Marketers' Generative AI Problem: Numbers vs. Words

2023 has been the year of generative AI hype, centered around large language models (LLMs). Despite an otherwise cool VC market, the percentage of VC dollars pouring into AI startups has more than doubled in 2023, to more than a quarter of all investment. But we all know the classic Gartner hype cycle for new technologies, and in fact Gartner itself is now warning that generative AI is at or near the peak!?

The signs of the oncoming Gartner “Trough of Disillusionment” are showing. Traffic to ChatGPT, the “Kleenex” of generative AI, dropped three months in a row, two of which were by a double digit percentage. A recent Gitlab survey showed that only a quarter of organizations are actively using AI in software development, despite that being one of the most concrete and widely discussed immediate use cases for generative AI.?

Promise vs. Impact

The significant gap between the promise of generative AI and its actual impact is showing itself. And marketing suffers from as wide a gap as does any function or use case. While 63% of marketers plan to invest in generative AI over the next 24 months, only 14% of organizations are currently using generative AI for marketing and sales today (according to McKinsey).

AI is not new to marketing, as marketing has been arguably been one of the most advanced and impacted sectors for AI for the last decade plus. Marketers use AI for targeting, optimization, personalization, analytics, and much more. Adopting generative AI seems like it should thus be an easy transition for marketers.

Yet so far it’s not. So what is causing this gap between promise and impact, and how do marketers close the gap? One important under-discussed issue has to do with the nature of the data LLMs use.

Marketing is Powered by Numbers

Marketing has spent the last two decades creating, aggregating, storing, optimizing, and leveraging structured tabular usually quantitative data. We track and quantify what we know about people, households, sites, apps, media sources, behaviors, locations, purchases, economic factors, psychographics, ad effectiveness… anything we can get our hands on.

LLMs are Powered by Words

But none of that is what powers generative AI. Generative AI largely leverages unstructured data such as text, images, videos, code, and audio. LLMs are not good with numbers. They’re bad at math, and they struggle to understand the relationships between numbers as well as they do between words. It’s intuitive to us humans that the number “10” is very different from the number “1000”. But to an LLM, those two numbers are just different strings of characters.

That’s why LLMs are mostly useful for words today, like customer service email responses, chatbots, product descriptions, or blog posts. But so much of what marketers spend their time and money on involves complex real-time decisioning that is based on databases of structured data. There’s still a significant gap we must bridge such that generative AI can personalize in that setting. Here are a few examples:

• Generating creative visuals based on ad versioning that has scored the best at attention metrics in the recent past
• Optimize ad copy in a given ad opportunity based on the household income and location-based segmentation data we have for that user
• Personalize a website experience based on the purchase and click history of that visitor

Marketers have access to a ton of incredible data. For generative AI to reach its potential in the marketing stack, we need to ensure we can personalize the creativity of LLMs using the best data assets to which marketers have access. Until then, LLMs may be limited to a bit role.