2nd/10_Wiki/Topics/AI_and_ML/Process-Automation-with-AI.md

---
id: wiki-2026-0508-process-automation-with-ai
title: Process Automation with AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI RPA, Intelligent Automation, AI Agent Automation]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [automation, ai-agents, rpa, workflow, llm]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python/JavaScript
  framework: n8n/Zapier/LangGraph
---

# Process Automation with AI

## 매 한 줄
> **"매 deterministic workflow + LLM reasoning step"**. 2010s RPA (UiPath, Automation Anywhere) 가 매 brittle screen-scraping 이었다면, 매 2026 현재 hybrid AI agent — 매 structured workflow node + LLM decision node — 가 default. n8n + Claude tool-use 의 dominant pattern.

## 매 핵심

### 매 spectrum
- **Pure RPA**: 매 deterministic, 매 scripted UI clicks. UiPath, Power Automate.
- **iPaaS**: API-first integration. Zapier, Make, n8n.
- **AI-augmented iPaaS**: 매 LLM step 의 추가. Zapier AI Actions, Make AI modules.
- **Agent automation**: 매 LLM 의 plans + executes. Claude tool-use, LangGraph, CrewAI.
- **Computer-use agent**: 매 LLM 의 screen + mouse 의 직접 control. Anthropic Computer Use, OpenAI Operator.

### 매 architecture
- **Trigger** (webhook, schedule, email, file).
- **Extract** (LLM parses unstructured → JSON).
- **Decide** (LLM chooses branch / tool).
- **Act** (API call, DB write, send email).
- **Verify** (LLM judges output, human-in-loop on low confidence).

### 매 응용
1. Invoice processing (PDF → ERP).
2. Customer support triage (email → ticket category + draft reply).
3. Lead enrichment (CRM + LinkedIn + LLM summary).
4. Code review automation (PR → AI comments).
5. Compliance monitoring (logs → policy check).

## 💻 패턴

### n8n + Claude (2026 standard)
```json
{
  "nodes": [
    {"type": "webhook", "name": "Email received"},
    {"type": "anthropic", "name": "Classify",
     "params": {
       "model": "claude-opus-4-7",
       "system": "Classify email: SUPPORT, SALES, SPAM. Return JSON.",
       "input": "={{ $json.body }}"
     }},
    {"type": "switch", "rules": [
      {"value": "SUPPORT", "output": 0},
      {"value": "SALES", "output": 1}
    ]},
    {"type": "zendesk", "name": "Create ticket"}
  ]
}
```

### LangGraph state machine
```python
from langgraph.graph import StateGraph, END
from anthropic import Anthropic

client = Anthropic()

def classify(state):
    r = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=200,
        messages=[{"role": "user",
                   "content": f"Category? {state['text']}"}]
    )
    return {"category": r.content[0].text.strip()}

def route(state):
    return state["category"].lower()

g = StateGraph(dict)
g.add_node("classify", classify)
g.add_node("support", lambda s: {"reply": "Support team handling"})
g.add_node("sales", lambda s: {"reply": "Sales team handling"})
g.set_entry_point("classify")
g.add_conditional_edges("classify", route,
    {"support": "support", "sales": "sales"})
g.add_edge("support", END)
g.add_edge("sales", END)
app = g.compile()
```

### Tool-use agent (Claude)
```python
tools = [
  {"name": "create_ticket",
   "description": "Create Zendesk ticket",
   "input_schema": {"type": "object",
     "properties": {
       "subject": {"type": "string"},
       "priority": {"type": "string", "enum": ["low","high","urgent"]}
     }, "required": ["subject"]}},
  {"name": "send_slack",
   "description": "Notify channel",
   "input_schema": {"type": "object",
     "properties": {"channel": {"type":"string"}, "msg":{"type":"string"}}}}
]

resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=2000,
    tools=tools,
    messages=[{"role":"user", "content": email_text}]
)
for block in resp.content:
    if block.type == "tool_use":
        result = dispatch(block.name, block.input)
```

### Computer Use (Anthropic, 2025+)
```python
# 매 LLM 의 screenshot 의 보고 click/type.
# Brittle UI 의 RPA 의 대체.
response = client.beta.messages.create(
    model="claude-opus-4-7",
    tools=[{"type": "computer_20250124",
            "name": "computer", "display_width_px": 1920,
            "display_height_px": 1080}],
    messages=[{"role":"user",
        "content":"Open SAP, navigate to PO #4521, approve."}]
)
```

### Verification + HITL (human-in-loop)
```python
def confidence_gate(decision, threshold=0.85):
    if decision["confidence"] < threshold:
        send_to_human_queue(decision)
        return None
    return execute(decision["action"])
```

### Idempotent retry
```python
@retry(stop=stop_after_attempt(3), wait=wait_exponential())
def safe_api_call(idempotency_key, payload):
    return requests.post(url, json=payload,
                         headers={"Idempotency-Key": idempotency_key})
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Simple SaaS-to-SaaS sync | **Zapier / Make** |
| Self-host + complex logic | **n8n** (default 2026) |
| Stateful multi-step agent | **LangGraph + Claude** |
| Legacy desktop GUI 만 | UiPath / Computer Use |
| Engineering team automation | Temporal + LLM step |

**기본값**: n8n self-hosted + Claude tool-use node.

## 🔗 Graph
- 부모: [[Workflow-Automation]]
- Adjacent: [[Tool-Use]] · [[LangGraph]]

## 🤖 LLM 활용
**언제**: 매 unstructured input (email, PDF, chat), 매 fuzzy classification, 매 multi-step planning.
**언제 X**: 매 high-volume deterministic ETL — 매 SQL/Airflow 가 fast + cheap. 매 LLM call 의 매 step 의 cost overrun.

## ❌ 안티패턴
- **LLM in tight loop**: 매 step 의 매 LLM call — 매 latency + cost. 매 batch / cache.
- **No HITL on irreversible**: 매 send email / charge card 의 human approval gate 의 필수.
- **Schema-less tool output**: 매 free-text 의 parse error. 매 JSON schema enforce.
- **Hidden non-determinism**: 매 prompt 의 minor change 의 production 의 break. 매 eval suite 의 필요.

## 🧪 검증 / 중복
- Verified (Anthropic agent docs, n8n.io, LangGraph docs).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — modern AI agent automation 의 full content. |