Laptop at dusk showing a payroll dashboard, a crumpled printed spreadsheet and an old calculator pushed aside

I'm not a developer. Our payroll system went live two weeks after I approved the spec

How I replaced a 3-4 hour monthly payroll run with a 15-20 minute one by directing AI coding agents - live two weeks after spec approval, used daily by 27 people.

Business 5 min
Mark Podliashetskyi
Mark Podliashetskyi CEO
Share:

Our payroll used to take me three to four hours every month. Now it takes 15 to 20 minutes. I am not a developer: I wrote the requirements and directed AI coding agents, with a second AI model reviewing their work. The system went live about two weeks after I approved the spec, and all 27 people in the company use it.

The result

27

people using it daily

3-4 h → 15-20 min

my monthly payroll run

~2 weeks

from approved spec to production

What was the problem?

Every month I ran payroll by hand from a spreadsheet. Hours, projects, people joining or leaving in the middle of the month, commissions - it all came together in my head and in formulas, and it took three to four hours each time. It was exactly the kind of work that is too specific for an off-the-shelf tool and too important to get wrong.

What did I actually do as CEO?

I did the part only I could do: I decided how our pay works and wrote it down as a requirements document. The hard cases were business decisions, not code:

  • Part-month pay: people who join or leave mid-month are paid pro rata.
  • Commission: paid from cash actually collected, never negative, and applied retroactively when payment arrives.
  • Time tracking: developers log their own hours and tasks per project, and managers check them.

Then I directed AI coding agents to build it. I did not write the code; I approved plans, answered questions and checked results.

Why can I trust code I did not write?

Because I did not ask one AI to check its own work. Before each feature, a model from a second vendor reviewed the coding agent's plan; after it, the same reviewer checked the code. Every finding was either accepted or rejected with a recorded reason, so I could see why something changed - or why it did not.

The other part of trust was the data. Payroll means real salaries. Early on, the agent created test data from the real spreadsheet and committed it to the repository. We cleaned it out and set a hard rule: real compensation data stays out of the repository and out of the AI tools, and tests run only on invented data. I wrote more about that in The guardrails we put on AI coding agents, and the incident that taught us.

I did not hand the work to AI. I handed it the parts that were never the real work.

- Mark Podliashetskyi, CEO

What changed after launch?

The system has been live since mid-September 2026 and is used every day. Developers record their hours and tasks per project almost daily, managers review them, and at the end of the month I run payroll in 15 to 20 minutes instead of three to four hours. Smaller fixes and additions continue as we use it.

The next step is client invoices: the same hours that drive payroll will generate our monthly invoices in the format each client needs.

What should another business owner take from this?

  • Your job is the rules, not the code. The most valuable document was the one describing how we pay people.
  • Pick a process you know by heart. You can only judge an agent's output in a domain you understand.
  • Never let one model review itself - use a reviewer from another vendor and record its findings.
  • Keep real personal data away from AI tools from day one.
  • Start with something internal. Your own team is the most forgiving and most honest first user.

This is how we work with clients too: the business owns the rules, AI agents do the building, and independent review plus human approval make the result something you can trust.

Want to Talk to Our Team?

Our engineers can help map this to your stack.

An unhandled error has occurred. Reload 🗙