The great AI mismatch: How to make sure LLMs pay off

A few years back, every brand was scrambling to get onto Alexa. From national pizza chains to leading rideshare firms, tons of companies introduced Alexa “Skills” all at once. This marketing blitz gave the voice assistant major buzz, but never really paid off for all the brands investing time and money in it.

Think about how we’ve truly used the device. I bet most of us rarely said anything like “Alexa, order me a pepperoni pizza.” Instead, we all stuck mainly to simple tasks like “Alexa, what time is it?” and “Alexa, set a timer for 10 minutes.” Unfortunately for many of the brands that tried to make the most out of Amazon’s voice assistant, the devices ended up being “a huge number of glorified clock radios.”

Today, we’re told that generative AI and large language models (LLMs) have leapfrogged rules-based assistants like Alexa and Siri—and are poised to totally transform how brands communicate with their customers. I’m a believer; in fact, I recently said on CNBC that generative AI will have as great an impact on business as the introduction of the PC. But I also think it’s important to reflect upon the lessons of inefficient implementations of earlier AI models, so that we don’t put time and money toward dead ends like Alexa.

What happens when models and data aren’t aligned

The two most important elements that go into AI implementations are the data model and the data set. The model (for instance, OpenAI’s LLM) and data set come together to generate outcomes. There’s currently a lot of focus on models because of big splash launches like ChatGPT. But without the right union of models and data sets, brands and consumers will never get to the outcomes they want. 

The outputs or outcomes generated by generative AI can only be as strong as the match between the model behind it and the data flowing into it. Let’s say you work for a healthcare company and want to provide customized recommendations about COVID safety ahead of the cold and flu season based on your customers’ city or state. Let’s also say that you’ve selected the right model and you feel confident your model can generate language in the right tone of voice, without discriminating against any group of people, and even integrating with your backend systems to send customer reminders or help them schedule appointments. 

But to generate the COVID information you send your customers, you draw from conversations on the public internet. If your goal was to generate trustworthy, helpful advice about healthcare, you’ve already failed; there’s simply too much misinformation and noise about COVID online. You may have selected the right model to communicate with your audience, but the data set was a mismatch. 

Similarly, Alexa’s Skills never worked for the brands that tried to use them because its model was not truly designed to work for any brand but Amazon. The real focus was always on getting you to buy more from Amazon, not to support other brands on the platform. The model remained focused on the Amazon use cases. Alexa was never meant to serve the purposes that restaurant and rideshare brands (and their end-consumers) wanted to achieve. 

Putting AI and your data set to work for your business

The next generation of successful enterprises will be led by those who avoid limited, closed ecosystems—and those who remember the importance of their own data. In other words, the winners will match expansive data models with precision data sets. 

Of course, not every brand can invest time and money in creating their own AI models. A small number of companies will provide those or will help non-AI brands build their own. But every brand has its own data set, and every brand should be thinking about how to leverage that proprietary, precision data set to get the most out of the large models for its own bottom line.

This is the real opportunity of generative AI. It’s not about the day-to-day, hyperpersonal use cases like having ChatGPT do your homework or write your LinkedIn posts. It’s about matching precision data sets with expansive models to generate outcomes that are usable, purpose-built, and drive the outcomes that matter to you and your customers. 

With incredible generative AI models now out in the world, understanding the possibilities of your proprietary data set is now a business imperative. What data do you have that every other brand in the world would be jealous of? And what will you do next with generative AI to make the most of it? 

By Rob LoCascio, founder of LivePerson