Where AI really helps in payroll, and where it doesn’t

The obvious way to bring AI into software is a chat. For the first feature our customers use AI with, we decided against it in the end. Why, and where AI makes the difference for us instead, is a story about asking the right question.

Behind it is the same test process we use for every feature: prototypes, real users, thinking aloud, iterating. What we learned didn’t confirm our direction. It changed it. And the feature that came out of it is available in Paymira today.

The obvious reflex: put it all in a chat

When people talk about AI in software today, they think of chat first. Type what you want, and the AI does the rest. Intuitive, flexible, modern. That is how we approached it too.

The more concrete we got, the clearer it became: for payroll, chat is often the wrong tool. Preparing a payroll run isn’t a conversation. It is a structured task with clear fields and defined values. Someone recording hours worked doesn’t want to phrase anything. They want to see what is recorded, what is missing and what changes. A structured interface gives that certainty. A prompt doesn’t. Quite the opposite: every time you rephrase the sentence, the time you wanted to save goes up in smoke.

The insight sounds simple. It wasn’t. A chat is tempting, and it would have been easy to do what everyone else does, just to show off a feature with AI quickly. But a feature that doesn’t help in daily work doesn’t keep our promise.

Where AI really makes a difference

The more interesting question was this anyway: where is the actual problem in payroll, and can AI help there?

The answer came from talking to our customers. Variable pay components (hours worked, expenses, bonuses, overtime) are part of almost every payroll run, often as many separate entries. They come from scattered sources: from time tracking as Excel, from accounting as PDF, sometimes in an email. Every customer has their own logic, and every month brings the same manual work.

That is the actual problem. Not the single change, but the recurring, unstructured batches of data that take the same route through the same manual bottleneck month after month: type it up, check it, next line.

And here, the very variety of formats that makes a chat unfit makes AI indispensable. This data can’t sensibly be typed in by hand. It has to be read, understood and assigned. That is what data import in Paymira does. You upload your file the way it comes out of your system. An agent works out what data it holds and for which period, and assigns it to the right employees and pay types. You check it and accept it. What used to be typed up line by line is done in a short check.

Reliable before fast

Data import went through our test process too: prototype, real users, thinking aloud, refining. Nothing goes live that we haven’t seen in the hands of people who work with payroll every day.

AI has its own bar: reliable before fast. In payroll, a confident mistake is worse than honest uncertainty. A feature has to show what it recognised, and just as clearly where it isn’t sure. That is why data import shows its source for every value. You see how reliably each value was recognised and, with one click, the place in your original file it came from. The hours, bonuses and expenses come from you, so you approve them. No recognised value changes a payslip before a person has seen it.

What this means for you

AI in payroll isn’t an end in itself. It is a tool, and as with any tool, what counts is knowing what it’s good for. We looked at both: where a chat seems tempting but costs certainty, and where AI makes the decisive difference.

When variable data from dozens of sources has to reach the system reliably month after month, AI isn’t just useful. It is the only way that scales. Not because it sounds modern, but because the problem is real, the volume is large and the manual work stopped being justifiable long ago.

Most promises about AI start with what the technology can do. We started with what the problem is. That makes the difference.