A South African platform shows how generative AI can handle sensitive financial data without cutting corners on security, accuracy, or oversight
By Mosepe Celesta
Artificial intelligence promises to save businesses time and money. But for the people who manage payroll — handling some of the most sensitive personal and financial data in any organisation — the promise has come with a nagging question: can we actually trust it?
A South African payroll platform believes it has found an answer. And it did not arrive by moving fast.
Deel Local Payroll, home of the PaySpace payroll platform, yesterday unveiled the thinking and engineering behind AI Assist, a generative AI tool built specifically for payroll environments. The system is designed to handle the relentless stream of queries that bogs down payroll teams daily — why did my net pay decrease, how much overtime did I earn last month, what taxes did I pay — while keeping sensitive data exactly where it belongs.
The backdrop to its development is a technology sector grappling with a trust deficit. A 2025 KPMG survey found that only 46 per cent of company workers trust AI, and that proportion is shrinking. More tellingly, 56 per cent of respondents said they were making more mistakes because of AI, not fewer. Against that backdrop, the instinct of many organisations has been caution. Deel Local Payroll chose a different response: build something that earns trust by design.
“Payroll staff juggle a lot of queries,” said Warren van Wyk, Director at Deel Local Payroll. “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.”
The challenge is not trivial. Payroll environments are dense with private information, regulatory obligations, and access restrictions. A tool that responds to keywords is not enough. Payroll queries are contextual — an employee asking about their deductions needs a different answer from a payroll administrator querying a tax code formula, even if the underlying words look similar. The AI has to know the difference, and it has to know who is asking.
AI Assist was built with that complexity at its core. Employees can ask specific questions about their own payslips — deductions, tax contributions, overtime calculations — and receive answers calibrated to their individual records. Payroll teams, meanwhile, can use the tool to surface specific components and formulas, interrogate the PaySpace knowledge base, and explore how the platform’s features apply to their organisation’s particular setup. The system provides answers relevant to each organisation’s payroll configuration, including calculation formulas and components linked to specific tax codes.
But the engineering team at Deel Local Payroll was equally focused on what the AI should not do.
Data security in payroll is not a secondary concern — it is the whole game. The team identified the critical failure points of generative AI in sensitive environments: unauthorised access to other employees’ salary information, sensitive data surfacing in unintended places, and the risk of private information being absorbed into an AI model where it becomes difficult to contain or audit.
Their response was architectural. Every time a user interacts with AI Assist, the system creates a temporary instance on secure Microsoft Azure cloud infrastructure. No data leaves that environment. What a user can see through AI Assist mirrors precisely what they are authorised to access within the PaySpace platform — nothing more. An employee cannot query a colleague’s salary if they do not have the system privileges to view it in the first place.
“We are very conservative with AI Assist’s features,” Van Wyk said. “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. If a feature cannot meet those criteria, it goes back onto the shelf.”
The team also confronted one of generative AI’s most persistent weaknesses: hallucinations. AI models can generate plausible-sounding but factually incorrect responses, a problem that becomes particularly costly when the subject is someone’s salary or tax liability. Rather than relying solely on the AI model to manage data retrieval and response generation, Deel Local Payroll built subsystems that handle specific queries and data access separately, passing curated information to the AI for processing.
“We cannot use our customers’ data to train the model,” Van Wyk explained. “Instead, we have systems that curate the right data and hand information to the AI, which then responds to the user. This 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.”
The development philosophy has been deliberately measured. In an industry where speed to market is often treated as a virtue, Deel Local Payroll has chosen to move carefully, test rigorously, and involve customers through pre-engagement sessions and beta testing before expanding AI Assist’s capabilities. Each new feature is evaluated against a clear baseline: does it make things genuinely easier, and can it be delivered with the accuracy and security payroll demands?
It is an approach that stands in deliberate contrast to the hype that has surrounded generative AI since it entered the mainstream. Van Wyk is direct about the gap between what AI is often promised to do and what it reliably can.
“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 is that simple, but it is 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.”
The early results, by the company’s account, have been striking. Time spent on routine query resolution has fallen significantly. Payroll staff report being able to direct their attention toward more complex and strategic work. And for employees — who typically interact with payroll only when something feels wrong — the ability to get a clear, contextual answer without waiting for a human intermediary represents a meaningful shift in how organisations communicate about pay.
What Deel Local Payroll has built is not the most ambitious AI project in the market. It is, by design, one of the most restrained. And in a sector where a single data breach or a miscalculated payslip can erode years of institutional trust, restraint may be precisely the right instinct.
Generative AI done right, it turns out, looks a great deal like patience.
Similar Posts by The Mt Kenya Times:
- Kipchoge completes the circle
- Who exactly has been governing Kenya?
- What the Job story teaches Christians about suffering and faith
- Kenya Police Bullets eye CECAFA crown after crushing Yei Joint FC in semis
- FKF opens talks on league kick-off after clubs push back
About the Author
