Don’t Leave Payroll to AI’s Probabilities: Workday’s Vision for “Lawful AI” and a New Way of Working

Payroll amounts change slightly every time you run the calculation. No business could use an application like that. Even if AI predictions are right 90% of the time, companies cannot confidently rely on them if the remaining 10% shows up in payroll or financial closes.

In his keynote at Workday Rising 2026, the company’s annual conference in Las Vegas, Workday co-founder Aneel Bhusri explained this problem as the difference between “probabilistic AI” and “deterministic business applications.” He positioned “lawful AI”—which keeps AI’s reasoning aligned with corporate rules and permissions—as the centerpiece of Workday’s new direction.

Why Bhusri Returned as CEO#

Bhusri co-founded Workday in 2005 and handed the CEO role to Carl Eschenbach in January 2024. He returned as CEO in February 2026. In his keynote, he said the new technological era ushered in by AI presents an opportunity for Workday to deliver products.

At the same time, he warned that falling behind on AI could threaten the business. Having been closely involved in building the company’s products, he explained, he could not simply watch the changes unfold from the sidelines.

The company has also been rethinking its organization and customer relationships, alongside its technology. One example is its appointment of Joel Hellermark, founder and CEO of Sana, the AI startup it acquired in 2025, as Workday’s chief AI officer (CAIO).

Putting AI on “Deterministic Rails”#

AI produces probabilistic answers based on its inputs and context. Business applications follow defined rules to return the same result every time. Drawing on this distinction, Bhusri stressed that payroll and quarterly financial closes cannot be left to probability. For AI to deliver value in work where mistakes are unacceptable, it must first earn trust.

Workday’s proposed solution is to layer AI on top of deterministic applications. The applications manage rules, approvals, and access permissions, while AI handles reasoning and prediction. This, the company says, allows AI to go beyond answering questions and making suggestions to execute work within established controls.

Bhusri cited Workday’s data and business context as the foundation for this approach. According to the company, it manages data for approximately 80 million employees using a common data model and processed roughly 1.4 trillion transactions on a single platform in 2025.

Workday says it has been working on data and machine learning since 2013 and was developing its own large language models (LLMs) by 2020. Information such as expense approvers, applicable policies, workflows, and security permissions also resides within its business systems. The company calls these mechanisms “deterministic rails” for AI. They define what AI can do and whose approval it needs.

17 In-House Agents and Four Ways to Use the Platform#

According to Bhusri, Workday has 17 “lawful” agents in operation across HR, finance, and IT. Around 50 were under development when he returned as CEO, but the company narrowed its focus to the most important ones.

Companies adopt AI in different ways. Workday has therefore provided four paths on the same platform.

• Use agents developed by Workday.
• Use the “AI API” to access Workday capabilities through another company’s front end.
• Build custom agents with “Extend AI” (formerly Extend Pro).
• Connect data from Workday and other systems through “Data Cloud.”

The company also plans to bring third-party agents into an “agent system of record” for centralized management and governance. It aims to extend Workday’s security and business process framework across these different ways of using the platform.

These announcements alone, however, do not establish how broadly the offerings can be applied to an individual company’s operations. Any deployment requires a clear understanding of the tasks involved, approval conditions, and how malfunctions will be handled, along with a design that grants only the necessary permissions.

Making Sana the Default Home Screen#

The keynote also included an announcement that Workday’s default home screen would switch to its AI offering, Sana. Making AI the entry point for users signals an effort to change how they request work and access information. Together with the Sana acquisition and Hellermark’s appointment to the leadership team, this shows that Workday is seeking to embed AI into how its products are used, beyond adding it as a feature.

Bhusri said AI should expand people’s capabilities rather than replace them. He argued that companies must acknowledge changing job roles and respond responsibly. He urged the approximately 30,000 users gathered at the event to spend less time on routine tasks and more on making decisions and moving their businesses forward.

Check Your Own “Rails” Before Deployment#

The core idea behind “lawful AI” is that business rules must be defined in a form that systems can work with before AI is entrusted with tasks. If approvers and policies remain unclear, the scope of what AI can execute will also be unclear. Before adding more types of agents, companies need to decide what to delegate and where human review should come in.

To apply Workday’s announcement to your own company, start by choosing one narrowly defined process, such as expense reimbursement or internal inquiries. Next, identify the data it uses, the rules that apply, and the required approvers. Then distinguish between actions AI can perform on its own and those that require human review. After a trial, assess processing accuracy and the time spent on review.

The question is not simply whether to adopt AI, but which tasks to entrust to it and under what rules. Choose one process in your organization and start by reviewing its approval flow and permissions.

Source: https://japan.zdnet.com/article/35253538/