I Automated My Client's 3-Hour Nightly Invoice Copy-Paste Ritual (CSV In, Reconciliation Out)

Every night at 10 PM, a distributor I know sits down with 40-80 PDF invoices and copies them into a spreadsheet. Vendor names, dates, amounts with thousand separators, the occasional duplicate from a re-sent email. Three hours. Every single day. He tried ChatGPT — it hallucinated line items. He tried paying a VA — turnover every two months.

The embarrassing part: none of this work was skilled. It was parsing, deduplication, and arithmetic. All things a deterministic script does perfectly, every time, with zero hallucination.

The manual invoice pipeline is a tax on small businesses

If you handle invoices for a small business, you already know the failure modes:

• Duplicates — a vendor re-sends an invoice, you enter it twice, and the P&L quietly lies to you
• Dirty numbers — 1,234.56 and 1234.56 and $1,234.56 all mean the same thing but Excel treats them as strangers
• Vendor aliases — "Acme Corp" vs "Acme Corporation" vs "ACME CORP" split your reporting into three phantom suppliers
• Reconciliation — matching what you entered against what your bank says is the part nobody has time for, so it never happens

None of this needs AI. It needs deterministic parsing — the same rules applied to every file, with the same result every run.

What I built: a no-AI invoice cleaning kit

I packaged a small Python kit that turns your nightly ritual into one command:

• invoice_cleaner.py — reads CSV exports of your invoices and:
  • dedupes exact and near-duplicate rows
  • normalizes dates (8/3/26 → 2026-08-03)
  • parses amounts with thousand separators and currency symbols ($1,234.56 → 1234.56)
  • normalizes vendor names so "Acme Corp" and "ACME CORPORATION" become one supplier
  • writes clean rows + an errors.csv for anything it couldn't parse (so nothing silently disappears)
• reconcile_report.py — produces a reconciliation workbook: totals by vendor, by month, by week, in one Excel file with three sheets
• sample_invoices.csv — 12 rows of realistic dirty data so you can see it work before touching your own books

The philosophy is the opposite of ChatGPT: no guesses, ever. If it can't parse a line, it goes to the error file instead of inventing a number. Deterministic, auditable, and 100x faster than copy-paste.

What it saved

On my test data: 12 rows → 10 clean, 2 duplicates removed, and the 1,234.56 thousand-separator rows parsed correctly on the first try. The report files generate in under a second.

For the distributor: his nightly 3-hour ritual becomes a 5-minute export + one command + a quick skim of errors.csv. That's ~15 hours a week back, doing work a human was never supposed to do.

The full kit (cleaner + reconciler + sample data + docs) is at AgentChip.

The rule that keeps it honest

Never let a tool guess a financial number. If your pipeline can't parse a line, it should surface it for a human — not round it, not skip it, not "AI-complete" it. That single rule is the difference between software that helps your bookkeeping and software that quietly corrupts it.

Automate the copy-paste. Keep the human judgment. Your evenings are worth more than data entry.

---

Originally published on the AgentChip blog.