Workday Scales AI Agents with Iceberg and Unity Catalog

A trust crisis waiting to happen

Workday faced a structural problem as it deployed AI agents across finance, HR, procurement, and supply chain functions: each agent with its own data store created exponential complexity. With 100 agents and 100 data systems, the architecture became what Phoenix Majumder, Senior Director of AI Engineering and Platforms at Workday, describes as "very complex spaghetti" — impossible to govern or monitor reliably.

The risk was existential. A single agent delivering incorrect information could collapse trust in the entire AI program, much like one faulty dashboard can undermine confidence in all business intelligence outputs. For an enterprise serving more than 10,000 organizations worldwide, that was unacceptable.

Workday's response was architectural: convert the multiplicative M × N integration problem into an additive M + N model by building a universal data layer where agents come to consume data rather than forcing data replication outward.

Building on Iceberg and Unity Catalog

Workday's solution combines Apache Iceberg as the open table format with Databricks Unity Catalog as the unified governance layer. Iceberg provides ACID compliance and snapshot isolation, giving multiple agents consistent views of data while they reason and act. Critically, it also bridged cultural divides between data science teams focused on rapid experimentation and platform engineers requiring deterministic systems.

Unity Catalog serves as the centralized control plane, enforcing fine-grained access controls, capturing end-to-end lineage, and ensuring comprehensive auditing. This prevents scenarios where, for example, a sales executive in one territory could prompt-inject their way into viewing accounts they're not authorized to see — a compliance disaster with sensitive enterprise data.

The architecture stores long-term agent memory in the Iceberg layer as persistent state, while short-term session context lives in vector stores. Reasoning traces are captured for auditability, allowing teams to trace any agent decision back through its logic.

Majumder emphasizes the modular design philosophy: "We have to think like we are builders of a Lego toy. You take the parts and build a toy. But equally, if a part breaks down in future, you should be able to swap it very quickly."

Why it matters

As enterprises move from single-agent pilots to coordinated multi-agent systems, governance and trust become the bottleneck — not model capability. Workday's architecture demonstrates that scaling AI agents safely requires solving the data foundation problem first. Without centralized governance, each new agent multiplies integration complexity and creates new vectors for compliance failures. The shift from multiplicative to additive complexity isn't just elegant engineering; it's the difference between AI systems that scale and those that collapse under their own weight.

Measurable business outcomes

With the governed data layer operational, Workday has deployed agents that accelerate finance close cycles, identify attrition risk earlier in HR workflows, optimize supply chain contract negotiations, and improve sales planning accuracy. The company progressed from single-agent experiments to coordinated multi-agent deployments without accumulating technical debt.

The unified approach eliminated redundant data pipelines, and the modular design enables rapid model swaps and cost controls at natural choke points. New agents onboard quickly because foundational data and governance infrastructure already exists.

Workday characterizes its progress as early-stage — "probably at a 3" on a scale where 10 represents full maturity. Near-term priorities include expanding interoperability as data sources proliferate, establishing registries and communication protocols for agent-to-agent interaction, and deepening governance automation to monitor and steer agents as "digital employees" whose reasoning can be analyzed and improved.

These details were first reported by Databricks in a customer case study on Workday's Unity Catalog implementation.