✍️ 필사 모드: Workflow Automation 2026 — Zapier, n8n, Make and RPA (UiPath, Power Automate): The Line Between iPaaS and RPA, and the AI Agent Crossover

Prologue — Automation is no longer one category

Ten years ago it was simple. Zapier was a small bridge between SaaS apps. UiPath was a robot that clicked buttons in Windows apps for you. Same word — "automation" — but the two products lived on different planets.

In 2026 the landscape has shifted.

• Zapier launched "Agents," LLM agents with tool-calling on top of 9,000+ integrations.
• n8n crossed 185k GitHub stars and became the de facto standard for self-hostable AI workflow platforms. LangChain, vector DBs, Claude and Gemini nodes are in core.
• UiPath pivoted the whole company to "agentic automation." LLM agents now orchestrate RPA bots.
• Microsoft Power Automate made Copilot — "build a flow with natural language" — the default builder.
• Make.com leans into visual branching and routing at 3 to 5 times the cost-efficiency of Zapier.

The lines have blurred, but they have not disappeared. iPaaS and RPA still solve different problems. The first question is still "does this system have an API?" — and AI agents are stacking a new layer on top of both answers.

The plan of this post:

1. iPaaS vs RPA — what's actually different
2. 2026 tool landscape at a glance — comparison matrix
3. The iPaaS trio in detail — Zapier, Make, n8n
4. The RPA camp — UiPath, Power Automate, Automation Anywhere, Blue Prism
5. AI agent crossover — where both sides meet
6. The pricing model trap — tasks, ops, executions, bots, credits
7. Self-hosting vs SaaS — how n8n changed the balance
8. The same workflow in three tools — real code
9. A decision tree for tool choice
10. Anti-patterns

Read it once and you should be able to pick a tool in five minutes.

Chapter 1 — iPaaS vs RPA: What's Actually Different

1.1 The essence of iPaaS — bellows over an API

iPaaS (Integration Platform as a Service) routes API calls between cloud SaaS apps. Zapier, Make, n8n, Workato, and Tray live here.

Core assumptions:

• Every system has REST, GraphQL or webhook APIs.
• Auth is standardised on OAuth or API keys.
• Data is JSON.

Workflow shape:

• Trigger (e.g. a Stripe payment event) plus transform (JSON parsing, filtering, mapping) plus action (Slack notification).
• Execution is event-driven, typically completing in seconds to minutes.
• Concurrency and idempotency are partly the platform's job.

1.2 The essence of RPA — fingers over a UI

RPA (Robotic Process Automation) automates the UI actions themselves — mouse clicks, keyboard input. UiPath, Power Automate Desktop, Automation Anywhere and Blue Prism are the canonical names.

Core assumptions:

• The target system has no API, or the API is restricted. Legacy ERPs, mainframes, in-house Windows apps, desktop Excel.
• The bot finds and clicks UI elements by coordinates, image matching or accessibility tree.
• The unit is a bot — it imitates what a human did.

Workflow shape:

• A bot spins up in a Windows VM, opens SAP, types into a cell, clicks "Save."
• Execution is schedule- or queue-driven, and a single job can take minutes to hours.
• The bot assumes the screen looks the same every time. The day the screen changes, the bot breaks.

1.3 The one-line split

iPaaS shines when the system has an API. RPA is the last resort when it doesn't.

Memorise that line and 90 percent of tool selection is done. Plenty of real workflows need both — which is exactly why each camp is now invading the other.

Chapter 2 — 2026 Comparison Matrix at a Glance

You don't need to memorise the table. Find the row that's your deal-breaker within five minutes. It's usually one of three: must self-host, must have a code escape hatch, or must be no-code.

Chapter 3 — The iPaaS Trio in Detail

3.1 Zapier — the gold standard for non-developers

As of May 2026, Zapier is still the friendliest iPaaS for non-developers.

• 9,000+ integrations — even niche SaaS apps already have a connector. Discovery cost is effectively zero.
• Trigger plus action linearity — readable in one line: "When X happens, do Y."
• Zapier Agents — GA in May 2025, with templates unified into the main Templates hub in February 2026. An Agent is a tool-calling LLM wrapper; you can hand it all 9,000+ integrations as tools.
• Copilot builder — natural language to draft a Zap. "When a new Stripe payment arrives, notify Slack and add to HubSpot" produces a working draft.
• AI guardrails — scans for prompt injection, PII, toxic language. The door that lets enterprises through.

Weak spots:

• Tasks are expensive — 3 to 5 times pricier than Make on average. Heavy polling workflows blow up the bill.
• Weak visual branching — Paths exists, but it's clunky next to Make's visual routing.
• Limited code step — 30-second timeout, restricted package set.

3.2 Make.com (formerly Integromat) — home for visual thinkers

Make is built around scenarios, with strong visual wiring between nodes.

• Visual branching and routers — draw if/else trees on a single canvas. The answer when Paths feels cramped.
• Operation credits — one operation per step. 3 to 5 times cheaper than Zapier for equivalent workflows.
• Maia AI builder — natural language scenario generation. Make AI Agents shipped too.
• 3,000+ connectors — fewer than Zapier, but all the load-bearing ones are there.

Weak spots:

• Learning curve — the visual canvas is powerful but slightly intimidating at first.
• AI nodes are basically HTTP calls — not as deep as n8n's first-class AI nodes.
• No self-hosting — cloud only.

3.3 n8n — the dark horse developers love

n8n is the fastest-growing tool in the 2026 iPaaS camp. Past 185k GitHub stars, driven by an AI-, developer- and self-host-friendly combination.

• Apache 2.0 (core) plus commercial license terms — "fair-code." Self-host is free; only multi-tenant SaaS resale is restricted.
• Native AI nodes — Claude Sonnet 4.6, Claude Opus 4.7, Gemini, OpenAI, a LangChain Agent node, and vector DB nodes (Pinecone, Qdrant, Weaviate, Supabase pgvector) are baked into core.
• AI Agent node — tool calling, memory and multi-step reasoning in a single node.
• HTTP nodes plus a code node everywhere — JavaScript and Python inline.
• Workflow as code — export/import as JSON, commit to Git, review in PRs.

Weak spots:

• A stretch to call it no-code — there's a real learning curve for non-developers.
• Operational burden — if you self-host, the DB, queue, scaling and backups are yours.
• Ecosystem is thinner than Zapier's — a niche SaaS connector may need building.

What makes n8n really interesting: it's the first tool to fully erase the line between iPaaS and AI agents. Workflows are themselves the tool-call graph for an agent.

Chapter 4 — The RPA Camp: People Automating UIs

4.1 UiPath — the category leader's pivot

UiPath spent 2025 and 2026 pivoting the whole company to agentic automation. It admitted the limits of click-only RPA and stacked an LLM-agent layer on top.

Core new products:

• UiPath Agent Builder — build LLM-based agents. Exceptions and decisions move from bots to agents.
• UiPath Maestro — orchestrates agents, bots and humans. Preventing "agent sprawl" is an explicit goal.
• ScreenPlay — agentic UI automation; AI "sees" the screen and clicks like a human, replacing coordinate-based fragility.
• RAG built in — agents pull real-time knowledge from Snowflake, SharePoint, Confluence, policy docs.

Target: large IT shops, finance, insurance, healthcare — places where governance, audit logs and an on-premises Orchestrator are deal-breakers.

4.2 Microsoft Power Automate — iPaaS, RPA and AI under one roof

Power Automate bundles three things under one umbrella.

• Cloud flows (iPaaS) — 1,400+ connectors, deep ties into M365, Teams and SharePoint.
• Desktop flows (RPA) — Power Automate Desktop, bundled with Windows 11.
• Copilot and AI Builder — natural-language flow generation, OCR, document extraction, GPT actions.

Pricing (April 2026 reference):

• Premium 15 dollars per user per month — cloud flows plus some desktop.
• Process 150 dollars per bot per month — unattended RPA.
• Hosted Process 215 dollars per bot per month — Azure-hosted bots.
• AI Builder bundles 5,000 credits with Premium; an extra unit is 500 dollars.

Strengths:

• Effectively free in M365-heavy orgs — partly bundled with E5.
• Tight Copilot Studio integration — chatbots, agents and flows live in one umbrella.

Weak spots:

• Cloud lock-in — desktop is local, but cloud flows are tied to Microsoft Cloud.
• Pricing complexity — per-user, per-bot and AI credits mixed in one invoice.

4.3 Automation Anywhere — cloud-native RPA

• Automation 360 — SaaS-based RPA. IQ Bot handles unstructured documents.
• AI Agent Studio — joined the 2025 / 2026 agent builder trend.
• Pricing is quote-based; estimates put enterprise contracts starting around 50,000 dollars per year.

4.4 Blue Prism (SS&C Blue Prism) — the regulated industries' fortress

• Since SS&C's 2022 acquisition, focused on highly regulated industries like finance and insurance.
• April 2026: SS&C Blue Prism WorkHQ launched — integrating agentic AI securely and transparently.
• Server-based architecture, on-prem strength and audit logs are the core selling points.
• Market share has slipped against UiPath and Power Automate; Gartner classifies them as challenger.

Chapter 5 — AI Agent Crossover: Where Both Sides Meet

The real 2026 story is this — iPaaS and RPA are converging on top of AI agents.

5.1 iPaaS reaching up — Zapier Agents, n8n AI Agent

Typical flow:

1. Trigger (e.g. new email) hands context to an LLM.
2. The LLM picks an action from its tool list (i.e. the iPaaS's 9,000 integrations).
3. Output goes back to the LLM, which decides the next step.
4. Stop at a branch needing human approval (human-in-the-loop).

n8n's AI Agent node expresses all of this as a single node. Declare tools, memory and model — done.

5.2 RPA reaching up — UiPath Agent Builder and Maestro

Typical flow:

1. A human submits work in natural language ("Review these 100 transactions and flag the suspicious ones").
2. Maestro decomposes the work and assigns each step to either a bot or an LLM agent.
3. Bots click through screens; agents use RAG to consult policy docs.
4. Results go to a human for review; the bot performs the final action.

UiPath's ScreenPlay is decisive here — a visual-understanding agent replaces the brittle coordinate-driven parts of RPA.

5.3 MCP ties both sides together

Model Context Protocol has become the de facto standard for connecting LLMs to tools since 2025.

• Zapier shipped Zapier MCP, exposing 9,000+ integrations as MCP tools.
• n8n can publish a workflow as an MCP server.
• Power Automate's Copilot Studio invokes MCP tools.

So a workflow built once can be used as a tool by any LLM host — Claude Desktop, Cursor, ChatGPT.

Chapter 6 — The Pricing Model Trap

The most common automation incident is misreading the pricing model and watching the invoice explode.

6.1 Tasks vs operations vs executions

• Zapier tasks — one task per executed action. Triggers are free. Tight polling burns tasks fast.
• Make operations — one operation per node that processes data. Branches and loops multiply cost.
• n8n workflow executions — applies on cloud. Self-hosted is free, but you carry DB and server costs.

6.2 RPA bot licensing

• Per-bot — UiPath, Power Automate Process. One license per bot that runs 24/7.
• Per-user — Power Automate Premium. Pay per user.
• AI credits — Copilot and AI Builder are separate, and heavy usage moves to a custom quote.

6.3 Hidden costs

• Data egress — every large payload through a workflow can be billed.
• Storage — execution and error logs pile up; default 30-day retention applies.
• Compute — UiPath Hosted Process bills Azure compute on a separate line.

6.4 Quote comparison example

Imagine 100,000 actions per month (each action equals one trigger plus three actions).

• Zapier: 300,000 tasks. Team or Company plan, 500 to 1,500 dollars per month.
• Make: 400,000 operations (four-node scenario). Teams plan plus add-ons, 150 to 300 dollars per month.
• n8n self-hosted: one server (around 30 dollars per month) plus ops time.
• n8n Cloud Pro: 50 to 200 dollars per month depending on volume.

A 3x to 10x difference. If the workload is mostly polling, moving to Make or n8n is the rational choice.

Chapter 7 — Self-hosting vs SaaS: How n8n Changed the Balance

7.1 Why self-host

• Data sovereignty — data handled by workflows never crosses an external cloud. GDPR, HIPAA, sector regulators.
• Unlimited executions — pricing is bounded by your server cost; one million runs cost the same.
• Custom nodes — write nodes that hit your internal APIs directly.
• Avoid vendor lock-in — workflow JSON in Git, portable any time.

7.2 Why SaaS

• No ops burden — upgrades, backups, scaling and monitoring are all the vendor's job.
• Fast start — first workflow live within five minutes of sign-up.
• Battle-tested integrations — official OAuth flows, webhooks and retry logic come verified.
• SLAs — enterprise contracts are possible.

7.3 n8n's compromise — lowering the cost of self-hosting

n8n drastically lowered the bar for self-hosting.

• Start with one line of Docker Compose.
• One-click deploy via AWS, GCP and Azure marketplaces.
• Managed-self-host options — vendor takes hosting and upgrades.
• Community Edition is permanently free with unlimited executions.

This compromise puts pricing pressure on SaaS-only Zapier and Make. A lot of n8n adoption is rumoured to be people fleeing the Zapier invoice.

7.4 Self-host operations checklist

# n8n production self-hosting minimums
infra:
  app: 2 vCPU / 4GB RAM or more
  db: PostgreSQL 15 plus (managed RDS recommended)
  queue: Redis (queue mode, mandatory above 100 runs per minute)
  storage: 30GB or more (execution logs plus binary data)

network:
  domain: workflows.example.com
  tls: Let us Encrypt or ACM
  reverse_proxy: nginx or Caddy

ops:
  backup: PostgreSQL daily snapshot, 7-day S3 retention
  monitoring: Prometheus plus Grafana, alarms on failure rate
  upgrade: monthly window with pinned Docker tag
  secrets: AWS Secrets Manager or 1Password Connect

Chapter 8 — The Same Workflow in Three Tools

Workflow: when a new GitHub Issue opens, post a summarised notification to a Slack channel called triage.

8.1 Zapier — trigger plus action

In the Zapier UI:

1. Trigger: GitHub plus New Issue. Pick the repo and connect via OAuth.
2. Action: Slack plus Send Channel Message.
   • Channel: triage
   • Message:

New issue: {{title}}
{{user.login}} - {{html_url}}

{{body}}

3. Test, then publish.

Pros: live in five minutes. UI is one screen. Zero code. Cons: limited message formatting. Anything fancier needs a code step.

8.2 Make — visual scenario

In Make:

1. GitHub: Watch Issues module, pick the repo.
2. Tools plus Set Variable: summary equals the bracketed repo name plus the issue title.
3. Slack: Create a Message module.

The three nodes are visually wired with arrows. Adding a router branches by label into different channels with one extra node.

Pros: branching and aggregation are intuitive. Cheaper than Zapier. Cons: one operation per node, multiplied by every branch.

8.3 n8n — workflow as code

The same workflow in n8n, as JSON:

{
  "name": "github-issue-to-slack",
  "nodes": [
    {
      "name": "GitHub Trigger",
      "type": "n8n-nodes-base.githubTrigger",
      "parameters": {
        "owner": "your-org",
        "repository": "your-repo",
        "events": ["issues"]
      },
      "position": [240, 300]
    },
    {
      "name": "Format Message",
      "type": "n8n-nodes-base.code",
      "parameters": {
        "language": "javaScript",
        "jsCode": "const issue = items[0].json.issue;\nconst repo = items[0].json.repository;\nreturn [{ json: { text: `*New issue in ${repo.full_name}*\\n<${issue.html_url}|#${issue.number} ${issue.title}>\\nby ${issue.user.login}` } }];"
      },
      "position": [460, 300]
    },
    {
      "name": "Slack",
      "type": "n8n-nodes-base.slack",
      "parameters": {
        "channel": "triage",
        "text": "={{ $json.text }}"
      },
      "position": [680, 300]
    }
  ],
  "connections": {
    "GitHub Trigger": { "main": [[{ "node": "Format Message", "type": "main", "index": 0 }]] },
    "Format Message": { "main": [[{ "node": "Slack", "type": "main", "index": 0 }]] }
  }
}

Commit that JSON to Git and workflow changes become PR-reviewable.

Pros: treatable as code; self-host is free and unlimited; the code node is powerful. Cons: bar for non-developers; you won't hand-author the JSON but you'll read it during reviews.

8.4 Layering AI on the same workflow — five minutes in n8n

Add a step where an LLM summarises the issue body and recommends a label.

[GitHub Trigger]
        ↓
[Anthropic Chat Model]   - Claude Sonnet 4.6 node
  prompt: "Summarise the issue in one sentence and recommend one label
          from bug, feature, or docs.\n\n{{ $json.issue.body }}"
        ↓
[Format Message]
        ↓
[Slack]   "[{{ $json.label }}] {{ $json.summary }}"
        ↓
[GitHub: Add Label]   - auto-apply if the label is bug

n8n's Anthropic Chat node exposes model, temperature and system prompt in the UI. Five minutes and the AI classifier is in place.

8.5 The same workflow in RPA — don't

This is not an RPA problem. GitHub and Slack both have first-class APIs, so iPaaS is the answer. Reach for RPA only when one side genuinely lacks an API.

Chapter 9 — A Decision Tree for Tool Choice

A simple tree to pick a tool in five minutes.

Does the system have an API?
- Yes
  - Users only non-developers?
    - Yes - Zapier (shortest learning curve)
    - No, lots of visual branching?
      - Yes - Make.com
      - No, AI plus self-host plus code needed?
        - Yes - n8n
- No (Windows app, mainframe, legacy ERP)
  - Already on M365?
    - Yes - Power Automate Desktop plus Premium
    - No
      - Large IT, finance, insurance - UiPath or Blue Prism
      - Cloud-native preferred - Automation Anywhere

Extra branches:

• Need AI agents? -> n8n (developers), Zapier Agents (non-developers), UiPath Maestro (enterprise RPA).
• Need data sovereignty? -> self-hosted n8n or on-prem UiPath Orchestrator.
• More than 10,000 runs per hour? -> n8n in queue mode or an RPA-grade bot pool.

Chapter 10 — Anti-Patterns

10.1 Faking an API with RPA

A team buys RPA for a legacy ERP, then automates the next system the same way. The next system has an API, and the bot is clicking screens. Bots are fragile and roughly 100 times slower than API calls. If an API exists, always use the API.

10.2 Every workflow on one tool

"We're a Zapier shop, every workflow goes through Zapier." Result: one giant workflow blows up the task bill, and a code step hits the 30-second timeout. Workflow size and frequency decide the tool. Small things in Zapier, large in n8n, UI automation in RPA.

10.3 Self-hosting without an operator

Spin up n8n on EC2, walk away. Six months later the disk is full, PostgreSQL has no backups, or the certificate expired. Self-hosting is not free. Compare ops cost against the SaaS bill before committing.

10.4 Giving the agent unlimited tools

A Zapier Agent is granted "use any of the 9,000 integrations." A single prompt injection later, Slack is full of spam. Least privilege for agent tools. Guardrails are not optional.

10.5 Changing workflows without review

"It's not really code, no need for a PR" — one person breaks a workflow and the entire team stops. Workflows are code. Make and Zapier both export scenarios. n8n submits JSON via PR.

10.6 Multiplying RPA bots forever

"This task gets a bot, that task gets a bot." Past 100 bots, every break locks up a human for a week. Bots are debt. Compute ROI per bot. If it does not pay back in six months, retire it.

Epilogue — Not "Which Tool" But "What Landscape"

Automation tool selection is not a single choice like "Zapier or n8n." It's landscape design.

Adoption checklist:

• List 10 systems to automate; mark which have APIs.
• Estimate workflow frequency per hour and data volume per system.
• Note the mix of non-developers, developers and RPA specialists.
• Check data sovereignty and regulatory needs (GDPR, HIPAA, sector rules).
• Simulate the six-month invoice — tasks times operations times bots.
• Predict if AI agents will be required within 12 months.
• First decision: iPaaS, RPA or both.
• Second decision: pick one tool and run a two-week PoC.
• Commit workflow JSON or exports to Git.
• Assign an operations owner — especially if self-hosting.

Short anti-patterns list:

• One tool for every workflow — costs explode.
• API exists but RPA clicks the screen — API first, always.
• Treating self-hosting as free — count the operator hours.
• Maximum-permission agents — never run without guardrails.
• Workflow changes without review — change tracking collapses fast.

Next post

The next post is a n8n self-hosting production guide — from a single EC2 box to queue mode, HA and observability, walked through with the real timeline of one company. PostgreSQL tuning, splitting out the Redis queue, execution data retention, failure alarming, the Git workflow — the things that broke over six months. If your Zapier invoice has just crossed 3,000 dollars a month, that post is the next decision frame.

Automation tools are not the destination. They are the tool. Design the landscape. When the landscape is right, the tools just do their job.

References

iPaaS

• Zapier — Automate AI Workflows, Agents, and Apps
• Zapier Agents — Combine AI agents with automation
• Zapier February 2026 Updates — AI guardrails & governance
• Make.com — AI Workflow Automation
• Make.com Pricing 2026 (G2)
• n8n — Workflow Automation Platform
• n8n on GitHub (185k+ stars)
• n8n Self-hosted AI Starter Kit
• n8n Pricing (Cloud vs Self-hosted, 2026)

RPA / Agentic Automation

• UiPath — Business Orchestration & Automation Platform
• UiPath Moves Into Agentic AI Realm (AI Business)
• Top 10 UiPath AI & Agentic Trends for 2026 (Accelirate)
• Microsoft Power Automate Pricing
• Power Automate 2026 Release Wave 1 (Microsoft Learn)
• Automation Anywhere — Automation 360
• SS&C Blue Prism — Agentic Automation
• SS&C Blue Prism WorkHQ Launch (April 2026)

Standards, Comparisons, Guides

• Model Context Protocol (MCP)
• Make.com pricing — Is it worth it? (Zapier blog)
• Blue Prism vs Automation Anywhere 2026 (Kanerika)
• n8n Review — AI Agents and Automation (Roborhythms)