As enterprises rush to add AI to legacy workflows, Workday takes a different view: intelligence must be engineered into the architecture itself, not layered on top. In this conversation, Indira Vidyaprakash, Vice President of Software Development Engineering at Workday, explains what “AI-native” means at a platform level, from redesigning product teams around agents, skills, and tools to embedding governance and auditability as core design disciplines rather than afterthoughts. Drawing on Workday’s research into the “copy/paste economy,” she explains why fragmented systems, not a lack of AI, are the real productivity drain and how Workday’s Anticipate, Advise, Act model determines when agents should act autonomously versus defer to human judgment on high-stakes decisions.
Vice President of Software Development Engineering
Workday
CIO&Leader: Workday talks about moving “beyond copilots” to AI-native platforms. In concrete terms, what does that shift actually change in how your product teams design and build software? Is it a different starting point, a different team structure, or a different definition of “done”?
Indira Vidyaprakash: It changes the starting point: The platform has to be agent-native from day one — not just for one interface, but for every agent in the portfolio. That means agent access, APIs, and business context get engineered as design principles into each domain up front. Workday agents are embedded directly into products and business processes, so they can leverage deep business context to surface more accurate insights, make better organisation-relevant decisions, and take higher-impact autonomous actions. Building this way works only because Workday agents can have both deep and broad access to the data and context of work—who can do the work, what strategy should be used, and how it will impact the business.
That also changes team structure. Once you have many agents rather than one, product teams have to design at three distinct levels, not just UX and data model: the agent as the overall digital worker, the skill as a high-level functional grouping (e.g., Time Off, Payroll), and the tool as the specific Workday task or API a skill can call. Sana remains the conversational interface people talk to. Still, the structural change is that governance, auditability, and lifecycle management now sit alongside UX and data modelling as core design disciplines for every agent a team ships.
Finally, it changes the definition of done. A process shouldn’t just record an outcome or move information along; it should help deliver the business outcome, with AI understanding context and taking action where appropriate. That’s what our Anticipate, Advise, Act model captures: agents that pick up work before anyone files a request, propose the right action with the policy and reasoning already applied, and execute inside real approval chains when it’s time to act, with every step logged. Unlike standalone assistants or copilots that mostly answer questions, Workday agents operate as teammates across that full arc, with governance built in, not bolted on. That’s the real move beyond copilots.
CIO&Leader: Many enterprises added AI as a layer on top of existing workflows. What did Workday learn, perhaps the hard way, that convinced you intelligence needs to be embedded at the architecture level rather than bolted on?
Indira Vidyaprakash: Our India research, “The copy/paste economy,” found employees are genuinely positive about work and about AI. 89% say AI has already improved their day-to-day work. But 89% also spend significant time coordinating or translating work between systems; 86% spend time moving information between them, and more than a third lose over seven hours a week to it. Only 32% say AI is actually embedded in their core systems.
Many agents and copilots live in disconnected systems, cut off from the data, context, and business rules they’re supposed to help run. This relegates them to doing surface-level tasks, rather than meaningful work. Employees have become the glue holding disconnected systems together, and AI sitting on top of that fragmentation becomes one more thing to coordinate.
That’s the architecture lesson: intelligence has to sit inside the systems where work happens, not around them. That’s why Sana operates inside Workday’s existing security, permissions, and configuration rather than as a separate tool to log into, making it one connected system. Fewer handoffs, less rework, more work completed in the flow itself. That’s what makes AI operational rather than peripheral.
CIO&Leader: You’ve spoken about AI moving from recommending actions to executing them. What guardrails or checkpoints does Workday build in before an AI agent is allowed to act on something like payroll, compensation, or a hiring decision, rather than flag it for a human?
Indira Vidyaprakash: AI shouldn’t have the same autonomy across every type of work. The level of automation has to reflect the decision’s potential impact. This is really our Anticipate, Advise, Act model in practice: agents earn the right to act as the stakes go down, and stay closer to advising as the stakes go up because Workday’s business processes are deterministic by design: they have a clear start, a clear end, and an outcome you can audit.
That’s not just a principle; it’s an operational checkpoint. We assess use cases by how closely they touch decisions affecting people and apply a tiered control framework accordingly. For routine operational work, agents can do much more: Workday’s Payroll Agent continuously scans for issues, flags what’s missing, and guides administrators to a fix before a run fails. That’s anticipating the problem and advising the fix before it ever executes inside an approval chain. That’s agentic execution on a deterministic, rules-bound process, the kind of lawful agent behaviour we design for from day one.
Sensitive decisions are handled differently by design. We wouldn’t automate a hiring manager’s call based purely on an algorithmic outcome; our AI supports the hiring process; it doesn’t make hiring decisions or reject candidates. For compensation, the Compensation Agent gives managers agent-guided workflows and budgeting guidance grounded in company policy, with built-in compliance checks, but the manager makes the call. Every action any agent takes is permission-aware and sits inside the customer’s existing compliance, security, and audit controls. The goal isn’t to limit AI. It’s to give it the right autonomy while keeping human judgment where the consequences are significant.
CIO&Leader: Understanding “business context” is often used loosely. How does Workday’s AI actually learn a specific customer’s context, their org structure, policies, workflows, without that becoming a data privacy or accuracy risk?
Indira Vidyaprakash: Business context isn’t about giving AI more data access — it’s about the agent understanding what that data means: goals, constraints, policy, workflow history, what people have already done. The mechanism behind that is what Workday calls its “World Model of Work”: application surfaces are built to capture and surface insight as work happens, fusing decades of transaction history, labour market signals, and the fingerprint of each company’s process design into context no generic model has access to. That’s how it learns a specific customer’s org structure and policy, not by reading more of the customer’s data, but by being built into the systems where the org structure and policy already live.
Sana is the clearest expression of this. It’s deeply integrated with Workday’s people and financial data. It runs inside Workday’s existing security, permissions and audit framework, so it inherits the same access boundaries a human user would have rather than reasoning over some separate, ungoverned copy of the data.
Sana Enterprise extends the same principle outward, letting agents act across other connected applications while carrying that same governance model. The goal is to give AI enough context to act appropriately, without creating a shortcut around the controls the customer has already put in place.
CIO&Leader: How is natural language interaction changing what an HCM or ERP interface even looks like? Is Workday designing toward a future with fewer dashboards and more conversation, or do you see both coexisting in the long term?
Indira Vidyaprakash: Natural language turns the interface from a place people navigate into a place they can ask, decide and act. Sana is built as that conversational, proactive front door—the one interface people talk to inside Workday —that routes work to the right agent and gets it done, without anyone needing to know which agent handles what. But not every action belongs in conversation. Reviewing information, approving something, handling an exception: some of that is still better served by structured screens and dashboards.
The direction is coexistence, not replacement, with context and permissions carrying over as people move between the two. We see the same pattern in learning: Workday Learning, powered by Sana, answers learner questions in natural language and adjusts to each person’s context, rather than having someone search a course catalogue themselves. Dashboards and course catalogues don’t disappear; they stop being the only door in.
CIO&Leader: What’s been the biggest surprise, good or bad, in how employees actually use conversational AI features in production, versus how your teams expected them to be used?
Indira Vidyaprakash: We expected conversational AI to speed up individual tasks mostly. What we’re seeing is people pointing it at the coordination burden itself, exactly the gap our copy/paste economy research surfaced, where employees spend most of their time glueing disconnected systems together rather than creating value.
When we asked how employees actually use AI agents in their core systems, the top uses weren’t simple Q&A: 55% for monitoring metrics and suggesting actions, 49% for onboarding, 44% for routing approvals, 42% for budgeting and forecasting. These are workflow use cases, not task use cases. It’s not just the conversation that matters. It’s whether the AI can connect the right information and move the work forward, which is exactly why we built Sana to orchestrate a process rather than answer one question and stop.
CIO&Leader: Sana is central to Workday’s story of bringing data, apps, and workflows together into a single experience. From an engineering standpoint, what’s the hardest part of making previously siloed systems feel seamless to the end user: data integration, permissions, or something else?
Indira Vidyaprakash: Data integration and permissions matter, but neither is sufficient alone. An agent also needs to understand the process a piece of data belongs to and what should happen next. Sana handles this as the orchestration layer, connecting Workday with 100+ enterprise applications and routing every action through the Agent System of Record so security and auditability stay unified even across multiple agents and systems.
The genuinely hard part isn’t the API connection; that’s largely solved. It’s making sure context, permissions and audit trail travel with the agent when it hops from a Workday process into a third-party calendar or CRM and back. Our research shows that handoff is exactly where employees lose the most time today doing this manually, so closing that gap, not layering another AI tool on top of it, is the real engineering goal.
CIO&Leader: For a large enterprise with dozens of legacy systems, how realistic is “one seamless experience” in the next few years versus being an aspirational north star?
Indira Vidyaprakash: It’s realistic, but it doesn’t mean every legacy system disappears. It means the complexity underneath becomes far less visible to the person doing the work. Sana already connects to 100+ enterprise applications and can orchestrate existing agents and third-party tools with human-in-the-loop handovers, instead of leaving them siloed.
Our research gives a concrete proof point: in Asia Pacific and Japan, employees at organisations with AI deeply embedded in core systems are 1.8 times more likely to report AI cutting task time by 25% or more than those where AI still sits on the periphery. That gap is already showing up workflow by workflow: payroll, approvals, onboarding, even if it isn’t arriving as one single seamless system overnight. The pieces, from Sana for Workday to Sana Enterprise, are already real and in production today.
CIO&Leader: When AI touches people, pay, and business-critical decisions, the cost of a wrong answer is much higher than a chatbot giving bad advice. How are governance, auditability, explainability, and bias checks actually built into the development lifecycle rather than added as a compliance layer afterwards?
Indira Vidyaprakash: Trust has to be designed in from the start- what we mean when we say an agent is “lawful by design.” We risk-rate AI use cases by how closely they touch decisions affecting people, and that scales the safeguards applied. Every action a Sana agent takes is permission-aware and runs inside Workday’s existing security and audit framework rather than a separate access model. We’ve strengthened logging and traceability specifically to support audits and give humans a real ability to contest an output, not just a theoretical one.
Technical feasibility doesn’t automatically mean something should be automated. For sensitive decisions, human oversight stays essential, and we continue to hold ourselves accountable as industry scrutiny of AI in hiring and workforce decisions grows. The goal is for governance, explainability, and bias mitigation to be present when a system is designed — built in, not bolted on after it ships.
CIO&Leader: As Site Leader for Workday’s Chennai GCC, how is this India-based team specifically contributing to the AI-native shift, driving core platform capabilities, or more localised innovation for the India/APAC market? And what does building trustworthy, high-stakes enterprise AI demand differently from your engineering talent here compared to a typical product engineering role?
Indira Vidyaprakash: Workday’s product and technology hub in Chennai drives AI innovation across the full stack: development tooling, product engineering, and the infrastructure underneath it all. The team builds core platform capabilities that power and scale next-generation agentic solutions for customers globally, including the APIs and tooling that give agents governed access to Workday’s data, the infrastructure that scales as workloads shift from answering questions to running always-on, multi-step workflows, and the context and permissions plumbing that lets something like Sana reason reliably without ever stepping outside a customer’s access boundaries.
That work has a natural pull toward India and APAC specifically, where our own research shows both the opportunity and the urgency: employees are highly engaged with AI, yet only 32% say it’s deeply embedded in their core systems, and more than half spend at least half their time coordinating across disconnected tools. Being close to that reality sharpens our urgency to build AI into systems, not layer it on top.
The engineering bar is genuinely different from typical product work. A typical product role rewards someone for shipping a feature that works. Our teams in Chennai raise that bar with a forward-thinking mindset: anticipating where a workflow is heading, and the judgment to know when a system should act versus defer to a human. That judgment has to sit alongside comfort working with legal and compliance from day one, given the regulatory and audit demands in HR, finance and IT. That combination sets the standard we hold ourselves to in Chennai: AI capable enough to take meaningful action, and dependable enough to earn the trust that enterprise work demands, at the scale of the 11,500+ organisations already running on Workday.