Payroll gets attention for two reasons: somebody has not been paid, or somebody is trying to sell an AI strategy. Neither is reassuring. The first is a human problem. The second is usually a slide deck looking for one.
Amit Levi's work at Papaya Global sits where those two worlds meet. The company helps employers hire, manage and pay people across borders. That means local rules, tax questions, contracts, payment cut-offs and documents nobody designed to match each other. It is not a category that improves because a chatbot sounds friendly.
He came to payroll by way of product problems
Levi's route to Papaya was not a straight line through HR software. Before joining the company, he held product roles at CoolaData and Yokee Music, then product and marketing roles at Anodot.[1] The common thread is less glamorous than a career story normally allows. Each role involved taking a complicated system, finding the bit that made users give up, and making it less annoying.
Papaya promoted Levi to Senior Vice President of Product in February 2024.[1] By then, the product had a difficult brief: make global workforce operations feel less like a collection of country-specific exceptions. Every new market adds rules about employment, payments, tax and termination. A system can hide that complexity. It cannot wish it away.
The useful part of AI is not the answer it writes. It is the bit of work it removes without creating a new thing to check.
The first AI plan did not survive customer contact
Papaya's early instinct was familiar: make AI an add-on and charge more for it. Customers did not bite. Levi told Bessemer Venture Partners that buyers now expect AI as part of the product.[2] That is a better result than a polite focus group. It forces a company to decide whether the technology is actually making the core product better, or merely giving sales a fresh noun.
Levi's team went back to the work. They listed the places where customers and internal teams were getting stuck, then matched those problems with capabilities that were already good enough to use: language models, document parsing, anomaly detection and forecasting.[2] No grand transformation programme. Just a search for tasks that were slow, repetitive, and expensive to get wrong.
Start with the tasks that ruin everyone's afternoon
One result was a compliance knowledge agent connected to Papaya's employment database. A manager can ask a country-specific question without joining the support queue. If the system is not confident, it sends the question back to a human expert. That last part deserves more attention than it gets. In payroll, an AI system that guesses smoothly is not helpful. It is just quicker at creating a problem.
Onboarding got the same treatment. A manager can describe the hire, country and required documents in plain language. The platform creates the form, sends it to the worker and checks what comes back. Bessemer reports that the process went from days to hours.[2] It is a company-reported outcome, so take it as evidence to test, not a number to paste into a procurement deck. But the before-and-after is at least clear enough to argue with.
The payslip story is the more honest one
Papaya also built an agent to match payslips to the correct employee across languages and formats. The first version used names. That did not work well. It turns out names can be abbreviated, transliterated or attached to the wrong person on a document. A shocking discovery only if you have never worked with payroll data.
The team added local rules and exclusions. Bessemer says the match rate began at 30% and later approached 85%, with 99% set as the point where the process would need very little human review.[2] This is the story worth paying attention to. The first version failed. The team found out why. The number improved, but they did not pretend 85% was good enough for a sensitive document.
What Levi actually changed
Levi changed the question the product team asked. The easy version was, “Where can we put AI?” His version was, “Where are people stuck, and what would make that job less fragile?” That sounds like a minor wording change. It is not. The first question produces features. The second produces a list of work that needs fixing.
He also appears to have been willing to let a commercial idea fail. Papaya tried selling AI as an add-on. Customers did not want it. Instead of forcing a premium tier around a buzzword, the team put the technology inside the existing work.[2] That is not a flashy founder story. It is a product decision made by someone listening after the pitch stops working.
The payslip project says more about his approach than a list of releases could. The first version was weak. It improved, but it was not waved through as finished. The team kept a hard reliability target and kept the people in the loop until the system earned more trust.[2] That is the work of product leadership in a regulated category: decide where a shortcut is safe, where it is not, and do not confuse a better demo with a better system.
The more interesting part of this story is Levi's product instinct. His background spans product, data and marketing, but he did not bring a marketer's answer to payroll. He did not try to defend the AI add-on once customers rejected it. He did not treat a 30% match rate as a launch. And he did not remove people from the process before the system had earned the right to ask for less review. The pattern is clear: start with the ugly bit of work, learn where it fails, and make the improvement prove itself. That leaves less room for theatre. It also explains why his work has changed the shape of the product.