From manual operations to autopilot · Family service station

The challenge

A family-owned SMB with a service station and bar spent hours every week on tasks that repeated without end: downloading invoices from email one by one and renaming them, reformatting POS exports so the accountant could understand them, updating the monthly cash journal and managing recurring office tasks.

None of these tasks were complex. But doing them correctly, every day, without anything being forgotten or copied wrong, consumed constant time and attention. The problem wasn't complexity — it was accumulation.

The solution

We developed four automations covering the full operational cycle of the business:

1. Invoice download and renaming: A script connected to the Gmail API automatically downloads incoming invoice PDFs and renames them with a consistent pattern (date, expense type, supplier). When a PDF's format is ambiguous, the Claude API interprets the content to classify it correctly.

2. POS Excel formatting: Daily point-of-sale exports are automatically converted into a structured Excel workbook with one tab per day, clear sections and correct number formatting. The script is cumulative: reprocessing a day only overwrites that tab.

3. Monthly journal update: The formatted POS data is automatically written into the accountant's monthly journal, filling the 23 corresponding accounting fields in the correct columns. If the month's journal doesn't exist yet, it's created from a template.

4. Recurring task manager: A desktop application shows each morning's tasks (daily, weekly and monthly), automatically carries over pending items and requires no internet connection or external accounts.

Code excerpt

# Invoice download and renaming — Gmail API
import re
from googleapiclient.discovery import build
from anthropic import Anthropic

client = Anthropic()

def classify_invoice(pdf_text: str, supplier_exceptions: dict) -> str:
    """Classify invoice as BAR or EESS using rules + AI fallback."""
    text_lower = pdf_text.lower()

    # Rule-based classification first
    for supplier, tag in supplier_exceptions.items():
        if supplier.lower() in text_lower:
            return tag

    bar_keywords = ['hostelería', 'alimentación', 'bebidas', 'makro']
    eess_keywords = ['combustible', 'carburante', 'gasoil', 'repsol', 'cepsa']

    if any(k in text_lower for k in bar_keywords):
        return 'BAR'
    if any(k in text_lower for k in eess_keywords):
        return 'EESS'

    # AI fallback for ambiguous cases
    response = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=10,
        messages=[{
            "role": "user",
            "content": f"Clasifica esta factura como BAR o EESS. Solo responde BAR o EESS.\n\n{pdf_text[:500]}"
        }]
    )
    result = response.content[0].text.strip().upper()
    return result if result in ('BAR', 'EESS') else 'EESS'

def build_filename(date: str, tag: str, supplier: str, total: float) -> str:
    mmdd = date.replace('/', '').replace('-', '')[:4]
    supplier_clean = re.sub(r'[^\w]', '', supplier)[:20]
    return f"{mmdd}_{tag}_{supplier_clean}_{total:.2f}.pdf"

Results

The impact was immediate and cumulative. Each automation on its own saved minutes; together they free up around 10 hours of manual work per week.

Measured impact: - ~10 hours/week saved on operational tasks - 0 manual errors in the accounting process since implementation - Invoices reach the accountant already organized and renamed - The monthly journal updates itself, without human intervention - Recurring office tasks run themselves, without depending on anyone's memory