Results
Payroll workflow:
from 5 hours to 30 minutes
A real story of how we automated payroll processing by combining human oversight + AI + automation. With a complete audit trail.
Before and after
Before (manual)
- ✗ Manual attendance data collection
- ✗ Copy-pasting data into email
- ✗ Waiting for accountant's response
- ✗ Manual payslip download from email
- ✗ Manual transcription of amounts to spreadsheet
- ✗ Manual distribution of payslips to employees
- ✗ No audit trail
~5 hours / month
After (with the operating system)
- ✓ Data prepared in Google Sheets
- ✓ AI composes and sends email to accountant
- ✓ Automatic payslip download (24/7)
- ✓ AI extracts data from PDFs into Sheets
- ✓ AI distributes payslips to employees
- ✓ Automatic preparation for the next month
- ✓ Complete change history
2 x 15 minutes / month
7 steps of the payroll workflow
AI-assisted + fully automated + 1 human step
Data loading
A Python agent loads attendance data, bonuses and deductions from provided inputs. The human supplies source information, the agent processes the rest.
Send to accountant
The AI agent autonomously composes an email with attachments and sends it to the external accountant automatically via email. No human confirmation needed.
Receive payslips
Google Apps Script detects the accountant's email → downloads ZIP → unzips → saves PDFs to Google Drive. Runs 24/7 without human intervention. A human verifies the output.
Write to spreadsheet
The AI agent reads PDF payslips, extracts "personnel cost" amounts and writes them to the correct column in Google Sheets. A human verifies the output.
Bank authorization
The only human step. Authorization of automatically prepared payment orders in the bank.
Distribute payslips to employees
The AI agent autonomously sends each employee their PDF payslip automatically via email. A human performs a final changelog review.
Prepare next month
Copy sheets in Google Sheets, reset temporary fields. Everything ready for the next cycle.
Complete audit trail
For every month there is a changelog: a record of every step, date and output. Every change is recorded and versioned. When an auditor comes knocking, you show the complete change history.
# processes/finance/payroll/changelog/2026-02.md
---
title: "Payroll - February 2026"
status: approved
---
# Payroll, February 2026
## Step 1: Data preparation
- Date: 2026-02-28
- Output: Data in Google Sheets
## Step 3: Receive payslips
- Date: 2026-03-01
- Output: Automatically saved to Drive
## Step 5: Distribution
- Date: 2026-03-02
- Output: 4 emails sent
Results in numbers
2 x 15 minutes
6 of 7 steps
bank authorization only
since deployment
What we learned
Start small
Skip the "big bang." Start with one process. When it works, add the next. We started with payroll.
Structured metadata is critical
Without metadata it is just a pile of Markdown files. With metadata it becomes a database AI can read.
Built-in approval for every change
Every process change goes through an approval workflow. The complete history is preserved forever.
AI needs structure
ChatGPT without context = generic answers. AI with structured company knowledge = relevant, precise answers.
People must buy in
The biggest obstacle is cultural, not technical. The answer: traditional documents cannot be read by AI or properly versioned.
Hybrid works best
Do not automate everything. Humans prepare data, AI handles the routine, automation runs 24/7. The optimal mix.
Want similar results?
In our workshop we demo the full workflow live and design a solution tailored to your company.
Book a workshop