How to implement generative AI for 400,000 employees

Executives conceptually understand GenAI, but they are the least likely to actually use it, according to anecdotal exercises performed by EY researchers. Three months into our journey, we unveiled a half-day hands-on exercise specifically to engage with this cohort of executives and elevate the “knowers” (100% of people) into “doers” (only 10%). We’ve gotten great feedback because participants stop merely talking about GenAI and actually start engaging.

Eventually this exercise evolved into an “AI master class” that we’ve also deployed to our clients. We talk about the technology and then leaders get their hands on the keyboard for something totally different than their day-to-day role: for example, creating new basketball teams with logos and brand names developed by publicly available GenAI tools, as well as using them to shape their marketing plans and predict sales projections. Client executives dive into the power to create, all while learning some of the leading practices.

In our AI survey, 73% of respondents revealed concerns that there won’t be sufficient training or upskilling in AI, and 63% expressed anxiety that they won’t have access to AI learning opportunities. Overall, 80% say they would be more comfortable using AI at work if they had more AI training and upskilling opportunities.

To address this anxiety around a lack of training in AI, EY launched a new EY Badges AI program in April 2023 offering certifications in responsible AI, applied AI and AI strategy. With a track record of investing in its employees, the organization had 14,000 people sign up for the EY Badges AI program in the first month. EY employees completed an all-time high of 24 million training hours across multiple learning program offerings in fiscal year 2023. On average, employees earn roughly 430 Badges each day for over 410,000 since the employee development program’s inception in 2017.

We also launched EY Fabric AI Ignite, a 30-minute weekly session led by engineering leaders to ignite our people’s curiosity, boost AI proficiency, drive demos and answer questions. Another recent hands-on webcast tackled “Prompt Engineering 101,” with two exercises in EYQ covering eight prompt-engineering techniques across four use cases.

Initially, we crowdsourced ideas for GenAI use cases to enable with EYQ — and had over 1,000 opportunities in just three weeks. We were left wondering: where should we start? We soon learned we weren’t the only ones running into this issue. In fact, one of our clients topped 2,500 different use cases in their crowdsourcing efforts.

As we combed through our own use cases, our clients’, and external perspectives from thought leaders, patterns began to emerge. We could group all the use cases into groups that relied on common technology building blocks. This would enable us to unlock the economic efficiencies of having a platform — we could build a “pattern” once and use this as a foundation for all other use cases that relied on the same technical principles. We have identified eight patterns in total, including multimodal systems and fine-tuned small language models (SLMs).

The first pattern we deployed is called retrieval-augmented generation (RAG) at scale. RAG is an approach that can reduce hallucinations and increase relevancy of responses by pointing the AI system to a specific corpus of documents or data sets (e.g., vector store). This is perhaps the most prevalent pattern we have seen across the market, and there are countless number of use cases across our organization that have been enabled as a result.

As a large global network of professional service firms that are highly regulated, the EY organization needed to take the responsible use of AI seriously. Whether we like it or not, employees of any company could use publicly available tools in ways that could be detrimental to their businesses.

EY leaders worked to create principles in how to use AI to drive fairness, accountability and reliability, tied to concrete ways of measuring each. Our platform uses a confidence index across its capabilities, maintains data centrally and stands up measured capabilities with enabled protections and a high degree of governance. We’re controlling focus, scaling and investment and keeping humans in the loop, so that fairness, accountability and reliability are applied with every use.

Over time, as a pattern of success is established and we gain confidence and momentum using the tool, we intend to move to more of an enablement model. The way brakes on a car are not just intended to slow it down but give the driver the confidence to go faster, this governance model can help EY safely accelerate into future AI opportunities.

The EY organization believes that those who embrace AI today can begin to build the future according to their companies’ values and ideals. While today’s AI-powered copilots and engineered prompts that augment human work offer powerful productivity enhancements, they still only scratch the surface of the technology’s true potential to transform.

Many enterprises will need to revisit their definition of AI initiatives to unlock value

AI is already transforming how EY teams operate and deliver for its clients, such as in third-party risk management (TPRM), where we are building AI tools to improve the quality of TPRM assessments. What will AI unlock tomorrow that we never imagined today? Will your tax professional become a bot, available for your questions 24/7? How will regulation impact the reality of this scenario? These are existential questions for us, and your business must grapple with the implications in your sector. For instance, in fast fashion, AI could scrape what’s buzzworthy from social media and automatically channel it into new designs. In a more mundane yet transformative use case, AI can also scan the risk and control landscape to assess gaps in a company’s current operations and show how they should evolve, as well as perform analytics and monitoring.