Workday has launched a dedicated AI research team focused on solving the hardest technical problems that enterprises face when deploying AI agents at scale, including how to build selective memory systems, coordinate multiple agents for better decisions, and ensure that deleted information actually stays deleted. The new Workday AI Research division is publishing peer-reviewed findings at top machine learning conferences and introducing a PhD fellowship program to deepen collaboration with academic researchers.

What Are the Real Challenges Holding Back Enterprise AI?

Enterprise AI is fundamentally different from the consumer AI that makes headlines. When AI agents handle HR decisions, financial transactions, or IT operations, the stakes are much higher. Organizations need systems that can explain their reasoning, follow strict compliance rules, and protect sensitive data. Off-the-shelf AI models, no matter how powerful, often fall short in these areas.

Workday researchers have identified three critical problems that enterprises must solve to deploy AI agents responsibly. The first is selective memory: AI agents need to remember context and history, but storing everything creates privacy risks, slows down systems, and clutters decision-making with outdated or duplicate information. The second is multi-agent coordination: can multiple specialized AI agents working together make better decisions than a single agent? The third is secure deletion: when an organization asks an AI agent to forget something, does it actually forget, or does the information linger in hidden copies ?

How Can AI Agents Learn to Remember Only What Matters?

Workday researchers developed a more selective approach to agent memory that filters out outdated, duplicate, or unreliable details while keeping information that actually matters. In testing, the method delivered 12% higher precision and approximately 8% better overall memory quality, while retaining 97% of the memories that mattered. The system also ran about 31% faster than the leading AI-driven comparison, showing that better memory can be both more useful and more efficient.

This finding challenges a common assumption in AI development: that bigger memory is always better. Instead, Workday's research suggests that smarter filtering and selective retention can improve both accuracy and speed. For enterprises managing millions of employee records or financial transactions, this efficiency gain translates directly to lower computing costs and faster response times.

Can Multiple AI Agents Make Better Decisions Than One?

Workday researchers tested whether splitting complex work among specialized agents could improve both answer quality and compliance with business rules. In the study, one agent explored different options, another focused on following compliance rules, and a coordinating agent directed their work. The approach improved accuracy by 5.8%, and every final answer met the study's defined constraints.

This finding is significant because it suggests that organizations do not have to choose between higher-quality AI recommendations and strong guardrails. By dividing labor among agents with different specializations, enterprises can achieve both better answers and stronger compliance, which is essential in regulated industries like finance and healthcare.

Steps to Implement Trustworthy Enterprise AI Systems

• Selective Memory Design: Deploy AI agents with filtering mechanisms that retain useful context while removing outdated or duplicate information, improving both accuracy and system speed.
• Multi-Agent Orchestration: Structure complex AI tasks by assigning different agents to explore options, enforce rules, and coordinate decisions, rather than relying on a single agent.
• Secure Data Deletion Protocols: Ensure that deletion requests remove information from all copies, including summaries and historical records, not just the original data store.
• Explainability and Auditability: Build systems that can explain their recommendations and provide clear audit trails for compliance and accountability.
• Adaptive Resource Control: Implement systems that adjust computing resources based on task complexity, improving efficiency without sacrificing accuracy.

Does Asking an AI Agent to Forget Something Actually Work?

One of Workday's most surprising findings concerns data deletion. Researchers discovered that deleting information from an AI agent's memory does not always remove it for good. When researchers asked an agent to forget something, a copy of it was still recoverable from an old summary about one in five times. Fully erasing the information meant also deleting every summary that mentioned it, showing that "forgetting" in AI agents requires clearing every copy, not just the original record.

This finding has major implications for privacy and compliance. Under regulations like the General Data Protection Regulation (GDPR), organizations must be able to delete personal data when requested. Workday's research reveals that AI systems require careful architectural design to ensure that deletion requests actually remove all traces of information, including indirect references in summaries and historical records.

How Is Workday Supporting the Next Generation of AI Researchers?

To fuel this research and deepen collaboration with the academic community, Workday is introducing the Workday AI Research PhD Fellowship. The program supports exceptional doctoral students working at the intersection of AI and enterprise software. Fellows receive $50,000 in annual research funding through an unrestricted gift to their university, dedicated mentorship and direct collaboration with a Workday AI researcher, and early access to relevant career opportunities at Workday.

"As AI agents evolve to remember context and take action on behalf of employees, enterprises are facing complex challenges around privacy, auditability, efficiency, and enterprise-grade accuracy that off-the-shelf models simply cannot solve," said Gerrit Kazmaier, president of product and technology at Workday. "Workday AI Research is dedicated to solving these exact problems, delivering the rigorous science needed to build intelligent, reliable systems that organizations can actually trust and deploy at scale."

Gerrit Kazmaier, President of Product and Technology at Workday

Workday's research has already been accepted by top conferences including the International Conference on Machine Learning, the International Conference on Learning Representations, the ACM Web Conference, and the Association for Computational Linguistics. This peer-review process ensures that the findings are rigorous and reproducible, not just marketing claims.

The launch of Workday AI Research signals a broader shift in how enterprises approach AI development. Rather than adopting generic AI models and hoping they work, leading companies are investing in specialized research to solve their unique technical challenges. For organizations considering AI deployment in critical business functions, Workday's findings offer a roadmap for building systems that are not just powerful, but trustworthy, efficient, and compliant with regulatory requirements.

" }