Open Lab, Hands-On Lab: Build a Bookkeeping Pipeline

⚡ Try it yourself. Don't just read about the architecture, run it. This lab gives you 12 months of (fake) bank statements and the skills to take them all the way to a trial balance and a bank rec, the right way: code computes, AI drafts, you decide. Everything is synthetic and safe to practice on.

By the end you'll have run the full cycle on a real-feeling dataset and felt where the AI helps and where your judgment is non-negotiable.

▶ Want it to coach you live? The Interactive Guided Lab walks you through each step, checks your work, catches your mistakes in real time, and won't let you move on until it ties, the fastest way to feel the principles. Use the full kit below to go deep.

The demo business

Evergreen Landscaping LLC, a sole-owner S-corp in Georgia. Seasonal cash flow (slow winter, busy summer), a truck loan, payroll, owner distributions, and a few transactions that are deliberate judgment calls. All fake, fake name, fake account number, fake address, so you can practice freely.

What's in the kit

Everything lives in labs/bookkeeping-automation/:

• generate_data.py, reproducibly builds the 12 statements (run it, or tweak and regenerate).
• data/statements/2025-01.txt … 2025-12.txt, the monthly bank statements (they foot).
• data/chart-of-accounts.md, the COA, with the judgment calls spelled out.
• skills/00…05, the six step-by-step skills below.

To (re)generate the data: cd labs/bookkeeping-automation && python generate_data.py

The architecture you're practicing

This whole lab is one idea made concrete (from Skills & Architecture):

Code computes. AI drafts. You decide. Mechanics → deterministic. Legwork → the AI proposes. Judgment & sign-off → you. The AI never decides a position; it gets the desk ready for you.

Watch which layer each step lives in:

Do it yourself

New to this? Run just January first to learn the loop, then decide whether to do all 12.

1. Redact the statement (00), tokenize the client + customer names, but keep the merchant names (they drive the coding). Practice the habit even though the data is fake.

2. Extract (01), and foot it: Beginning + activity must equal the statement's Ending Balance. (January begins at $36,000.00.)

3. Categorize (02), the AI proposes accounts and flags the judgment calls (capitalize vs. expense, loan principal vs. interest, the IRS payment, ambiguous vendors). You decide each one.

4. Workpaper (03), group by account; it must tie to the bank.

5. Trial balance & P&L (04), debits = credits; recompute the math yourself.

6. Bank rec (05), reconcile to the penny with the two timing items provided.

The answer key (so you know you nailed it)

• December ending cash (bank): $36,799.25, your full-year workpaper must tie to this.

• Book cash after the rec's timing items: $37,949.25 ( + $2,300 deposit in transit − $1,150 outstanding check ).

• If a step doesn't foot or tie, the AI made an error you just caught, which is exactly the skill this lab is teaching.

What you just proved

• AI is genuinely fast at the legwork (reading statements, proposing codes), and wrong often enough that footing and tie-outs are non-negotiable.

• The judgment calls (capitalize, principal/interest, distribution vs. expense) are yours, the AI's job was to surface them, not settle them.

• A reliable pipeline is deterministic where it can be, AI where it helps, human where it counts.

Guardrails (this is a lab; real client work adds more)

• Real statements are full of PII, redact locally first, use a firm-approved tool, and keep it inside your WISP / §7216 posture (Regulatory Foundation).

• No silent posting, extracted numbers are drafts until footed and reviewed.

• You're the reviewer of record. The trial balance is yours when you sign it, not when the AI prints it.

Open Lab track · the applied companion to Skills & Architecture and Workpaper Automation.