Marketing teams want leads scored and routed. Ops wants tickets summarised and tagged. Engineering wants the same flows to survive audit, handle 10k runs a day, and call Claude with tool use, without a spaghetti of Zaps nobody can version-control.
At PAI Technologies we ship automations across the spectrum: n8n and Make for fast internal glue, FastAPI and LangGraph when branching, evals, PII redaction, and EU hosting requirements dominate. This article is how we choose, and how clients avoid rebuilding everything at month three.
The wrong question is “no-code vs code.” The right question is what must be testable, deployable in your VPC, and observable when legal asks what happened to user data last Tuesday.
What no-code orchestration is great at
n8n, Make (Integromat), and Zapier excel at connecting SaaS systems: CRM → Slack → spreadsheet, form → email → Notion. Time-to-value is hours when auth and field mapping are straightforward.
Self-hosted n8n appeals to teams that want EU-friendly deployment without writing a worker fleet from scratch. We have wired n8n to webhooks that call Claude for classification when the flow is linear and volume is moderate.
Use no-code when stakeholders are non-engineers who must own changes, when logic is mostly linear, and when compliance scope is low (no sensitive health data, no complex retention rules).
Where visual tools start to hurt
Branching hell: ten IF nodes for model routing, language detection, and escalation paths become unreadable. Code expresses state machines; canvases hide edge cases.
Version control and review: PRs with tests beat exporting JSON graphs nobody diffed. For SOC2-minded clients, change management expects git history.
Evals and quality gates: “call Claude and hope” is not production. LangGraph or FastAPI services can run golden-set checks before side effects execute (send email, update CRM).
Secrets, rate limits, and idempotency: retries that double-charge or double-email are common in naive Zap chains. Code gives explicit idempotency keys and backoff.
When we reach for FastAPI and LangGraph
FastAPI services host webhooks, auth, structured logging, and Claude/OpenAI calls with Pydantic-validated IO. Deploy to client VPC or eu-central alongside Gordon-class apps.
LangGraph fits multi-step agents: retrieve policy, draft reply, human approval, then send, with checkpointing and replay. PlaywithDB-adjacent flows often need governed DB tools, not only HTTP nodes.
We add Postgres for job state, dead-letter queues, and audit tables. Ops automations that touch GDPR or HIPAA data get retention jobs in the same codebase, not a forgotten Make scenario.
- Requirement for custom tool use (internal APIs, SQL, vector search)
- p95 latency or cost caps that need caching and model routing in code
- DPO asks for immutable audit log of inputs/outputs per run
Hybrid pattern we use often
n8n triggers on schedule or form submit → calls PAI-hosted FastAPI endpoint that runs RAG + Claude + eval → returns structured JSON → n8n routes to Slack/HubSpot. Business users keep the edges; engineering owns the brain.
Make handles marketing experiments; production path migrates to code after validation. We document the promotion criteria so “temporary” Zaps do not run for two years.
Zapier remains fine for solo-founder stacks with low risk. We rarely recommend it for EU-regulated AI at scale.
Compliance and observability
Map subprocessors in the flow: n8n cloud vs self-hosted, model provider, CRM. EU clients need this table in week one, same as Gordon onboarding.
Log with redaction: store run IDs, model version, retrieval sources; avoid raw PII in third-party automation logs when possible.
For ops automation without breaking compliance, pair this article with our GDPR guardrails piece, retention and lawful basis still apply when “it is just internal.”
Decision checklist
Start no-code if you need proof in days and data is low-sensitivity. Plan code migration when volume, branching, or audit requirements appear. We can estimate both phases.
Start in code if EU hosting, health/finance data, or agentic tool use is day-one. Three-to-four-week vertical slices still apply.
Send your current stack and flows in a brief; we will recommend n8n/Make vs FastAPI/LangGraph honestly, including when you should keep what already works.