To Choose Is to Lose: The Future of AI Lies Not With One Vendor, but With Many

This article was originally published in Dutch by ITDaily. It has been translated into English below.

AI is a story of many vendors and models, Cognizant knows. In the critical environment of life sciences, the company is working to implement both classical and generative artificial intelligence. What works for the pharmaceutical industry can immediately be a useful guide for other organizations looking for a way to embrace AI in a future-proof way.

Dr. Pierre Marchand started his career in AI even before it was cool. "I've been working around data analytics and data science all my life, in the banking, manufacturing and chemical industries, among others," he says. "When you've been in the business for 30 years, you start to see the big picture." Six months ago, the expert made the jump to Cognizant, where today he is working as Chief Data Strategist.

Clear Vision in a Crucial Sector

That move was no accident. With the rise of generative AI, Marchand wanted to once again be at the center of the action. "Cognizant was already talking about Gen AI before the competition," he knows. "The company is taking a leadership role and embracing the right plan of action.”

We talk to Cognizant in the context of AI in the life sciences and pharmaceutical industries. That sector is home to enterprise-level organizations where business operations are laced with a scientific approach, and competition is keen. On the one hand, AI has enormous potential to accelerate development and production within pharmaceuticals; on the other, trust in AI and protection of sensitive data is almost nowhere more crucial. What is good enough for life sciences will almost certainly be good enough for other sectors, so we are curious to see what role AI plays there today.

Classical or Generative?

Marchand emphasizes that AI in life sciences and pharmaceuticals is nothing new. "Classical AI is already embraced by all major players within the industry, spread across many processes." Traditional or classical AI uses clearer rules, created by experts. Already today, the algorithms optimize all kinds of tasks, and have the great advantage of not being mysterious. The results produced are the result of understood programming.

The revolution and hype today concerns generative AI. It differs from classical AI because the intelligence does not come from distinct algorithms, but from a neural network trained from large amounts of data. During training, the network develops into a smart and useful AI model, but what happens under the hood is less transparent. Such neural networks have the potential to revolutionize processes, but have important limitations.

Data for the Black Box

"To train a generative AI model, you need huge amounts of qualitative data," Marchand says. "These are not always available. For example, there is limited data available on some rare diseases, so predictions lose accuracy. When data is scarce or unbalanced, the reliability of generative AI fails."

“When data is scarce or unbalanced, the reliability of generative AI fails.”

That reliability is another thorny issue. Marchand refers to the black box dilemma. The neural networks behind generative AI work, but it is often not clear why certain inputs lead to specific outputs. "That lack of transparency hinders acceptance of AI-generated results," he knows. "Especially in the scientific world, where decisions have an impact on people's lives."

From Suspicion to Trust

"Without integrity, there is no trust," the AI guru knows. "No one wants to hire a corrupt accountant. Integrity comes from traceability. As a human being, you want to be able to trace where an answer comes from. Specific to AI, it helps if you can verify why a model gives a particular answer. That creates trust."

"In most cases, we see that people completely check answers once or twice and then trust slowly appears. No one wants to risk their job by just assuming a prediction from AI, but when you get the chance to test and investigate yourself, however, confidence does start to grow."

Test First, Then Use

In the pharmaceutical industry, no matter how good a suggestion is, it is not possible to just accept it anyway. "When generative AI suggests a new molecule for a particular treatment, it is almost like suggesting a new musical note that has never been heard before," Marchand illustrates. "The suggestion is exciting, but still needs rigorous testing to ensure both safety and efficacy. The road from AI insights to useful applications is paved with extensive validation, and that brings down the speed of implementations."

“The road from AI insights to actionable applications is paved with extensive validation.”

Yet generative AI is already playing a role in the industry today. "Atomwise uses generative AI to predict molecular activity," Marchand knows. "The company is embracing deep learning models to potentially discover drugs. Insilico Medicine is doing something similar. Tempus deploys generative AI to suggest tailored treatments based on patient data. Deep Genomics attempts with the help of deep learning models to predict how genetic mutations can lead to disease.

The possibilities are virtually limitless. In time, Cognizant expects generative AI to be deployed throughout the pharmaceutical industry's value chain. The technology can assist experts in the discovery of new active molecules and other research and development, the preparation and processing of clinical trials, as well as more classic areas such as content creation and sales and marketing of finished drugs. However, the ultimate goal remains clear: to achieve ever better results in patients.

Papers with a Side of Fruit Juice

Cognizant is working with those and other companies to integrate AI in a structurally sound way. That requires a thoughtful approach, because the field is evolving rapidly. Marchand: "To keep up, you have to devour scientific papers for breakfast. A tsunami of progress takes place every week. I have never seen such a pace of development in my entire career.

Marchand and his colleagues take on that sufferer role (pun intended) so they can assist customers. In practice, he sees most organizations begin their generative AI journey with what he calls employee GPT. An internal version of GPT-4 or a similar model is fed a bunch of data and can now answer questions and assist employees. "That's good as a proof of concept, but not where the real value is," warns the expert.

Strong Together

"The real value comes when you can use AI to figure out better internal processes." To do that, you need to approach the rapidly evolving AI field correctly. "AI is not a one vendor story. You used to be able to invest based on Gartner's magic quadrant. You picked right up what fit within your budget, and waited for your promotion. You can't do that with generative AI, because the playing field changes every day."

The solution, according to Marchand and Cognizant, lies in a broader agile approach. "For optimal results, you don't need one AI model, but several," he knows. "For a chatbot, for example, you can start the conversation with OpenAI's technology. Then, when the conversation evolves toward a medical topic, MedPalm is better suited."

Different models have different things to contribute. "Orchestration is important. AI works best within a model where different models can work together as a whole. For that, though, the models need to be able to call on a shared memory: you don't want to re-explain your problem when you're sent from OpenAI to MedPalm."

Chasing the Brain

Cognizant therefore built Neuro R AI. That's a framework that is model-agnostic. It allows companies to combine models into one, but also to swap them out for new and better versions. The framework ensures that the most appropriate model is deployed in the right context, while a combination of long- and short-term memory ensures consistency. When Marchand explains it, we are reminded of our own brain, and how it is made up of specialized parts that still work together as a whole.

"Moreover, Neuro AI is not limited to just generative AI," Marchand notes. "The framework can perfectly combine new generative AI with classical AI and analytics." For Marchand, Neuro AI was one of the reasons for moving to Cognizant. The field is evolving at lightning speed, but he sees the agnostic orchestration approach as the right one based on his experience.

We are living through turbulent times. Generative AI is shaking up the world and is going to have a major impact in just about all sectors, not just life sciences. However, everyone is facing similar challenges. The almost untraceable evolution of AI right now is an important one, as is the issue of trust. There are still some open questions, but for the implementation of AI, Marchand and Cognizant already see a clear solution. Multiple models, algorithms and types of AI need to work together for the best results, and that requires an agnostic framework.