I Used an AI Agent to Manage My Family Budget for Two Months — Saved Enough for an iPad

Summary: I connected an AI agent to my bank accounts, credit cards, and Alipay/WeChat Pay for 60 days. The agent categorized every transaction, flagged unusual spending, suggested optimizations, and tracked progress against a monthly budget. Result: my family spent ¥2,400 less per month (¥4,800 total over two months) — enough to buy a new iPad. This is a practical, data-rich account of what AI agent-driven budgeting actually looks like, what tools to use, and how to set it up yourself. Spoiler: you don't need a finance degree, you need the right agent workflow. And yes, KaiheAiBox A1 running 24/7 is the ideal host for a budgeting agent.

The Setup (Before You Ask "Is This Safe?")

Let me address the security question first, because it's the only reason most people don't try this.

I did not give an AI agent access to my bank login credentials. Here's what I actually did:

Exported transaction data from my bank's mobile app (most Chinese banks let you export 3-6 months of transaction history as CSV)
Uploaded the CSV to a local-only agent running on my KaiheAiBox A1
Gave the agent read-only instructions — it could analyze, categorize, and flag, but could not move money or make payments
Set up daily email forwarding — my bank sends transaction alerts via email; the agent reads those emails and updates the ledger

The agent never had my passwords. It never initiated a payment. It was an analyst, not an operator.

If you're building a budgeting agent, read-only access is the right starting point. You can always add write capabilities later, once you trust the agent's judgment. Start with insights, not actions.

文章配图

Month 1: The Wake-Up Call

I thought I knew where my money was going. I was wrong.

The agent categorized 247 transactions in Month 1. Here's what it found:

Category	Amount (¥)	% of Total	My Guess (Before Agent)
Dining out	4,820	31%	~¥3,000
Subscriptions (forgot to cancel)	387	2.5%	¥0 (thought I had none)
Online shopping (impulse)	3,150	20%	~¥2,000
Transportation	1,240	8%	~¥1,000
Utilities + rent	3,200	21%	~¥3,200 (correct)
Entertainment	1,450	9%	~¥1,000
TOTAL	¥15,247	100%	~¥12,200 (I was off by ¥3,000)

Three things jumped out immediately:

1. Dining out was 31% of spending. My wife and I were ordering delivery or eating out 4-5 times per week. We told ourselves "we're too busy to cook." The agent's comment: "If you cooked 3 more dinners per week, you'd save ~¥2,000/month."

2. Forgotten subscriptions. ¥387/month in subscriptions I had stopped using but never cancelled: a video platform premium tier, a cloud storage plan I'd replaced with a cheaper alternative, and a fitness app I opened exactly once in the past 6 months.

3. Impulse shopping was 20%. The agent flagged 23 "impulse pattern" transactions — defined as: purchased after 10pm, no price comparison found in browser history, not on the pre-approved shopping list. Total: ¥3,150.

Month 1 was painful to read. But it was also the first time I had accurate, categorized, complete visibility into our spending.

Month 2: The Course Correction

Based on Month 1's analysis, I gave the agent three new instructions:

Instruction 1: "Guard the Groceries"

"Before any food delivery order over ¥80, send me a WeChat message asking: 'You already have groceries at home. Cook?' If I don't reply within 10 minutes, cancel the order."

This one intervention reduced dining-out spending from ¥4,820 (Month 1) to ¥2,100 (Month 2). Not because the agent "blocked" me — I overrode it twice — but because the friction made me conscious of the choice.

The most effective AI agent interventions are not restrictions. They're friction. Make the undesirable behavior slightly harder, and the desirable behavior becomes the path of least resistance.

Instruction 2: "Kill the Forgotten Subscriptions"

"On the 25th of each month, list all subscriptions renewed in the past 30 days. I will confirm which to cancel."

Month 2 saved ¥387 by cancelling 3 subscriptions. Small? Yes. But ¥387 × 12 months = ¥4,644/year. That's a meaningful chunk of an iPad.

Instruction 3: "The 24-Hour Rule for Impulse Buys"

"For any online purchase over ¥200 that isn't on the pre-approved list, add it to a 'waiting list.' Send me a reminder in 24 hours. If I still want it, I'll confirm. If not, it gets removed."

This single rule eliminated ¥1,890 in impulse purchases in Month 2. The agent didn't need to "persuade" me not to buy things. It just added a 24-hour cooling-off period. Most of the time, I didn't even remember what the item was by the time the reminder arrived.

The Numbers: ¥4,800 Saved in 60 Days

Metric	Month 1	Month 2	Change
Total spending	¥15,247	¥11,830	-¥3,417 (-22%)
Dining out	¥4,820	¥2,100	-¥2,720 (-56%)
Subscriptions cancelled	¥0	¥387	+¥387 saved
Impulse purchases	¥3,150	¥1,260	-¥1,890 (-60%)
Two-month total saved	—	—	¥4,800

¥4,800 is enough to buy: - A new iPad (10th gen, 256GB): ~¥3,999 - Plus a nice case: ~¥400 - Plus Apple Pencil (USB-C): ~¥799 - Total: ¥5,198 — we're only ¥398 short after 2 months

Month 3 is on track to push us over the line.

The Toolchain: What I Actually Used

Here's the exact setup, in case you want to replicate it:

Tool 1: KaiheAiBox A1 (The Host)

The agent needs to run 24/7 to: - Check email for transaction alerts (every 30 minutes) - Update the spending ledger (daily at 11pm) - Send WeChat reminders (in real-time) - Generate the monthly report (1st of each month)

A laptop could do this, but then I'd need to keep my laptop on all the time. KaiheAiBox A1 draws 15W — about ¥10/month in electricity. It runs OpenClaw, which handles the email monitoring, WeChat integration, and scheduled tasks.

Why not just use a phone app? Two reasons: (1) Phone apps can't run continuously in the background without being killed by the OS. (2) I wanted the agent to have access to my full financial history (6 months of CSV data), which is easier to manage on a small server than a phone.

Tool 2: OpenClaw (The Agent Framework)

OpenClaw handles the orchestration: - Email monitoring (IMAP, checking every 30 minutes) - WeChat message sending (via WeCom API) - Scheduled tasks (cron-style) - Data persistence (SQLite ledger)

The agent logic itself is a Python script that OpenClaw calls when new transaction emails arrive.

Tool 3: The Analysis Script (Python)

The actual "AI" part is a Python script that: 1. Parses transaction CSVs (bank format) 2. Uses a local LLM (running on A1) to categorize each transaction 3. Compares against budget targets 4. Flags anomalies (unusual merchant, unusual amount) 5. Generates the daily/weekly/monthly report

I used Claude Code to write the first version of this script in about 40 minutes. The prompting session went like this:

"Write a Python script that reads a bank transaction CSV, uses an LLM to categorize each row into {dining, shopping, transport, utilities, entertainment, subscriptions}, detects anomalies (amount > 3× the 30-day rolling average for that merchant), and outputs a daily spending report as Markdown."

Claude Code wrote the script. I tweaked three things (merchant name normalization, the anomaly threshold, and the output format). Total time: 40 minutes to working; 2 hours to polished.

Tool 4: WeChat (The Interface)

The agent sends alerts and reports via WeChat (WeCom API, personal account). This was the right choice because: - WeChat notifications are hard to miss - I can reply to the agent's messages to confirm actions ("Yes, cancel that subscription") - It works on both my phone and my wife's phone (shared family budget)

Three Lessons Learned

Lesson 1: Categorization Accuracy Improves Over Time

In Week 1, the agent miscategorized about 15% of transactions. By Week 4, that was down to <5%. The improvement came from two sources:

Feedback loop. I corrected miscategorized transactions. The agent (well, the LLM) learned my patterns. "Oh, this merchant is actually 'dining,' not 'shopping'."
Merchant name normalization. Banks use inconsistent merchant names: "Starbucks #0028 SH" vs "Starbucks (Shanghai Huaihai)." The agent learned to normalize these into a canonical name.

If you're building this: don't expect perfection on Day 1. The agent gets better with use.

Lesson 2: The Agent Needs a "Cooling-Off Period" Too

In Month 2, the agent sent me 23 "are you sure?" reminders for purchases. I overrode 21 of them. The two I accepted saved me ¥680.

But here's the thing: the overrides were training data. By Month 2, the agent had learned which types of "are you sure?" reminders I usually override vs. which ones I usually accept. It started prioritizing the high-savings reminders and de-prioritizing the low-savings ones.

Your agent is not static. It learns your patterns — both the spending patterns and the "do I listen to the agent?" patterns. This is why a 60-day run produces better results than a 7-day run.

Lesson 3: The Biggest Savings Come from Behavior, Not Optimization

I thought the agent would find clever ways to optimize my existing spending — better credit card cashback, cheaper alternatives for regular purchases, that kind of thing.

It didn't. Or rather, that accounted for maybe ¥200/month in savings.

The real savings came from behavioral change: cooking more often, cancelling forgotten subscriptions, and adding friction to impulse buys. The agent didn't need to be clever. It needed to make invisible patterns visible.

The best financial AI agent isn't an optimizer. It's a mirror. It shows you what you're actually doing with your money. The behavior change is up to you — but you can't change what you can't see.

Can You Do This Yourself? (Yes, Here's How)

You don't need to be a programmer. Here's a 3-step path to building your own budgeting agent:

Step 1: Get Your Data (30 minutes)

Log into your bank's mobile app
Export the last 3 months of transactions as CSV
Do the same for Alipay and WeChat Pay (both have "export statements" features)
Combine them into one CSV

Step 2: Set Up the Agent (1-2 hours)

Option A (Easy): Use a no-code AI agent platform (Zapier with AI, Make.com, or n8n with an LLM node). Connect your email, set up categorization, and have it send you a weekly report.

Option B (Better): Run the agent on a KaiheAiBox A1 with OpenClaw. This gives you 24/7 operation, private data, and full customization. The learning curve is steeper, but the result is a system that runs indefinitely without maintenance.

Step 3: Give It Guardrails (30 minutes)

Start with read-only. Let the agent analyze and report. After 2-4 weeks, if you trust the analysis, you can give it limited write actions (cancel subscriptions, set spending alerts).

The most important guardrail: the agent should never be able to initiate a payment. Analysis = yes. Money movement = no. Not because the agent is untrustworthy, but because you want a human in the loop for anything involving transferring money.

Why KaiheAiBox A1 Is the Right Host

I tried running the budgeting agent on three different platforms before settling on KaiheAiBox A1:

Platform	Uptime	Power Cost	Privacy	Setup Difficulty
Laptop (always on)	Unreliable (updates, crashes)	~¥80/month	Medium	Medium
Raspberry Pi 4	Good	~¥5/month	High	High
Cloud VPS	Perfect	$10-20/month	Low (data in cloud)	Low
KaiheAiBox A1	Perfect	~¥10/month	High (local)	Low (pre-configured)

The A1 hits the sweet spot: always-on, low power, private data, and pre-configured OpenClaw. The cloud VPS was the runner-up, but I didn't want my financial data leaving my home network.

A budgeting agent is a 24/7 workload that benefits from private data handling. That's exactly the use case KaiheAiBox A1 was built for.

Your Turn

The math is simple. If an AI agent can help you find 10-20% in savings on a ¥10,000/month spending budget, that's ¥1,000-2,000 per month. After a year, that's ¥12,000-24,000 — enough for a serious upgrade to your tech setup, a vacation, or a meaningful contribution to your savings.

The barrier isn't technical. It's behavioral. Most people don't want to see where their money is going, because they suspect the answer will make them uncomfortable.

An AI agent won't judge you. It'll just show you the numbers. And sometimes, that's exactly what you need.

KaiheAiBox · AI Agents