Workday Executives Argue for Data Context and Controls for AI Trust

Workday is making its case for trustworthy enterprise AI at Workday Rising EMEA in Barcelona this week, arguing that agent trust rests on two engineering pillars: governed data context and enforced agentic controls. Executives Yasmeen Ahmad and Gabe Monroy discussed Data Cloud and the C1.ai identity acquisition in interviews with diginomica.

Enterprise software’s “SaaSpocalypse” panic has cooled, but the pressure it exposed hasn’t. Vendors must now prove they are the right home for agentic AI — not just the smartest models, but the best-guarded ones. Workday’s answer, laid out by its tech leadership ahead of Rising EMEA in Barcelona, is that trust is an architecture problem, not a model problem.

That framing came from Yasmeen Ahmad, who left a role at Google to join the company. Her reasoning, in an interview with diginomica at Workday Rising Las Vegas, cuts to the core of the enterprise AI debate: “It’s not just a model that you stick alongside some data and it works and delivers value. There’s an entire operating system that you need around the model.”

Why context engineering beats model size

Ahmad’s argument is blunt. Fundamentally, those problems are not the problems of making models smarter, as she told diginomica, noting that it is a governance challenge.

Four layers workday data cloud
Photo: blog.workday.com

Workday’s wager is that twenty years of HR and finance transactional data — payroll rules across jurisdictions, ledger structures, accounting enforcement — constitutes what Ahmad calls a “golden source” the company traditionally never opened up.

The stack is flipping, she argues: agents arrive headless, and whoever holds the authoritative context wins. A third-party agentic vendor could theoretically scrape similar knowledge from the internet, but, as she put it, it is not the same as a system that enforces the context.

Workday’s own engineering blog adds a concrete layer to this: Early Access to pre-configured Financial and HCM Semantic Models, announced at Rising. These aren’t auto-generated schemas — they package codified KPIs, dashboard logic, and business rules covering objects like Headcount, Cost Centers, and Voluntary Turnover.

The company’s claim is specific and testable: a question like “what’s our regrettable attrition rate for people managers in EMEA this quarter” should return the same answer from any agent, any tool, every time, because the definition lives in the semantic model rather than “in a prompt somebody wrote last Tuesday.” That consistency, per the blog post, is “the difference between AI you can demo and AI you can govern.”

Inside the Agent Context Service

Context delivery is where the engineering gets hard. Workday CTO Gabe Monroy confirmed to diginomica that Workday Data Cloud functions as a context layer for “people and money” data, blended with external sources via zero-copy ingestion through the Prism product — badging data merged into an org chart, for instance, to power return-to-office analytics agents.

But Monroy went further, describing an Early Access capability called the Agent Context Service as very unique and differentiated engineering we’ve done to make sure we’re grabbing the right context at the right time, with layers of caching and machine learning and scoring on that data.

Translation for the non-specialist: rather than flooding an agent’s context window with everything, the service scores and ranks data so the agent harness receives only the relevant slice, efficiently. Policy documents, Monroy noted, can be injected separately — Data Cloud is a major source, not the only one.

The most interesting open thread is “decision traces” — the informal policy decisions humans make that never get hardened into written rules. Monroy described this as a very interesting area for Workday, given our position as the authoritative home for the org chart, pointing to decisions made from a people and organizational perspective, mapped to goals and financial outcomes — the kind of question buyers should ask in Barcelona.

The harness: deterministic guardrails, not prompt hope

Context gets an agent the right information. It doesn’t stop the agent from doing something unauthorized. Ahmad’s example is deliberately mundane: an expense-approval agent permitted to execute approvals up to $100, with anything above routed to a human-in-the-loop.

“We’ve already developed these systems in the past. We developed approval workflows. We developed identity model security in the human world. Now, the human world to agent world doesn’t translate one to one. But it’s a misnomer to assume you can just let agents run without these, without the governance, without the security.”

What this means for enterprise buyers

  • Ask for enforcement, not policies. A model that “knows” your approval policy is not a control. A deterministic workflow that blocks the transaction is. Workday’s architecture bet is that the latter is the only defensible position.
  • Semantic models are the real lock-in battleground.
  • Zero-copy ingestion changes the data warehouse conversation. Data Cloud is a major source for agentic context, but not the only one: policy documents and other external data can be injected separately.

What’s still unresolved

Two gaps remain. diginomica’s reporting framed the open question as whether Rising EMEA would deliver answers on cost, ROI, and systems of agentic control.

The practical takeaway for anyone evaluating agentic AI this quarter: the differentiator isn’t parameter counts.

Full context on the semantic model announcement is available in Workday’s own breakdown of Data Cloud’s layers, delivered at Rising.

Photo of author

Technology Desk

Technology Desk at Archyde.

NETS to Pilot Tap-to-Pay QR Payments in Toa Payoh with Nationwide Rollout Planned

Luca Brecel One Win Away from 2027 German Masters Main Stages After Beating Yao Pengcheng