AI Data Analysis SOP: Build a Sleepless Data-Analysis Digital Worker on Coze

Stop hand-grinding reports. While you nest VLOOKUPs to eternity in Excel, a few have built an AI data-analysis digital worker that never sleeps. It auto-cleans data, runs statistical models, writes insights, and proactively WeChats you on business anomalies. This isn't sci-fi-it's 2026.

Your Data Workflow Is Getting "Dimensionally Struck"

9 AM: export channel traffic/conversion -> Excel pivot ROI -> manually copy to PPT -> draw charts -> write "up 2.3% MoM." This occupies 70%+ of a data analyst's time. You're "human ETL + human BI."

New script: an "AI digital clone"-toss raw data into chat, it auto:

• Python cleans missing/anomaly values
• Computes core metrics (ROI/LTV/retention) per business definitions
• LLM attributes changes: "up 2.3% MoM because Douyin Qianchuan creative A's CTR spiked"
• Generates visualizations + Markdown weekly report
• Proactively alerts via DingTalk/WeCom when DAU crashes

ByteDance Coze lowers the bar to "drag-and-drop." This piece teaches you to build a production-ready data-analysis agent on Coze.

Want a "Digital Worker," Not a "Chatbot"

Pure chatbots have three hard flaws: no memory/state, can't execute complex workflows, limited tool calls.

A real data-analysis agent is a modular "digital worker," four layers:

• Perception (data ingest): Excel/CSV/DB/Feishu tables/API (Douyin/Business Advisor). Coze: "File Box" and "Plugins."
• Brain (smart analysis): Orchestrated chain: data exploration -> statistical modeling -> anomaly detection -> attribution. Coze "Workflow" nodes call LLM + code interpreter.
• Execution (model management): Python code interpreter for stats/hypothesis tests/ML. Coze "Code Node" native Python sandbox with pandas/numpy/scikit-learn pre-installed.
• Output (presentation + alerts): matplotlib/seaborn charts + HTML/PPT reports + Feishu/DingTalk/WeCom/email alerts.

Platform Comparison: Why Coze?

Domestic users going fast pick Coze: no network fuss, no code, full chain from ingest to alerts.

Hands-on: Clone an "E-commerce Weekly Agent"

Need a daily "Douyin Livestream Operations Report": GMV/viewers/conversion/Top 3 hits/YoY trend/anomaly alerts. Previously 2 hours/day manual; now 20 minutes to build the agent.

Step 1: Persona & Prompt Engineering

Coze new Bot "E-commerce Data Analyst Xiao D." Prompt template:

# Role
You are a 10-year e-commerce data analyst expert, skilled at finding growth via data.

# Core Responsibilities
1. Receive uploaded Douyin livestream data CSV.
2. Auto-clean: handle missing values, remove anomalies (GMV≤0 test data).
3. Compute key metrics: daily avg GMV, MoM growth, conversion (buyers/views).
4. Identify Top 3 hit products, analyze growth reasons.
5. Detect anomalies: single-day GMV波动 > 2 std dev, output alert.
6. Generate report with charts and text insights, Markdown format.

# Output Rules
- Chinese output, professional but accessible.
- Numbers to 2 decimals.
- Charts via Python, upload image links.
- Conclusions data-based, no fabrication.

# Limits
- No non-data-analysis questions.
- No investment advice.

Step 2: Configure "Arsenal"

Coze Bot "Skills" add plugins:

• Code Interpreter: core weapon, runs Python.
• Feishu Message + Feishu Bot: scheduled push and alerts.
• Image Understanding (optional): analyze data screenshots.

Coze code interpreter sandbox pre-installs pandas/numpy/matplotlib/seaborn/scikit-learn-no pip install.

Step 3: Build Core "Workflow"

Coze "Workflow" page visual drag-and-drop, zero code. Create "Report Generation" workflow:

Node 1: Start - Input var file (File type), upload CSV.

Node 2: Code Node (cleaning) - Python:

import pandas as pd
import numpy as np

def main(file_url: str) -> pd.DataFrame:
    df = pd.read_csv(file_url)
    df = df.dropna(subset=['GMV'])
    df = df[df['GMV'] > 0]
    df['日期'] = pd.to_datetime(df['日期'])
    return df

Node 3: Code Node (metrics) - Compute daily avg GMV/MoM/conversion/Top 3.

Node 4: LLM Node (attribution) - Feed metrics to LLM, generate "up 2.3% MoM because Qianchuan creative A CTR spiked" human-language conclusion.

Node 5: Code Node (visualization) - matplotlib line/bar charts, upload image links.

Node 6: Feishu Bot Node - Integrate Markdown report + images, scheduled push to Feishu group.

Step 4: Test & Tune

Upload a real CSV. Common issues: field name mismatch -> add field mapping in code node; chart Chinese garbled -> matplotlib add plt.rcParams['font.sans-serif'] = ['SimHei']; LLM attribution unreliable -> add "must cite specific values" to Prompt.

Step 5: Scheduled Trigger

Coze "Scheduled Trigger" plugin, set daily 9:00 auto-run, push report to Feishu group. Just check your phone in the morning-report's ready.

Pitfalls

1. Don't expect first-run success: Test with 10 rows of fake data first, then real. Large data may timeout code node (Coze limits 10 min/execution).
2. Define business metrics upfront: LLM doesn't know your "active user" means launch or login. Write business-metric docs in the knowledge base for Agent reference.
3. Data security: Coze data passes ByteDance servers. Sensitive data (user privacy/financial core) -> Dify private deployment or desensitize first.

Conclusion

AI data analysis isn't "use ChatGPT to write SQL" single-point efficiency-it's workflow automation. Coze lets you build a sleepless data-analysis digital worker with zero code. You shift from "human ETL + human BI" to "AI architect," focusing on business judgment and strategy.

The future belongs to those who can build tools with AI, not just use them.