU, generate kill-signals for losing ad sets, and post a Telegram digest at 8 a.m. Volume: ~2 000 daily computations + 4 cron jobs + 1 LLM call per kill-signal explanation.

We built it on custom TypeScript on Railway + Supabase + Browserbase. Why:

Reason Detail
Volume >>50/day n8n/Make would hit cost or throughput ceiling in month 2
Browserbase + Stagehand integration No native nodes in n8n/Make; would require custom HTTP nodes anyway
Daily P&L computation Complex SQL + denormalized tables; n8n’s data manipulation tools collapse
Operator wants kill-signals via Telegram Custom message formatting + threading; trivial in code, painful in n8n
3-year horizon Run-cost on Make Enterprise would be $300–$500/mo; custom is $150/mo

Result after 6 months: ~5 100 $ ad-spend recovered in month 1, run-cost stable at $150/mo, dropshipper owns the code on his GitHub. This is the custom win pattern.

What’s the comparison table on the variables that matter?

We graded all 3 options on the 8 variables that actually drive the decision. Bias-free, based on production observation.

Variable n8n (self-host) Make.com Custom code
Visual builder ✓ (best UX)
Self-hostable
Version control (Git) ✓ (JSON export) ✓ (native)
AI-native nodes (LLM tools) Partial (HTTP wrappers) Partial ✓ (full SDK)
Cost per 10K executions ~$0.05 (infra amortized) ~$3.50 ~$0.10
Time-to-build (1st workflow) 4–8h 2–4h 16–40h
Time-to-modify (existing workflow) 30 min 15 min 1–4h
Failure recovery / debug Painful at scale Best built-in Best (your own logs)
Scales past 50K exec/day Yes (with tuning) No (cap) Yes
Vendor lock-in None Total None

The tradeoff is clear: Make.com wins on UX and time-to-build, custom wins on scale and cost-at-volume, n8n wins on the middle ground (technical owner, low-medium volume, no lock-in). Pick by use case, not by ideology.

What’s the recurring failure mode we see?

The most common failure is “we’ll use n8n / Make to ship fast, then migrate to custom if needed.” It almost never happens in practice. What happens instead: the workflow runs for 4–8 months in n8n/Make, accumulates implicit business logic in 27 nodes nobody documented, and then the migration to custom takes 3–4× longer than building custom from scratch because you’re reverse-engineering rather than designing.

If you suspect from day 1 that the workflow will exceed Q1’s 50/day threshold or hit Q3’s >3 LLM calls, just build custom from day 1. The 8–10 days of upfront engineering save 4–6 weeks of migration pain in month 8.

The exception: prototypes for product validation. If the workflow is “we want to test if our ICP cares about X,” n8n cloud or Make is correct — the goal is learning, not scale. Throw it away after 30 days regardless of result.

What’s the 2026 update — does AI-first tooling change the calculus?

Yes, modestly. Two new patterns matter:

  1. Code-generation IDEs (Cursor, Continue, Codex CLI) have collapsed the time-to-build for custom code by 40–60%. A workflow that took 16h to code in 2024 takes 6–9h in 2026. This shifts the break-even point earlier — custom now wins at lower volumes than before.
  2. n8n added native AI agent nodes in v1.40+ (LangChain integration, vector store nodes). For 1–2 LLM call workflows, n8n is now genuinely competitive with custom on AI tasks. But it still collapses on >3-step agentic loops with branching, because the visual debugger isn’t built for that.

Net effect: custom wins more often than it did 18 months ago, n8n closed some of the gap on AI workflows, and Make.com remains best-in-class for non-tech operator handover. Make.com hasn’t moved much — that’s their strategic position and they’re sticking with it.

What should you ask before picking a platform in 2026?

The single question that beats every other framework: “Who will own this workflow 12 months from now, and what will they do when it breaks at 3 a.m. on a Saturday?”

If the answer is “the founder, who can read JSON and wake up at 3 a.m.” → n8n. If the answer is “an ops manager who’ll click around and call a freelancer” → Make.com. If the answer is “a dev team with on-call rotation, monitored alerts, and a runbook” → Custom code. If the answer is “I don’t know” → Don’t ship the workflow yet. The “who owns it” question is the entire decision.


Sources: Openclaw internal data on 47 production workflow deployments (2024–2026). n8n self-hosting docs. Make.com pricing. Operation cap math verified against Make Pro plan limits. Cost-per-execution benchmarks measured on identical workloads (10K runs, simple HTTP+LLM pipeline) across the 3 platforms in March 2026.

Alexandre Bloch
Solo founder · Openclaw. I build custom AI agents for SMBs and B2B companies. Measurable ROI, fixed-cost build, code you own.

← All articles