Human capital management leader Workday serves more than 11,000 customers worldwide, supporting a community of 75 million users. In FY25, over 90% of deployments were completed on time. After debuting on both the Fortune 500 and S&P 500 in 2024, Workday introduced its Agent System of Record in 2025 to help enterprises manage the emerging wave of AI agents.

In the first part of this interview with HPE’s North American AI GTM Leader Clyde Gillard, Joe DosSantos, VP of Enterprise Data & Analytics at Workday, explored the shift from generative to agentic AI and the data and security disciplines that make it real.

In this second part of the conversation, DosSantos delves into the limits of today’s AI agents, the critical role of a semantic layer in ensuring accuracy, strategies for teaching agents to acknowledge “I don’t know,” what Workday’s Agent System of Record means for treating agents as true coworkers, and why ROI and clear prioritization are key to unlocking AI’s highest value.

From LLMs to the enterprise: Why structured data (and semantics) matter

DosSantos argues that today’s well-known agents inherit strengths and weaknesses from the data they’re trained on. “It’s  interesting that the agents we know and love as brand names were largely trained on unstructured, publicly available data.”

That helps with summarization and stylistic tasks, he notes, but exposes gaps in precision: “These models aren’t very good at math. They can’t count very well. They don’t know; they don’t understand some of the constructs.”

Closing that gap, DosSantos says, requires a durable business vocabulary. “The secret sauce to that is around a semantic layer.” He frames it with a deceptively simple example: “How do you help people understand a phrase like ‘churn’?”

DosSantos explains that when organizations agree on meaning and where the data lives, they can bind access controls, reduce hallucinations, and translate natural-language questions into trustworthy analytics.

Teaching agents to say “I don’t know”

Human analysts naturally ask for clarification; AI agents need to develop that same instinct. “There’s always going to be a first time somebody asks something; there’s always going to be a moment when you need to say, ‘What do you mean?’” DosSantos sees the solution in a strong feedback loop — capturing ambiguous queries, enriching the semantic layer, and encoding refusal behaviors when the system’s knowledge falls short.

He says it matters because “the very first time that you get a bogus answer, you’ll never use it again.” That is why his team experiments with interface guardrails — up to and including hard-coded fallbacks — because, as he puts it, agents can be “notorious, hallucinating liars.”

Agents as coworkers: Defining roles, responsibilities, and the future of work

Workday sees AI agents as more than just software — they are designed to function like coworkers. As DosSantos explains, “An AI agent acts and talks and operates a lot like a human.” To support that view, Workday has introduced what it calls the “agent system of record.” This framework, he says, acknowledges that “an agent can’t just exist. It needs to be brought in for a purpose, working in an organization with a job and function, like it needs to have a boss.”

DosSantos emphasizes that agents require the same kind of structure as employees: “You need to understand the ability of these agents. How do they talk to each other? What are their roles and responsibilities?” Just as companies hire people into departments and expect them to collaborate across teams, agents too must be given clearly defined roles and expectations.

Looking ahead, he believes agents will increasingly take on specific functions. “The idea of an agent taking over a specific function is the future,” he says. But that shift doesn’t mean human work disappears — it evolves. Drawing an analogy, he notes, “It’s just like when we moved from horses to cars; it didn’t blow up everything. But did people need to become mechanics? Yes. It changes the nature of the economy.”

DosSantos is optimistic that if deployed correctly, agents will free people to focus on what only humans can do. “The jobs that we can do become more human. The decision-making, the influence, the actual alignment of people — these are not things that robots do.” In his view, AI agents should handle the repeatable, less enjoyable tasks, leaving space for people to build culture and drive decisions. Ultimately, he sees agentic AI as part of a larger automation journey.

Measuring ROI and choosing the right use cases

Moving ahead, DosSantos frames ROI for AI agents the same way organizations evaluate human roles. “If we agree that in many respects an agent is doing the work of a person, some of us at work sit in call centers and some of us sit in profit centers,” he says. The difference is in how their impact is measured: “If you sit in sales, you make us money. If you sit in purchasing, you help us spend less.”

By extension, the ROI of agents must be assessed along the same lines. For example, if an agent does Business Development Representative (BDR) outreach, measure the pipeline created; if it negotiates, track cost savings.

When it comes to prioritization, DosSantos cautions against chasing AI for its own sake. “Thinking about AI in a vacuum is a fool’s errand. You need to start thinking about what’s important to your company,” he says. For him, ROI always ties back to function — revenue generated or costs reduced — and the key is anchoring value in metrics executives already recognize.

To make those choices, the company maps use cases against value and feasibility: “If you think about cost and value, the magic quadrant, as Gartner would call it, is where we have the data, the know-how, and the assets to deliver, and it’s a high-value impact.” This is reinforced by “value engineering exercises that help us weigh effort against value, but the value is always anchored to projected cost savings or revenue generation,” DosSantos explains.

CDO Magazine appreciates Joe DosSantos for sharing his insights with our global community.