During the Frankfurt Roadshow of the Dataiku 2023 Everyday AI Conferences, Philipp Wendland, Senior Data Scientist at Deloitte, opened his session examining a survey of executive leaders across global companies — a survey that included a remarkable statistic.
94% of global business leaders surveyed see AI to be critical for an organization’s overall success over the next five years.
Implementation of Generative AI in larger business organizations served as the foundational question of Wendland’s session. After businesses and industries cross the first major hurdle of recognizing the explosive growth of AI, what are the real-world applications in an organization, as well as the challenges and strategies that emerge?
Generative AI's History Is Still Being Written
Wendland then gave a brief history of AI, pulling a quote from John McCarthy, renowned computer scientist and one of the founders of the discipline of AI. “AI is the science and engineering of making intelligent machines,” Wedland began. Soon after, “we were able to extend neural networks into deep learning.” The acceleration of the AI industry has been almost exponential, and Wendland is quick to point this out, but notes that as quickly as Generative AI entered the collective conscience, questions and challenges about its implementation emerged. “Two percent of enterprises in Germany are currently using Generative AI. Only 13% of organizations actually have a plan to use Generative AI in their businesses.”
Wendland believes this is the next big milestone, shifting from understanding AI to implementing it at the enterprise level. Though growth in the AI space has skyrocketed, global industries are still largely in the exploratory phase, trying to map out and understand the impact that AI has on internal processes and how it can be leveraged for long-term growth and productivity.
Just as guests were beginning to see the driving force behind his talk, Wendland dropped another statistic that caused his audience to take notice.
80% of jobs in some way or another will see and feel the impact of Generative AI.
While people have been interacting with AI through technologies like OpenAI’s ChatGPT or Google’s Gemini, these types of interactions are relatively surface-level, Wendland notes. “Enterprises need to focus on everything below the tip of the iceberg. What are the business cases, the legal implications, or the risk and regulatory implications on creating a scalable and robust framework to really implement these solutions?”
The foundational models that Generative AI is built on are just that–foundational. Today’s models are getting faster and increasingly better at not just understanding code, as an example, but also writing it. Some models are being trained to do even more. But even with the tasks that AI can accomplish, from creating images or videos from prompts, or recreating very convincing audio of human voices from small samples, Wendland recognizes that there are limitations that need to be understood.
Understanding AI's Limitations
He points to the example of large language models (or LLMs) that are capable of understanding or generating text. “LLMs are trained on a lot of data, requiring a lot of GPUs and a lot of time, meaning that there is a cutoff date for a model.” Wendland points out that models are usually trained on a given dataset and infrequently retrained, resulting in a model that doesn’t have access to new information, lacking context in many cases.
This lack of information and breakdown of the models reveals itself in anomalous ways–ways that large organizations and enterprises should be wary of as they implement AI with large, shifting datasets. “We’ve heard of the term “hallucinations,” where models generate factually wrong statements. Of course, this is being worked on so it happens less and less frequently, but LLMs produce text based on probabilities based on what it has read,” Wendland explains. Naturally 100% accuracy is incredibly difficult to achieve, but he notes that this is because the models aren’t capable of humanlike reasoning.
He also mentions a very important point about biases that might emerge from the insights that LLMs produce. “We are all biased, subconsciously or not. In some applications this might not make much of a difference, but it’s important to consider that these models can have a bias and this should always be taken into account.”
...and Working Through Challenges
As enterprises look to include AI into not only their processes, but also their business planning, Wendland asks organizations to be mindful of the initial challenges in doing so. Among these major challenges facing organizations looking to start or augment their ML strategy is the cost benefit in doing so. The bigger and more complex models get, the more compute they need — the more compute they need, the more expensive they become.
In addition to the cost of the model itself, enterprises also need to consider which vendors to do business with, and the deployment and maintenance costs. In addition, there are deeper questions that Wendland poses. “In which pipelines will we deploy ML? What are the legal implications? Who owns the output?” Answering these questions and accounting for these challenges before undertaking any enterprise-grade AI exercise helps organizations succeed in the long term.
Wendland then pointed to what he called “three major plays” where Generative AI helps to transform businesses.
Efficiency
“Generative AI can help realize huge efficiency gains in your organization,” Wendland says. Not only can well-implemented AI make lengthy or complicated processes and workflows more efficient, it can make everyday tasks faster, too. Dataiku offers solutions like enterprise-grade prompt engineering to operationalize data across organizations to optimize efficiency.
Experience
Here Wendland speaks to the concept of personalization. “It’s now easy to hyper-personalize content on specific customer segments, or specific purposes.” As an example, he cites the task of writing product descriptions before the advent of Generative AI. This was a very tedious, manual task, requiring writing that accurately matched the vision and style of the company. With a generative system being fed thousands of descriptions in a certain style, it can generate very accurate, brand-aligned languages for multiple customer segments. “That wasn’t feasible before,” he says.
Capability
“We have technology that can read large amounts of text, summarize it, and do this quicker than any human can,” Wendland says. The value that AI provides is that it can now use these massive datasets and generate insights that are not only usable for developers or data scientists, but other teams as well. “Business and management can access data using Generative AI solutions,” he says. This makes insights much easier to understand and act upon, accelerating the overall speed and optimizing the efficiency of the organization. Dataiku provides data visualization solutions to help non-technical team members leverage AI.
Deploying Enterprise-Grade AI Solution Takes a Village
“It makes sense to include everyone to really stand up across disciplinary teams to really leverage the benefits of Generative AI,” Wendland continues. For an organization to be successful with Generative AI, all parts of the organization should be aligned with not only its implementation, but also its usage and goals. Perhaps even more critical is the concept of continuous improvement.
As an organization becomes more adept at using Generative AI, the insights it provides for the organization grow even more useful. As the insights become more useful, the organization performs better, and the cycle continues. To learn more about how models can continually improve over time, see Dataiku’s solutions for validation and evaluation of models.
Regarding the relatively near future, Wendland says, “The workforce will change. The way our everyday lives will change.” Different, changing, new skills will be required in an era when Generative AI is seamlessly integrated into everyday life. It’s up to organizations and enterprises of all sizes to understand and respond to the challenges of AI deployment so they can reap the benefits of a well-integrated AI system that is truly enterprise-grade.