Workday and Sana Present an Agent-Led Vision for Enterprise Software

Joel Hellermark opens by describing his interest in building teams. Studying groups such as the early Macintosh team and Xerox PARC, he saw small interdisciplinary teams working differently from traditional command-and-control organizations. He believes agentic AI creates an opportunity to rethink how companies and products are organized.

Sana uses a five-level model of autonomy to describe that transition. The starting point is manual work: tickets move among people across many enterprise systems. Copilots then draft suggestions, but employees still move among multiple interfaces.

At the third level, Hellermark says agents begin doing work and bring people in when needed. The fourth is policy-bounded autonomy, where systems use organizational knowledge and rules to automate more of a process. The proposed fifth level adds greater autonomy and self-improvement within guardrails; he says reaching it will take a long time.

His first proposed foundation is a multiplayer AI experience, in which teams of people and agents share context and route work across departments. The second is living memory: company knowledge and policies that can update, rather than systems that forget context between interactions.

The third foundation is ambient agents that work continuously in the background. Hellermark imagines them checking compliance, updating rules and developing workflows. He presents this as the product direction rather than a claim that every such workflow is already complete.

Hellermark introduces Sana for ITSM, intended to handle employee IT requests across the lifecycle, and a Travel Agent for conversational booking and associated expenses. He then starts a demonstration of a unified interface for recruiting, travel, IT and custom agents.

In the demo, Hellermark asks the system to book a Sunday flight home. He says it uses personal memory, company policy and calendar information while continuing the task in the background. The task list also shows recruiting and a Salesforce access request.

The recruiting example concerns hiring five account executives in Singapore. The interface shows a candidate pipeline and traces of actions such as job requisitions, postings and interviews. Hellermark says the agent escalates decisions, including offer letters, for a person to approve.

After an approval, the demonstration shows the system noticing a repeated decision and asking whether to update company policy for similar cases. Hellermark also reviews an out-of-band salary example. He presents this as the proposed living-memory pattern, with a human deciding whether to grant further autonomy.

The IT example involves Salesforce access for a new sales lead. Hellermark approves the request, and the interface suggests automatically approving similar in-region requests in future while escalating broader cases. The screen says the proposed policy can be reversed.

The demonstration returns to travel: Sana suggests a flight based on the calendar, then shows expense filing and calendar work after the booking. Hellermark describes a future in which employees supervise background tasks and agents surface decisions when necessary.

Hellermark contrasts this agent-operator model with employees managing many separate copilots. He describes a shift toward proactive systems with organizational and personal memory, and calls it a re-founding moment for Workday. He then invites Workday co-founder Aneel Bhusri on stage.

The pair discuss the so-called SaaS apocalypse. Bhusri says he initially felt nervous about the implications but came to view AI and enterprise software as complementary. He calls AI a reasoning engine and says he has not seen a comparable technological shift in his career.

Bhusri argues that reasoning models alone are not well suited to deterministic processes such as payroll and closing books, where organizations need reliable, complete results. In his view, combining reasoning with established process systems is the opportunity.

He thinks software centered on static orchestration across applications may be vulnerable as agents take on workflows. Deep domain applications may face pressure later, he says, but Workday can also use models and tools such as Claude Code and Cursor to improve productivity and compete.

Asked about Sana for ITSM, Bhusri does not predict immediate market leadership. He expects a gradual path: start with a narrower product, add ticketing and workflows, then move upmarket. He expects a similar development pattern for travel.

Bhusri says Workday must succeed in AI as well as cloud computing if it is to endure across technology generations. He describes returning to work with Hellermark and colleagues, and the conversation turns to a role announcement.

Bhusri announces that Hellermark is being named Workday's chief AI officer, ahead of a formal announcement later that day. He describes his own role as coordinating experts through what he sees as a reinvention of work, not merely an AI feature added to an existing product.

Hellermark says the opportunity is to sell completed work as well as software. He compares progress toward autonomy with self-driving cars: edge cases took longer than many expected, and people must remain involved while systems are tested and improved.

He identifies policy and knowledge graphs as a proposed foundation for agents, describing a ‘world model of work.’ Bhusri says Workday is trying to return to startup habits while making ownership of AI vision, applications and APIs clear.

Hellermark expects boundaries among software categories to blur as shared knowledge, connections, agents and human collaboration become common foundations. He sees employee onboarding and offboarding as reasons for Workday to expand into adjacent areas such as ITSM.

He says the combined company can increase agent autonomy over time and bring products to more employees. Responding to criticism of Workday's user experience, he sets an internal ambition to make it the ‘Hermès of enterprise software.’

Hellermark points back to the demo's single interface as an example of AI changing how users interact with enterprise applications. Bhusri closes by expressing confidence in the combined teams and jokes that his advice to the new AI chief is not to screw it up.