Artificial Intelligence Done Right

Generative AI can be a powerful business knowledge service, saving considerable time for employees. Yet, concerns grow around AI accuracy and how it handles sensitive information. According to a 2025 KPMG survey (https://apo-opa.co/3UgIjBb), only 46% of company workers trust AI, and that proportion is shrinking, especially since 56% say they are making more mistakes because of AI.

Despite such growing doubts, the technology can be an exceptional information source and time-saver, especially for sensitive, technical, and contextual queries. A South African payroll platform developed these advantages, using AI to provide detailed and contextual information for customer employees and payroll teams.

Payroll staff spend considerable amounts of time answering routine questions, such as why someone’s net pay decreased, what amount of taxes they paid, or how much overtime they earned last month. Said staff also have queries around formulas and components, regulations, payroll statuses in different business units, and creating new payroll structures.

“Payroll staff juggle a lot of queries,” says Warren van Wyk, Director at Deel Local Payroll, home of the PaySpace payroll platform. “Can generative AI handle those queries while adding real context? We believe it can, but that’s not enough. There has to be trust, accuracy, and oversight. How do we create those things? That was the challenge we set for ourselves.”

Building an AI assistant worthy of payroll

Generative AI excels at conversing in natural language, changing how we communicate with technology. Already, search engines produce AI-compiled answers that users can expand with follow-up questions.

Yet, search answers are relatively simple. An AI that handles payroll queries must meet specific parameters. Van Wyk didn’t want another chatbot that simply responded to keywords. Payroll information is very contextual, and the AI must reflect that. Payroll environments are also full of private information, security limits, and regulations that require strict oversight and privileges.

These stipulations define the baseline for a useful generative AI payroll assistant. Thus, the engineers at Deel Local Payroll developed AI Assist, a remarkable assistant that can answer payroll questions tailored to the person asking. Employees get specific answers about their payslips, from what they pay in tax to deductions and overtime.

Payroll staff additionally use AI Assist to surface and study specific payroll components and formulas, handle detailed queries, and explore the PaySpace platform’s knowledge base. AI Assist provides answers relevant to an organisation’s payroll setup, such as calculation formulas or listing components sharing specific tax codes. Payroll departments also use AI Assist to understand and exploit the PaySpace system’s features.

Building trust in AI

However, being helpful is not enough. Generative AI can make mistakes. Firms worry about AI’s access to information. Can someone use AI to learn other people’s salaries? Will the AI unearth forgotten and under-supervised data? What about the current data? Payroll information is incredibly sensitive. What if the AI causes a data leak or embeds that information in its model?

Developers understand how crucial it is to answer those concerns and use them as their baseline, says Van Wyk.

“We are very conservative with AI Assist’s features, focusing on several important things from the start. We host pre-engagements with our customers to identify specific use cases, and we established several rules. The AI will be native to the platform, not an integration. It will conform to ISO and SOC standards, and it will show a user only what they have authority to access. Nobody can ask AI Assist about other people’s salaries if they don’t have the authority to see those details. If a feature can’t meet those criteria, it goes back onto the shelf.”

The team designed AI Assist’s infrastructure to ensure sensitive data doesn’t appear to the wrong user or end up inside an AI’s model. Whenever someone interacts with the AI, it creates a temporary instance on secure Microsoft Azure cloud infrastructure. No data leaves that space, and access to data depends on the user’s profile as determined by the PaySpace platform.

To reduce the risk of hallucinations and other mistakes, the system doesn’t rely solely on the AI model. It runs subsystems that handle specific queries and data access, passing information to the AI.

“We cannot use our customers’ data to train the model,” says Van Wyk. “Instead, we have systems that curate the right data and hand information to the AI, which then responds to the user. This is very important and has two advantages. It stops data from leaking into the AI model, and it improves answer accuracy because we directly control the mechanisms that link the data with AI Assist.”

Intuitive and effective

Hype and misconceptions cloud generative AI’s potential. Deel Local Payroll avoids these issues by focusing on the technology’s most obvious value and deliberately slowing its development pace to ensure maximum, focused benefits.

“We start from a basic premise: how can generative AI make things easier for our customers, whether they are working on payroll or need answers from payroll. How do we create thoughtful communication that gives them answers they can trust? It’s that simple, but it’s still very difficult to figure out because this is a new technology with many unknowns. So, we move carefully, test thoughtfully, and involve our customers through user engagements and beta testing.”

The results are incredible: “It’s amazing; there’s nothing like it. This will change payroll and every aspect of how we engage with business information. Generative AI done right is changing business for the better.”