Choosing the Right AI Model for Your Business

Enterprises are captivated by the potential of large language models and generative AI. But how should organizations go about implementing their AI strategy? Should they go with large models, or do smaller models work better?

Speaking at the 7th Chief Digital and Data Officer (CDDO) Asia Summit, Carlos Queiroz, managing director and global head of Data Science Engineering at Standard Chartered Bank, noted that there is no right or wrong answer. Instead, he suggests that the use case should drive the decision on the AI models to use.

Go big or go small?

Queiroz noted that there are “obvious” solutions in some cases, such as having small models deployed for specific niche tasks. The use case will ultimately determine if a solution is a good fit, however. He also cautioned against using too many disparate models—either large or small—as it can require substantial upkeep.

“This might mean we are creating another arm of technology in order to support [multiple models]. If that makes sense for your use cases, then yes, go ahead. If not, then you need to think of a different approach,” he said.

“There are use cases which are using large-scale processing of data, which requires significant compute or development. I would rather [use] what is already available because the amount of investment I have to put in to fine-tune a [smaller model] would be humongous,” said Sudhanshu Duggal, CIO, digital leader at Forbes Technology Council.

Duggal offered a different take on the debate, noting that reusing what is already in production as part of the organization’s AI Factory will make more sense than attempting to develop AI models new from scratch. He also pointed out that smaller models designed for niche use cases might be difficult to scale for broader tasks.

Mind your data

When it comes to data, Queiroz suggested that businesses avoid proprietary models and go with open-source ones to ensure control. The reason is simple, says Queiroz. While organizations can and should have contracts to protect their interests, there is no way that a data leak can be rectified.

“Once the data is out the door, you lose control over it. I think you should be very concerned about this. I would not consider using proprietary models for businesses creating intellectual property or products that give you a competitive advantage. I would recommend the use of open-source models where you have full control,” he said.

“You want control over the model, control of the data that has been used to train the model, control of the answers that the model gives. And data is at the core of the enterprise. If you’re using a model that allows your data to go outside your control, I'd say you should be careful because you don't know where the data is going.”

However, Queiroz concedes that some financial institutions have weighed the risks and have decided they are fine with it. “The enterprise needs to decide what's best for them, and based on their use cases, decide what they want to adopt,” he summed up.

Start with the use case

So how should enterprises get started with AI? Both Queiroz and Duggal were unequivocal that everything starts with the use case.

“In this day and age, the most important thing for you is to start with the use case and I think I'm coming back to this, but it's important that you look at what use case you're going after,” said Duggal. “Depending on your use case, big AI factories may make sense for you and pay off in the long run. But if there are not enough use cases or you are just testing the waters, then that may not pay off.”

When it comes to deciding whether to go with a proprietary model such as OpenAI’s GPT-4 or Databricks’ DBRX open-source model, cost is a factor that might not be obvious at the beginning, says Duggal.

“Cost is a major consideration when you work with these big models on Microsoft Azure or Google Vertex. You have to keep in mind that there’s a certain cost structure you're signing up for,” he said.

Ultimately, it is clear that the key to successful AI implementation is aligning the choice of model with specific use cases and strategic goals.

Image credit: CDOTrends