FEATURED Internal AI Operations System

AI-Assisted Payroll Operations System

Payroll arrived as spreadsheets from different people in different shapes, and an incorrect entry was only discovered after payment. The fix isn't a single script — it's a financial dashboard, an AI reconciliation loop that's still being trained, and a hard rule that the system never guesses with someone's paycheck.

ROLE: Requirements, validation rules, workflows, QA
TOOLS: FastAPI · SQLite · spreadsheet processing

The constraint

There was no clean API to build against.

Most systems like this start with an API to pull data from. The third-party payroll platform running payroll for around 46 employees here didn't offer one, so the first job was building the bridge itself — moving data out by hand and connecting it into a backend built to receive it, rather than assuming a clean integration would already be there. The dashboard that resulted behaves the same way a live, API-backed system would from the outside; getting there meant building the plumbing an API would normally provide.

The dashboard

One place to see the whole financial picture.

Once the data had a home, the dashboard became the single point of visibility into the financial system — payroll, what's being withheld and paid in employee taxes, and direct deposits, with an expense tracker sitting alongside it. It works closer to something like QuickBooks than a single-purpose payroll tool: bank statements can be connected, more than one business can run through it at once, and every expense is tracked against the correct business rather than pooled together.

Payroll & taxes

What's being paid out and withheld, visible in one view instead of reconstructed after the fact.

Direct deposits

Tracked alongside payroll rather than in a separate system someone has to cross-check manually.

Expense tracker

Bank statements connect in, and every expense is attributed to the correct business.

Multi-business

Built to run more than one business through it side by side, not just a single entity.

Teaching the AI to reconcile

The manual side is by design, for now.

Reconciling expenses still has a real manual step: a person verifies each match and corrects it when it's wrong, and that correction is what the AI is learning from. The plan is for that supervised phase to shrink as the AI earns confidence — proven by getting it consistently right, not assumed because the automation exists.

The same caution carried into a later, separate experiment: light browser automation around that third-party payroll platform itself, aimed at cutting down the repetitive clicks involved in manual entry and processing. It was explored deliberately rather than shipped as full autonomous automation — the same earn-it-first approach applied to reconciliation applies here too.

The part that matters most

When the AI isn't sure, it doesn't guess.

This is payroll and money — the cost of a confident wrong answer is too high to treat like any other automation. Every uncertain case is escalated to a person instead of resolved automatically, with clear boundaries on what the system can decide on its own versus what always needs a recorded human checkpoint.

No hallucinated answers

Uncertainty routes to a person instead of the system inventing a plausible-looking number.

Escalation over autonomy

Autonomy is earned per task by a track record, not granted by default because the automation is live.

Security first

The same discipline applied to reconciliation logic applies to how the system is secured and reviewed.

The pipeline

Import, validation, exception queue, approval.

Editorial representation of the architecture, not a client interface. Click a step to see how it works.

Have a financial or payroll workflow that needs this level of discipline?

Tell us what's manual today and where the risk actually is, and we'll map what a practical, guardrailed implementation could look like.

Plan an AI project See other systems we've shipped