n8n + Ollama 帳單自動化：把 12 份信用卡帳單餵進 AI，三分鐘出一份全年分析報告n8n + Ollama Billing Automation: Feed 12 Credit Card Statements Into AI, Get a Full-Year Report in Three Minutes

I had 12 months of credit card statements as plain-text files, 30–40 entries each. Manual sorting was possible, but I wasn't going to do the same thing twice.

12 months × ~35 entries each = ~420 rows to categorize. That's what "manually doable" actually looks like — one cell at a time, think about which category this merchant fits, type, Enter, next cell. And I definitely wasn't uploading those files to a cloud AI — merchant names, amounts, timestamps are enough to reconstruct a family's entire spending profile.

So I spent an afternoon wiring together n8n + Python + Ollama into a fully automated pipeline. One webhook trigger, three minutes later: a categorized Excel file with charts, plus an AI-written monthly summary for each of the twelve months.

This is how the pipeline works, where it broke, and what the numbers said.

Architecture: three layers, one job each

n8n Webhook
    ↓
Python Flask endpoint (parse → classify → Ollama → Excel)
    ↓
JSON back to n8n

n8n is the conductor: receives the trigger, calls the Python endpoint, formats the response. Three nodes total. Five minutes to build.

Python Flask does the actual work: parses the statement files, runs keyword classification, calls Ollama for each month's summary, generates the Excel with openpyxl. All the real logic lives here.

Ollama translates numbers into sentences: each month's spending data goes into gemma4:e2b with a prompt asking for an 80-character summary — the month's notable characteristic, one thing worth watching, one recommendation.

The n8n side: three nodes

Webhook trigger (GET /billing-annual)
    → HTTP Request (POST → Flask :5001/analyze)
    → Set node (clean up response fields)

Webhook node: receives the GET request. responseMode: lastNode holds the response open until the last node finishes — important when the pipeline takes minutes.

HTTP Request node: the critical one. Settings:

• Method: POST
• URL: http://host.docker.internal:5001/analyze (container → host machine)
• Body: {"year": "2026", "skip_ai": false}
• Timeout: 30 minutes (the default 5 minutes will time out before Ollama finishes)

Set node: extracts status, elapsed_s, summary.total_transactions from the Flask JSON response and reshapes it into something cleaner.

That's it. n8n's role here is thin by design — a trigger entry point and an HTTP proxy. The complexity lives in Python.

The Python side: three scripts

All the parsing, classification, and Excel generation happens in Python — split across three scripts.

1. billing_parser.py: parsing statements

HSBC Traveller card statements arrive as plain text in this format:

匯豐旅人御璽卡
89,845 TWD
2025/12/03
連加＊全聯福利中心台北民生東路
3,800 TWD
2025/12/05
ＡＰＥ＊ＰＡＲＫＩＮＧ Ｔａｉｐｅｉ ＴＷＮ
7,200 TWD
...

Two gotchas:

Gotcha 1: fullwidth Unicode. Merchant names come in fullwidth characters (ＡＢＣ instead of ABC). You need to normalize these before any keyword matching. The conversion shifts codepoints in the 0xFF01–0xFF5E range back by 0xFEE0:

def fullwidth_to_ascii(text: str) -> str:
    result = []
    for ch in text:
        code = ord(ch)
        if 0xFF01 <= code <= 0xFF5E:
            result.append(chr(code - 0xFEE0))
        elif ch == '　':   # fullwidth space
            result.append(' ')
        else:
            result.append(ch)
    return ''.join(result)

Gotcha 2: merchant name suffixes. Parking garages show up as PARKING Taipei TWN. Strip the country/city suffix naively and you're left with an empty string. Fix: only apply the strip if something remains afterward:

stripped = re.sub(r'[A-Z]{2,}\s*$', '', name).strip()
if stripped:
    name = stripped

2. Classification: 15 categories

Keyword matching against 15 categories, ordered by priority — first match wins:

CATEGORY_RULES = [
    ('手續費',   '其他',    ['手續費', '國外交易']),
    ('訂閱服務', '個人數位', ['NETFLIX', 'DISNEY', 'KKBOX', 'SPOTIFY', 'GOOGLE', ...]),
    ('育兒教育', '育兒',    ['安親班', '才藝班', '游泳', 'LEGO', '夏令營', '學校', ...]),
    ('旅遊出行', '旅遊',    ['長榮航空', '華航', 'Agoda', 'Booking', '旅館', '飯店', ...]),
    ('交通停車', '日常交通', ['中油', '加油', 'PARKING', '停車', '捷運', '高鐵', ...]),
    # ... 15 categories total
]

Merchant name is uppercased, then matched sequentially. After a few rounds of tuning, misclassification rate across 405 transactions is under 5.

3. generate_excel.py: four worksheets

Charts are rendered by matplotlib as PNG, then embedded into the worksheet via openpyxl:

fig, ax = plt.subplots(figsize=(10, 5))
ax.plot(months, totals, marker='o', color='#3498DB', linewidth=2.5)
# ...save as PNG, then wb.add_image(img, 'A16')

The big gotcha: n8n Write Binary File is blocked

The original design had n8n writing the Excel file directly to a local path — one less HTTP endpoint. Turns out n8n v2 in Docker blocks all local path writes from the Write Binary File node. /data, /tmp, /home/node/.n8n/ — all return "Not writable."

This is intentional security behavior in n8n v2, not a bug.

The fix: move all file generation into the Flask endpoint. n8n never touches binary data — it only fires the trigger and reads back JSON. Annoying to discover mid-build, but the resulting architecture is actually cleaner.

Second gotcha: the n8n Code node sandbox doesn't allow fetch or $helpers. No direct Ollama calls from inside a Code node. That's why the Ollama call also lives in the Flask endpoint, using Python's requests.

Results: NT$847,538 for the year

405 transactions. NT$847,538 total. NT$70,628 monthly average.

Category breakdown: childcare takes a third

Childcare alone accounts for 34.1%. Breaking it down: after-school programs for three kids — NT$165,000. Swimming and music lessons — NT$53,500. Those three line items total NT$218,500, nearly a quarter of the full year's spend.

Transport and parking comes second at NT$133,020. Parking garages alone: NT$86,400 (NT$7,200 per month on average). Add two gas stations at NT$36,890. Driving in Taipei costs more than it feels like it does until you see it itemized.

Supermarkets third at NT$128,866. Costco NT$49,330 + PXMart NT$38,000 + Carrefour NT$24,446. The Costco membership pays for itself.

Monthly peaks: each one has a reason

August's spike isn't runaway daily spending — it's one family trip pushing the whole month up. Strip out the travel category and August looks normal.

September (NT$82,555) is high because of back-to-school: new backpack, desk, stationery NT$8,500 + school registration fees NT$4,500 + summer camp balance payments. All concentrated in late August / early September.

Subscriptions: NT$9,468 for the year

Four subscriptions, NT$9,468 total. Lower than expected — several are split through Apple Family Sharing.

Ollama analysis quality

Each month's summary comes out at 80–120 characters in Traditional Chinese. The prompt feeds in both the category totals and the top 12 transactions by amount:

請務必使用繁體中文台灣用語回答。

以下是台灣家庭 2026-08 信用卡消費（共 35 筆，合計 NT$109,248）。

類別合計：
  旅遊出行 NT$54,950、育兒教育 NT$13,500、超市量販 NT$10,960、交通停車 NT$9,160

最高消費前 12 筆：
  2026/07/04 長榮航空...   NT$ 32,500 [旅遊出行]
  2026/07/06 暑期夏令營第一週費用   NT$  9,000 [育兒教育]
  ...

請以 80 字以內分析：1) 本月消費特色 2) 值得注意的支出 3) 一句建議。
直接分析，不要分點，不要重複數字。

August's output (translated):

This month's spending is dominated by travel and childcare education. The single airline ticket stands out as a high-value item — future trip planning would benefit from more precise budgeting to avoid unexpected travel costs.

Reasonable quality. It correctly identifies the month's dominant pattern and the recommendation is on point.

One consistent weakness: across all 12 months, the advice converges on some variation of "budget more carefully." The wording shifts but the core is the same. This makes sense — spending structure is stable, fixed costs dominate, and there's only so much the model can say about a pattern that doesn't change.

The fix would be adding "compared to last month" context to the prompt — force the model to describe the delta, not just the state. I left this for a future iteration.

Performance

Running on MacBook M1 16GB:

gemma4:e2b on M1 16GB peaks at around 7GB RAM usage. Fourteen seconds per monthly summary, twelve months in a row, under three minutes total. Faster than I expected from a CPU-only inference run.

No GPU. Pure CPU. Still fast enough.

Should you build this

If you have one statement to sort through, just open a spreadsheet.

This pipeline makes sense when:

• Your statement format is consistent and you'll process it every month
• You don't want financial data hitting any cloud service
• You want AI summaries but not the API bill
• You already use n8n or have been looking for an excuse to try it

What you need: n8n (Docker is fine), Ollama (pull gemma4:e2b), Python 3.12+ (openpyxl + matplotlib).

Total setup time across all three tools: under 30 minutes. After that, one webhook trigger, three minutes, done.

Scripts are in tests/scripts/. Statement format reference in tests/data/.