DEV Community

Maryan K
Maryan K

Posted on

I Built an AI Employee That Files Invoices and Refuses to Guess

I Built an AI Employee That Files Invoices and Refuses to Guess

The "AI agent" promise got way ahead of the product. Nika is the opposite: a deliberately narrow AI employee that does one job, files invoices, and is designed to stop and ask rather than guess.

Why an "employee" and not an app

The founder who still does their own bookkeeping does not need another dashboard to learn. They need the invoices handled. So Nika is positioned as an employee you hire in five minutes. She watches a mailbox, pulls every invoice, files it into your records, and asks before she ever guesses.

You pay per finished invoice, $0.40, and $5 gets her started. No subscription, no card required to look around.

The one design decision that matters: ask, do not guess

Every AI bookkeeping demo shows a model confidently extracting a supplier name and a VAT number and being wrong. The interesting engineering problem is not extraction, it is confidence. Nika routes anything she is not sure about to a human with a specific question, instead of filing a guess.

A filed error is expensive to unwind. A held item waiting for an OK is cheap. The copy on the site says it plainly: "Anything she isn't sure about waits for your OK." That sentence is the product.

Per-task pricing instead of seats

A part-time bookkeeper for four hours of weekly invoice work does not justify a salary, holiday, and training. Per-finished-task pricing matches the actual job: it is four hours a week, not a role. No work done, no charge.

What is live and what is not

Nika is the only live employee today. The site teases more employees (a speed-to-lead caller, a win-back agent, a review manager), but those are clearly marked "coming soon", and I will not pretend they are shipped. One real employee doing one job well is the honest version of this product.

The stack is a Next.js app with a three-language site (English, Greek, Ukrainian). If you are building an "AI agent" product, my advice is to shrink the scope until the agent does one verifiable job and nothing else. That is where the trust is.

Top comments (0)