Workday Illuminate: The Future of Enterprise AI Is Here
At Workday, we believe that AI can fundamentally transform the enterprise. Here’s why.
At Workday, we believe that AI can fundamentally transform the enterprise. Here’s why.
In this article, we discuss:
The first phase of this generation of AI, which is characterized by the use of large language models (LLMs) in products such as ChatGPT, is ending. As it ends, the initial excitement and enthusiasm we have seen over the last two years is being replaced by healthy skepticism and key questions, especially in enterprise applications.
However, it remains true that the potential benefits of AI, if applied thoughtfully, could be truly staggering.
In this next phase, the approaches to AI between the consumer market and the enterprise will increasingly diverge. Organizations are subject to evolving regulatory requirements wherever they operate, other types of liability, and pragmatic concerns with return on investment, given the costs associated with LLM-based solutions. Enterprise AI must shift from attention-grabbing use cases to being able to demonstrate real value, and therefore customers need their vendors to thoughtfully address the challenges that LLM-based AI solutions deliver.
Inherent to LLM-based AI are several challenges that will need to be addressed to enable enterprises to move from proof-of-concept to widespread adoption of AI-based solutions. These challenges can include: 1) accuracy, 2) repeatability, 3) bias, 4) latency, 5) cost, and 6) memory.
Accuracy. LLMs excel at pattern recognition, leveraging vast datasets to identify correlations often imperceptible to humans. This makes them “popularity engines” instead of “accuracy engines.” An AI, such as one that provides the next word in autocomplete, will provide a popular answer, whether or not it is correct. While this might be tolerable in consumer markets where popularity often equates with accuracy, the enterprise landscape demands precision. Incorrect outputs can lead to legal and compliance risks, particularly in critical applications.
Repeatability. LLMs are inherently stochastic, yielding varied outputs even with identical inputs. This variability, while contributing to their human-like quality, poses challenges in enterprise scenarios where consistency is paramount. Applications in medicine, finance, and regulatory compliance can necessitate repeatable results.
Bias. LLMs mine the past to predict the future. They consume massive amounts of data, and that data contains both intended and unintended patterns. The fingerprints of past bias and discrimination are inherent in the data and, when left alone, an LLM trained on the data will simply perpetuate those biased patterns going forward. It is critical that enterprise vendors developing AI systems start by examining and understanding the training data, and use techniques to identify and ameliorate bias.
Latency. LLMs can take substantial time to generate results, and higher performance results can take longer. In many use cases this is not a deal breaker. However, as the technology is extended to perform actions and “think” more deeply using methods like Chain of Thought, multiple LLMs will need to be combined. Each action will then likely involve numerous back-and-forth exchanges between these LLMs. When that happens, each delay adds up, and the overall response time could become unacceptably long, depending on the requirements of the use case.
Cost. LLMs are massive software engines sitting on top of massive computing resources. They are breathtakingly expensive to train and use, both in hard costs and in impact on resources such as energy and the environment. As LLMs are enlarged and combined to produce better results, these costs increase rapidly.
With innovation as a core value, Workday is continuing to responsibly and intentionally build upon the Workday platform to support our investments in core applications, Workday Iluminate, Workday Extend, and user experience.
Workday’s approach always starts with the data. Workday and our customers have one of the largest, cleanest, and differentiated enterprise data sets on the planet. Encompassing a wide breadth of applications around the people and money in organizations, this data can unlock patterns that can be used to accelerate work, to assist completion of tasks, and even transform how enterprises operate. The data is also where Workday starts addressing right-to-use, regulatory, legal, and ethical issues. Workday has built an industry-unique machine learning development platform that builds in these concepts from the ground up.
The data is not just valuable because of its sheer volume. It is also valuable because of its variety. It represents more than 10,500 organizations that vary in size from tens to millions of employees, in dozens of industries and countries. When you slice a small data set to ask and answer specific questions, you can quickly run out of enough data to feed an LLM. Workday’s data set is so large and varied that it can be sliced while still retaining enough data to drive high-quality results.
This ability to use smaller data sets aligns with our model approach. In general, we use smaller models (although still staggeringly large when compared to the LLMs of just three years ago) derived (“distilled”) from larger models. We train each model with these smaller data sets, and add essential context, to produce results that are better than that which can be achieved by larger generalized models in terms of speed and accuracy, for far less time and money. With a smaller model we can also increase the context window, allowing the LLM to take on far more complex tasks. We can then combine these models—each an “expert” in a particular area such as expenses, payroll, or contracts—to allow us to solve complex problems that touch on multiple areas.
Context is also critical to how we leverage LLMs to solve problems. While high-quality data is essential, it is not sufficient to overcome the enterprise challenges highlighted above. You also need context to allow the LLM to disambiguate answers and drive to correct outcomes. Context needs to be provided in training, when fine-tuning, and also in active use through techniques such as advanced prompting and RAG (retrieval augmented generation).
So what do we mean by context? Context is a combination of both business context and user context. Business context includes taxonomies and ontologies, the language that is unique to a specific industry or area of expertise. We encounter these differences in everyday life, where subject matter experts in fields such as medicine and legal use language in a different way than that of the general public (which is what foundational LLMs are trained on). Business context can be more specific, to the specific organization and its terms, or standard documents, such as a canonical standard contract that a contract LLM should use when analyzing or creating a new contract. Given Workday’s deep data across different use cases and industries, we are able to supply relevant business context when the LLM needs it to improve performance across every dimension.
Often ignored in these approaches is user context. The correct answer the LLM should provide is not just dependent on business context, it also depends on who is asking and in what role they are in. Recognizing that the person asking is a subject matter expert versus a casual user is critical to providing a response that has value for each of them. Workday has this data for more than 70 million active users—who they are, their title, role, location, and so much more. This allows our AI to tailor its responses to increase relevance to the person asking.
Given this, our vision for the next generation of Workday AI, Workday Illuminate, is to move AI value from improving efficiency (i.e., generating an accurate job description) and toward full enterprise transformation.
Workday Illuminate isn’t just about speeding up your work. It’s about enabling a profound shift that redefines how businesses operate. Workday Illuminate will accelerate the user to make current work faster and smarter. They will write better and faster, find insights automatically, and automate tedious tasks–such as completing knowledge articles, contracts, anomaly detection, auto-filling, and more.
Workday Illuminate also assists the user by providing expertise in discovery and action. They can instantly find anything, get personalized help, navigate complex processes, stay informed, and automate approvals–just like having an assistant alongside them each day.
And while that is very valuable, the largest impacts of Workday Illuminate lie in its transformative power. For example, integrating the capabilities of HiredScore into Workday, we can reimagine an end-to-end hiring process—from intake to offer-—so that talent is found in less time and recruiter capacity is increased by 25%. We can also do this with processes like expenses, succession planning, accounting, and more. By fundamentally rethinking entire workflows, Workday Illuminate will enable organizations to achieve exponential value from their AI and thus more easily embrace its adoption. It’s not just doing things faster—it’s about doing things differently and better, and empowering people to achieve more, whether they’re a Workday expert or an occasional user.
Workday is not only continuing to build upon our Workday Illuminate vision, but we have opened up access and are allowing customers and partners to build on Workday to extend these capabilities themselves. Customers and partners can leverage our AI capabilities and build their own use cases and custom applications via Workday Extend with AI Gateway. We have over 400 AI-enabled applications built via Workday Extend to date.
With innovation as a core value, Workday is continuing to responsibly and intentionally build upon the Workday platform to support our investments in core applications, Workday Iluminate, Workday Extend, and user experience. We are actively fine tuning our AI models to ensure optimal performance and continuous evolution, and Workday’s unwavering commitment to innovation, particularly in AI, translates to a platform that is shaping the future of enterprise technology–lighting a path forever forward for our customers.
Gain more insights on how artificial intelligence and machine learning are driving the future of work. Learn more.
More Reading
Earlier this year, Professor Taha Yasseri was appointed the inaugural Workday Professor of Technology and Society at Trinity College Dublin and Technological University Dublin. In this episode of the Workday Podcast, we speak to Professor Yasseri about this new and exciting partnership and how it will help to bridge the gap between technology and society.
Workday is introducing enhancements across an array of products focused on hyper-automating processes for financial management, planning, and spend management while improving the speed and accuracy of data insights.
AI developers and the companies that deploy and use their technology must work together to ensure AI is leveraged for good. All parts of the AI value chain have a role to play in the responsible development and use of AI.