5 tips for choosing the right AI model for your business

The rise of generative AI (gen AI) has been largely driven by high-profile large language models (LLMs) like Open AI's GPT-4o, Google's Gemini, and Anthropic's Claude. However, amidst the buzz surrounding these giants, small language models (SLMs) are quietly gaining ground. Some industry experts suggest that SLMs could be the future of gen AI.

According to research firm Gartner, while LLMs have traditionally led the charge in language model development, SLMs present viable solutions to critical challenges such as budget constraints, data protection, privacy concerns, and risk mitigation. As businesses navigate the world of gen AI, leaders may find themselves choosing between these two types of models. Here's what five business leaders have to say about the future of AI models.

1. Consider domain-specific opportunities

Claire Thompson, group chief data and analytics officer at financial services giant L&G, believes that both small and large models will find their place in business operations. She envisions a future where LLMs could be fine-tuned for specific applications, enhancing their utility.

"I can see a situation where some of the LLMs could start to be trained on specific topics to get more detail out of them, and I can see that beginning to happen more and more," she told ZDNET.

While there's a clear demand for domain-specific models, Thompson is skeptical about companies dedicating significant resources to in-house development.

"I don't know whether you'd build your own," she said. "When I talk about building models, it's more about leveraging existing models internally and using your data in a secure environment to achieve results."

Regardless of size, Thompson sees the future leaning towards domain-specific models.

"I think we will start to get more tailored models," she said. "You could see, for example, how you might tailor a model around medical information, climate topics and ESG, and asset markets. It's those specific use cases where you could get more bespoke models coming out."

2. Pick the right horse for the course

Nick Woods, CIO at MAG Airports Group, emphasizes that the future of gen AI likely involves a mix of large and small models, tailored to specific business needs.

"I don't think it's one size fits all," he said. "And I think the model you select depends on the use case in your business."

Woods warns against jumping into AI development without a clear strategy.

"No, it's the last thing we should do," he said when asked about launching an AI program.

Instead, he believes executives should focus on their broader business transformation goals and identify the tools, including gen AI, that can help achieve them.

"So, for example, we may want to run a small, specific model on the edge to go and solve a particular use case around something like spotting when an air bridge has docked," he explained.

For broader, more complex questions like global air traffic patterns and weather impacts, different models might be necessary.

In essence, Woods believes choosing the right model is akin to picking the right horse for the course.

"I think you will see many small models deployed at the edge at scale for particular use cases," he said. "That's almost inevitable. However, I still think you'll see some big models prevailing."

3. Consider the context

Gabriela Vogel, senior director analyst in the Executive Leadership of Digital Business practice at Gartner, highlights the growing interest in small, domain-specific models among CIOs.

"The clients I speak with are trying to find and create models applied to a specific context," she said. "They're not necessarily big, general models, but ones specifically tied to small databases for a particular application."

Vogel notes that companies are increasingly shifting from exploratory phases to deploying gen AI services using SLMs.

"They're making this shift because they've tested a lot," she said. "They've seen what works and doesn't with bigger models, and then they're trying to go more specific and apply that approach. That's what I've personally seen with my clients."

4. Go small to reduce hallucinations

Ollie Wildeman, who leads customer satisfaction at Big Bus Tours, believes the choice between SLMs and LLMs hinges on the specific use case, with many companies likely opting for smaller models.

Big Bus Tours leverages Freshworks Customer Service Suite, which includes AI-powered chatbots and ticketing, and a virtual assistant from Satisfi Labs that handles basic customer queries.

"Satisfi's AI technology only takes data from the specific companies they work with," Wildeman explained. "The company's technology is not connected to large-scale AIs, like ChatGPT or other tools -- they're doing it themselves."

This focused approach, Wildeman believes, offers significant business benefits, ensuring data is used responsibly and reducing the risk of AI "hallucinations."

"In that way, your data is safer because you know where it's coming from and what processes they're using," he said. "Also, you get fewer hallucinations because you know the model you're using is designed for the type of business you're in."

Wildeman concludes that smaller, domain-specific models will play a crucial role in enterprise AI strategies.

"I think for businesses, the choice of model is going to be more specific, whereas probably for the general user, these massive free models that you see everywhere will be popular."

5. Focus on your first-party data

Rahul Todkar, head of data and AI at Tripadvisor, believes the ideal model for a company isn't solely about size but about customization.

Professionals may experiment with both large and small models, but Todkar sees the future in purpose-built and customized models.

"Take the example of Mistral 7B, which is a relatively small model in the context of other LLMs, but it does fantastically well when you look at specific tasks," he said. "So, to me, the future is about customizable models."

Todkar emphasizes the importance of leveraging a company's first-party data effectively.

"It's not the training size or the features in the model that matter, but rather it's about taking that model and applying it in your context with your first-party data," he said. "That's when you move beyond off-the-shelf models and can use the insights from your data. So, the answer is going to be somewhere in the middle."