(Ise is pronounced “ee-say”)

In the last article, we discussed how to align Gen AI tools with business outcomes to make its ROI case clear. In this article, we will tackle the second objection for Gen AI adoption that we hear from enterprise buyers -- skepticism against Gen AI model output and quality control.

The challenge

Over the past year with Gen AI buzz all over the news, we sometimes hear of cases where the AI models come up with outputs that deviate significantly from reality and provide information that is objectively wrong. For example, ChatGPT came up with “bogus judicial decisions with bogus quotes and bogus internal citations” that ended up in a legal brief that was submitted to federal court. This phenomenon is called “hallucination,” and it happens when Gen AI models receive insufficient or biased training data, have incomplete understanding of the context, and/or are not structured to handle the full complexity of tasks at hand.

Clearly, these incorrect outputs are problematic across all fields, but especially in retail where decision making relies on accurate, granular information, and marketing content must reflect the product with high fidelity. The solution is using AI models that are custom-built for retail -- rather than off-the-shelf, generic AI models that do not understand the retail context and hallucinate easily.

Thesis: retail needs its own AI models

The solution to reducing hallucination and unlocking the full power of AI is leveraging vertical AI models. But first, let’s define what a vertical AI model is. A vertical AI model is specifically designed and trained for a particular industry or domain, leveraging industry-specific data and expertise. In contrast, a horizontal AI model is a more general-purpose model that aims to cover a broad range of applications across various industries -- examples of which are Midjourney and ChatGPT.

We believe retail needs its own vertical AI model for two main reasons:

First, retail wins on personalization -- not just personalizing the shopping experience to the end consumer, but also personalizing content creation to the brand itself. Generic horizontal AI models cannot meet this requirement because their model architectures are not designed to accommodate retail-specific contexts, needs and preferences, and behaviors and vocabulary. What you end up seeing is AI-generated marketing content that falls in the AI “uncanny valley” -- where AI models with disfigured body parts are shown and marketing content that do not look like the actual item designed and sold by the brand.

For a more technical explanation of the AI “uncanny valley” and how vertical AI models overcome it, check out Ise Founder and CEO Vanessa Yan’s LinkedIn post.

Second, retail professionals have a specific way of working that is often cross-functional in nature and follows a seasonal calendar. When a merchant has to collaborate with stakeholders across 5+ functions and generate different reports and dashboard summaries per meeting, the last thing they want to do is to ask ChatGPT to run an inventory data summary 4 different times with high variation in output format, until the report is refined the way they want. Most horizontal AI companies only provide basic chat interfaces -- which do not yield the productivity boosts and user satisfaction often advertised. Instead, with a vertical AI model, retail users will benefit from a UX flow that is designed with their use case in mind. Additionally, the seasonal nature of retail planning and go-to-market often means any copilot features must understand the retail calendar and contextualize tasks accordingly.

Want to hear more from Ise AI? Follow our Ise AI Insights series and sign up for a demo.