n8n vs Make.com vs a Custom Python Agent: Picking the Right Automation Layer Before You Regret It
Most UAE SMEs pick an automation tool because a salesperson demoed it or a YouTube tutorial covered it. Six months later they're rebuilding. Here's the part nobody tells you upfront: the tool you can switch away from cheaply matters more than the one that's fastest to start. This guide hands you the real decision criteria — pricing that holds up under load, PDPL data residency exposure, where each tool's logic ceiling actually sits, and what it costs to migrate when you guessed wrong. Choose for the workflow you'll be running in 18 months, not the one on your desk this week.
Make.com: Fast Start, Real Ceiling, Genuine PDPL Risk
Make.com's Core plan runs $9 a month for 10,000 credits at the entry tier. Cheap, until you do the arithmetic. A three-step workflow burns three credits every run. AI module steps cost extra on top of that, billed by token volume, file size, and page count, and Make Code adds 2 credits per second. So a brokerage parsing 50 deals a month can drain Core in days and find itself on Pro ($16) or Teams ($29) before the quarter closes. What you buy with that money is genuine speed: a fast drag-and-drop builder and more than 3,000 integrations. Watch the moving target, though. Make switched from operations to credits on 27 Aug 2025, then adjusted limits and costs again on 6 Nov 2025, so re-check your tier before you commit.
Data residency is where Make stops being an option for regulated sectors. You get US and EU regions only, chosen when the org is created and never movable afterward, with no UAE region at all. The EU choice does not close the PDPL gap. Health data routed offshore breaches Federal Law No. 2 of 2019, and PDPL (Decree-Law 45/2021, UAE Data Office) sets the same expectation for mainland personal data. Non-personal operational data is fine here. The moment personal data enters the workflow, check where it lands.
The Billing Mechanic Nobody Reads Until the Invoice: Operations vs Executions
This is the one cost driver that decides your bill, so it is worth understanding before you pick a tool. Make bills per operation, meaning every module on every run. n8n Cloud bills per execution, where one run counts as 1 no matter how many nodes fire, so a 2-step flow and a 200-node agent cost exactly the same. n8n self-hosted bills nothing per run: Community Edition is free with unlimited executions, and you pay only for the VPS underneath it.
Run the crossover on a 5-step flow at 1,000 runs a day. On Make that is 150,000 operations a month, which means Core's 10,000 are gone by day two and you are paying your way up the slider. On n8n Cloud it is 30,000 executions a month, clearing Pro's 10k and landing on Business 40k at roughly EUR 667/mo on the annual plan. Self-hosted, it is 30,000 runs on a flat $20-50/mo VPS, where the step count costs you nothing. One detail in n8n Cloud's favour: it counts only successful production executions, so tests and failures are free. The rule that falls out of this is simple. For 3-5 step, low-volume flows, Make is the cheapest. For many-step, AI-agent, or high-volume work, n8n self-hosted wins on the billing model itself, not on vendor preference.
n8n Self-Hosted: Data Residency Solved, DevOps Required
n8n Community Edition is free with no execution cap. A production VPS (2 vCPU, 4GB RAM, 40GB SSD plus managed PostgreSQL) runs $20-50/mo, around AED 73-184. Put that on AWS me-central-1 (UAE, live 8 Nov 2022, 3 AZs) and every byte stays in-jurisdiction. AWS is not the only door, either: Azure UAE North/Central and Oracle UAE give you the same in-country residency, so this is not a single-vendor recommendation. That setup satisfies PDPL and the Federal Law 2/2019 health-data requirement for clinics on NABIDH (National Backbone for Integrated Dubai Health) or Malaffi (Abu Dhabi).
On capability, n8n carries roughly 400 integrations, JS code nodes, and more than 70 native LangChain AI nodes right on the canvas. Workflows export as JSON, which gives you version control and a real audit trail. The price of that flexibility is operational, and the next section is the honest invoice for it. Where n8n fits: you have a technically comfortable person on hand, an 18-month horizon, and non-negotiable residency. Without those, weigh the maintenance burden carefully before you own a server.
What "Self-Hosted" Actually Puts on Your Plate (Beyond the AED 73 Server)
Four things are not optional, and together they decide whether the deployment survives production.
First, the encryption key. N8N_ENCRYPTION_KEY encrypts every stored credential, so set it explicitly and back it up. Leave it to auto-generate and it can change on restart or vanish entirely, at which point every saved credential becomes permanently undecryptable with no recovery path. Second, the database: use PostgreSQL, not SQLite. SQLite locks under load and cannot run queue mode. Third, queue mode itself, which spreads concurrent execution across worker containers and needs Redis as its broker to scale horizontally. Fourth, the front door. Webhook endpoints need signature or token auth, all of it behind a reverse proxy (Caddy, Nginx, or Traefik) terminating TLS on a stable domain.
This is exactly why the previous section sets the "one technical person plus 18-month horizon" test. For a clinic with nobody on DevOps, the server rental is the trivial line item. The on-call burden is the real cost. Price the on-call, not just the VPS.
Custom Python Agent: Full Control, Real Engineering Cost
It is worth correcting a common assumption: doing AI does not mean leaving n8n. n8n ships more than 70 LangChain nodes plus a LangChain Code node. The catch is that they are JavaScript-only and self-hosted-only, with no Python on the canvas. The actual case for going code-first is the work that turns brittle inside a visual builder: dynamic branching on unpredictable runtime data, runtime-enforced typed I/O, native on-prem vLLM endpoints, and audit-grade state that survives failures.
Two tools cover most of this. PydanticAI (stable 1.0, Sep 2025) handles linear typed flows. LangGraph handles stateful branching, human-in-the-loop pauses, and audit logging, and a PydanticAI agent composes neatly inside a LangGraph node. There is also a hybrid worth knowing, because it defuses the fear of a migration cliff: keep n8n as trigger and glue, then call out to a small Python service for the one hard reasoning step. A full-custom build runs AED 10,000-20,000 plus maintenance. That spend is justified when a stateful, multi-path process has outgrown the visual canvas: a 12-step legal conflict-check, say, or a clinic intake/lab/billing pipeline. And only when the organisation can own what gets built.
The Migration Cliff: How to Avoid Building Three Times
The rebuild cycle, Make to n8n to Python, burns time, money, and institutional knowledge. It is worse than it sounds, because Make scenarios do not import into n8n. There is no automated converter; the visual model has no node-graph equivalent, so every migration is a rewrite from scratch. Documented Zapier-to-n8n moves show the prize and the price together. One shop went from roughly $1,247/mo on Zapier to roughly $79/mo on a VPS and saved about $14,000 a year, with smaller workloads still clearing $5,000+. But every one of those migrations cost real developer hours, which the headline number never mentions.
The decision falls straight out of an 18-month roadmap. Simple, non-personal flows go on Make Core at $9. Anything touching PDPL or health data goes on n8n self-hosted in a UAE region (AWS me-central-1, Azure UAE North, or Oracle UAE) as the floor, but check the client's jurisdiction first: DIFC runs under DP Law No. 5 of 2020 and ADGM under its DP Regulations 2021, both separate regimes carved out of federal PDPL. Anything with 4-5 or more conditional branches, persistent state, or on-prem LLM orchestration goes straight to the Python agent or an n8n-plus-Python hybrid, skipping the dead-end middle tiers entirely. Planning this once costs far less than building it three times.
Questions about your setup?
We help UAE SMEs build AI systems that are compliant, on-premise, and actually useful. Free initial conversation.