Manual salary calculations for 30,000 annual new hires made overpayment a constant risk

The City of Copenhagen hires 30,000 full time, part time and temporary workers—a nurse, a teacher, a social worker—bound  by 57 labor agreements, 442 salary grades, and 928 wage subsidies, complex enough that overpaying was nearly guaranteed. In 2023, the city set out to let AI calculate those salaries instead, but only once it earned the trust of the humans who'd sign off on every number. For a city budget, that trust is the difference between money well spent and money that could have hired more of them.

Susanne Ninna Hansen, the City’s Head of Automation and Development, made a commitment that sounded ambitious at the time; she’d use AI to solve the salary calculation problem. Back then, generative AI was still in its early days. The technology wasn’t proven. The city’s payroll rules were convoluted. But Susanne believed that if she kept humans in control of the outcome, AI could handle the complexity underneath.

“We started with a simple principle,” she recalls. “Let AI do the work. Keep humans in control.”

The city did the unglamorous groundwork before AI touched a single case

The city had already implemented ServiceNow HR Service Delivery, handling one million cases a year. What they needed was a way to stop manual salary calculations. The solution, ServiceNow AI Agents that read résumés in any format, interpret each candidate’s prior experience, and recommend a salary range within a governed workflow. Deterministic calculation logic ensures every figure is accurate and traceable, and recruiters review and verify each recommendation before an offer is made. AI speeds up the starting point without taking the decision out of human hands.

“We set out to use AI for almost everything, and today I think it's 50-50. We've scripted a lot of it, like the mathematical counting of days,” says Jacob Teudt Jørgensen, IT Architect, City of Copenhagen.  Using ServiceNow enables AI to deliver the most value when it works within a governed workflow, where AI interprets the information, and the platform applies precise, auditable logic to the calculations.

With 57 collective labor agreements and more than 100 union contracts, Copenhagen knew it couldn’t swallow everything in one hit. It started small, focusing on four of the most common labor agreements.

It also built on a culture of data discipline. The city had painstakingly documented and standardized years of existing manual records and processes; the case-handling problem existed with or without AI, not in anticipation of it.

“AI is only as good as the unglamorous groundwork underneath it, and many organizations skip that groundwork,” says Susanne. “If you put garbage in, you get garbage out... with AI, you get the same garbage, just faster.”

Eight specialists checked every AI calculation until the system earned trust

The first version wasn’t perfect. For a public agency in the EU, one that must follow the General Data Protection Regulation (GDPR) and AI Act laws, this was a difficult balancing act.

But Copenhagen didn’t abandon the project. Instead, the team iterated. During the pilot phase, up to eight business specialists validated every AI calculation by comparing it against their manual work and flagging errors. The team fed those errors back into the system, refined the prompts, and updated the knowledge articles. This tightened the feedback loop.

Jade Sax Kristoffersen, a City of Copenhagen Payroll & Personnel data analyst was one of the initial test users. “In the beginning, we sometimes had eight, 10 minutes waiting time, and that's a long time, because many of us had this feeling that I could have done the case by now manually. But we also understood that this was a project in process, and for us to gain the benefits later, we had to have these experiences.”

“We took beatings in the beginning,” Susanne says. “But we committed to the partnership. We treated this as a learning opportunity, not a test the vendor had to pass.”

Payroll specialists traded data entry for compliance and exception review

Slowly, the accuracy improved. User confidence in the figures went from 70% to 90%, then higher. The test group of eight has expanded to 30 regular users, all providing feedback to keep the refinements coming.

“From day one we’ve seen this tool as a personal assistant. It is there only to help us,” says Jade. “The project leads have made it very clear that we’re here to verify and check all outputs.”

The payroll specialists who used to manually calculate every salary didn’t disappear. They stayed but their work transformed. Instead of data entry, they reviewed exceptions, handled edge cases, and ensured compliance. They did the thinking while the AI did the data processing. “Now we don't have to go in and look at all the singular documents. The AI assistant does this for us,” says Jade.  Salary calculations have dropped from 10 minutes to less than two minutes.

“We freed 2.5 FTEs,” Susanne says. “But it wasn’t a headcount reduction. It was a redeployment. Those people are now doing strategic work.”

Salary calculation changes freed up to 4,000 hours in the simplest cases alone

In its first phase, the salary composition automation delivered measurable outcomes:

• 2 days faster in approximately 75% of new-hire salary calculations.
• 3,000–4,000 hours saved in the simplest cases alone, just from eliminating manual salary grade interpretation.
• 2.5 FTEs redeployed to higher-value work, with no job losses. The team size stayed the same; the work shifted.

The payroll fix became a model for 30-plus other AI projects citywide

The payroll project became so successful that the City of Copenhagen is hoping to expand it to other departments.

“We’ve proved you can extract meaningful data from paper records, and have an AI interpret the information and make a business evaluation,” says Jacob. “We’re now exploring use cases involving the analysis of data breaches and criminal record checks.”

Future projects will follow the same principle as payroll; humans in control, AI handling the complexity.  “Work with the data you have today, don’t wait for ‘perfect’,” says Susanne. “Start small, learn, then scale; otherwise, you’ll be stuck making no progress for five years.”

Copenhagen has flipped the narrative. Instead of focusing on “Will AI replace our staff?” the city asked, “How do we free our staff from error-prone work, so they do what they were hired to do?” That shift in perspective made all the difference. It’s now the model for public sector AI adoption.