Joel has been with Avensia since 2014 and has worked on some of the most successful e-commerce sites in the Nordics. He’s also an AI passionate, leading innovative initiatives within the Avensia business as well as in customer projects.
What Are the AI Use Cases for E-Commerce Businesses Today?
If you attend industry events, you’ll surely hear AI visionaries talk about all the amazing things this technology will do for your business in five or ten years. But if you look at what makes sense today, most use cases include creating and editing text. Such as:
- Creating product descriptions based on PIM data.
- Enriching product descriptions with data from different sources.
- Adding or editing categories according to your product catalogue.
- Translating content for your international sites.
- Creating social media posts.
- Adapting product descriptions to different channels.
- Editing generic text to ensure consistency with brand style and tone of voice.
What Are the Typical AI Options Available for E-Commerce Businesses Today?
There are three main approaches to explore, and a fourth one that most don’t even consider as an alternative but that could be a perfect fit for many e-commerce businesses. Let’s go through the approaches:
- Vendor supplied AI modules, meaning built-in AI modules in existing solutions, such as the PIM, CMS and commerce systems. In the last few years, the AI FOMO (Fear of Missing Out) has made product tech companies mindlessly toss money to their developers and say, “Build AI!” This has resulted in many of them now supplying AI in some form, built into the product.
These system built-in-tools often use an existing AI model (like Open AI’s ChatGPT) in the background, adding prompt-engineered custom prompts and a user interface to fit the purpose of the solution. Unfortunately, these modules often come with a hefty price tag for the value they bring. - Using AI APIs directly (such as ChatGPT and Anthropic’s Claude) combined with prompt engineering to fit the specific business needs. Since these models are generic and can handle any type of task, the data sets “loaded into the model” are huge which increases the usage cost. Also, data and training capabilities of the AI model are owned by the supplier which limits the possibility to customize. You can buy fine-tuning capabilities on e.g. GPT-4, but it’s important to understand that you’ll be fine-tuning a gigantic all-purpose model, which means costs will increase accordingly.
Another challenge to overcome for this approach is prompting. Prompt engineering is not just a fancy buzzword for typing text into a chat. The difference between a well-engineered and iterated prompt compared to a single first-try prompt can be huge. Most of the initial AI rejectors I’ve talked with have judged the capabilities of the model(s) by their anecdotal usage after prompting it once or twice. Unfortunately, this may lead to dismissing AI when in fact their use case has been demonstrated to be a perfect fit for LLM’s (Large Language Models), with the correct prompting. Prompt engineering is an art, similar to programming. It requires experience and know-how.
This puts further complexity on your organization. Do you have the knowledge and experience to manage prompt engineering in-house? Alternatively, do you have the sourcing experience to hire the right consultants for this purpose? - Open-source LLM AI models let you build and host your own AI models (like Meta’s Llama) that you can yourself continue to train or fine-tune for your specific needs. This inherits all the complexity from previous steps and adds some new ones: hosting, maintaining & training. It opens some previously closed doors, namely data security and integrity. Your model can be hosted within your data center, and you can make sure the data, both training and usage, never leaves the data center. In fact, you could even host the LLM on a machine with no internet access.
The Niche Fourth Option
There’s a new alternative on the market. One of our partners has developed an LLM, from the ground up, that’s designed for e-commerce. Because it has a specific purpose rather than an AI tool that can do anything and everything, this model can work way faster and cheaper as it requires much less training and hosting capacity. In essence, you get:
- Full control over which data the model uses, both training and inference.
- The possibility to host it yourself or use it as an API.
- A specific-purpose model that makes it extremely fast and cheap to run.
- Cost efficient training of the model, due to its size.
What Are the Benefits of a Custom-Built LLM for E-Commerce?
Because a custom-built model can be much smaller than a generic AI tool, it will be both faster and cheaper to operate. With smaller data sets it requires less server space and hence can be run on standard machines rather than costly huge AI specific servers.
When we say small in this case, we mean the model is limited to support requests related to e-commerce operations rather than providing information about everything and anything. So, you won’t be able to ask this AI about the weather forecast, the history of Australia or the research from the latest Nobel prize laureates in Physics. That’s where generic tools like ChatGPT will be the better option. However, using ChatGPT to enrich product data for an e-commerce site is like using a bulldozer to drive a nail in a plank of wood. The result is the same, the nail will be driven into the plank, but there are better tools for the job than a bulldozer.
In relation to generic models a custom-built LLM can be small, but that doesn’t mean it can’t handle the large volumes of data needed in e-commerce businesses. Think about the hundreds of thousands of products an e-commerce site may include. Multiply that with the number of channels and languages to support with accurate product descriptions. Content that will need to be updated frequently. This is where you’ll see the true power of a custom-built LLM – handling these data volumes in a cost effective and secure model where you can control the data.
What is Your Advice to Businesses Experiencing AI FOMO?
Don’t run into an AI initiative or implement a tool just because “we need to work with AI”. Take your time and research where an AI tool would make sense in your business, which tasks could be conducted by AI and how it would fit into your existing system landscape.
Evaluate the AI tools you consider with a business-first mindset. Does the AI tool give real value to the customer? Will you get an ROI on the extra license cost it’s often bundled with? These are very relevant questions and often quite hard to answer. The often-unexplored jungle of AI possibilities may seem grand, but it might be just filled with exciting stuff that doesn’t help your core business. At least for the price. My advice is to first identify needs and strategies, without AI in mind, and then to see if there are any AI tools that can help you on this journey.
In my opinion, the biggest challenge for e-commerce businesses today related to AI is understanding the role of AI in their business – which tools to use, how to use them and how to integrate with other solutions.
This is where Avensia’s advisors can help navigate the AI landscape and support with a recommended way forward. Contact us to discuss your business needs.
