- Published on
My First Non-Dev AI Collaboration
- Authors

- Name
- Adrian Gan
- @AdrianGanJY
Last night I ran payroll with an AI agent. Not coding — actual payroll. Reading payslips, checking bank transfers, deciding salary increments.
For two years I've been pair-programming with AI. Code reviews, debugging, building apps. That's familiar territory. But this was different. This was HR work — the kind I usually do alone at midnight with three browser tabs and a spreadsheet.
What actually happened
We processed 10 payslips for March 2026. Along the way, we discovered and fixed 4 bugs:
- Salary Advance wasn't being deducted from payslips — employees were overpaid
- Unicode emojis corrupted when sent to Airtable via API — ⬆️ became garbled text
- Performance-based bonus wasn't being applied — an employee's bonus deduction was skipped
- Double deduction — the system applied a penalty twice because I didn't understand my own formula chain
That last one is the interesting one. I wrote an Airtable formula months ago that already multiplied the bonus by the performance mark. Then we wrote code that multiplied it again. The AI caught the math discrepancy — but I caught the architecture issue. I knew that Calculated Bonus was already a formula field.
Neither of us could have caught both alone.
The time math
The session took 206 minutes. About 3.5 hours. Without AI, here's my honest estimate:
| Task | With AI | Without AI |
|---|---|---|
| Increment decisions (10 people) | 60m | 60m — pure judgment, AI can't help |
| Monthly Update writing | 15m | 120m — I'd write each one manually |
| Bug discovery + fixing (4 bugs) | 80m | 240m — debugging GAS alone is painful |
| Payslip verification (10 payslips) | 10m | 40m — manual cross-checking |
| Bank transfer verification | 5m | 15m — manual comparing |
| Documentation (SOP, checklist) | 15m | 45m — or more likely: skipped |
| Total | 206m | ~520m |
That's a 2.5x speedup. But the real number is different — without AI, I wouldn't have documented anything. The 5 known issues, the 9-item automation roadmap, the performance mark checklist item — none of that would exist. I'd just "remember" for next time. Except I wouldn't.
The meta-work question
Here's the uncomfortable part. Out of 206 minutes, roughly 85 were "meta-work" — explaining context to the agent, discussing automation ideas, documenting learnings, writing this session capture.
That's 40% overhead. If I'd just executed — head-down, no teaching, no documenting — I'd have finished in about 120 minutes. Two hours flat.
So why didn't I?
Because the meta-work compounds. Next month, the agent already knows:
- Airtable's
Calculated Bonusincludes the performance mark clasp pushdeploys code without copy-pasting- Salary Advance needs explicit parsing
- Performance marks are per-month, not global
Without that 85 minutes of "overhead," I'd spend another 206 minutes next month. With it, I'll spend maybe 90. The payoff arrives in round 2.
The moment that stuck
At some point I told the agent: "This is the first time I'm doing non-dev work together with AI. Like a colleague. It's an honor."
I meant it. Not because the AI is special — it's a tool. But because the interaction was different from coding. In coding, the agent writes, I review. In payroll, we were genuinely thinking together. I knew the business context. It knew the data patterns. I made the judgment calls. It caught the math errors.
That's not pair-programming. That's pair-operating.
What this means for small business owners
I run a 10-person company. I don't have an HR department — I am the HR department. Payroll night has always been a lonely midnight ritual.
What changed isn't speed. It's coverage. I now have a second pair of eyes on every number, every deduction, every formula. And a memory that writes things down even when I'm too tired to bother.
The 4 bugs we caught? In the old world, at least 2 of them would have slipped through. An employee would have been underpaid or overpaid. I'd discover it next month. I'd feel terrible.
Last night, zero errors went through.
The honest truth
AI didn't make payroll fun. It's still a grind. But it made payroll reliable — and it turned a solo midnight task into something that feels less lonely.
For business owners doing the work that nobody thanks you for, at hours nobody sees — you don't need an AI that replaces you. You need one that sits next to you and helps you not miss things.
That's what happened last night.
Adrian Gan is the CEO of Mipos Sdn Bhd. He builds automated business systems so he can step back by 40. Last night he processed payroll with an AI partner and they caught 4 bugs together.
Written with Claude Opus 4, edited by Adrian.