Product data and generative AI: Together a profitable combination!

Whether or not you find, buy and return a product after receiving it depends largely on the quality of the product data. These must be accurate, complete and clear to properly inform your customers during the buying process and then keep logistics running smoothly. Delivering wrong products damages the image and results in negative reviews.

For collecting, managing and enriching product data, a product information management (PIM) system is the solution. The key idea is that there is one truth for all product data. PIM allows you to optimize product data centrally, but creating and translating product descriptions is often a tedious and time-intensive task. With the advent of generative AI, we held conversations with our customers to hear what problems with managing product data were most common. Together, we worked out the possibilities and needs of generative AI in PIM systems.

Best practices for using AI to enhance your product data management

1. Logical errors within data

A banana cannot be one meter long and a bicycle cannot fit into a small shipping box. These are logical errors that a human sees immediately when he or she looks at them. The crux is that in many PIM systems these errors are not found with standard checks: you must find them in the thousands of product records. This was the most frequently cited issue by our customers, regardless of market.

Generative AI can process the fields and descriptions and identify at which items something seems off, such as incorrect dimensions or inconsistencies. An employee can then review those reports and make corrections.

2. Make translations and change the context

Fully automated translation of product data has been around for some time, for example with a cloud service such as Google's Cloud Translation AI or Deepl Translate API. This allows you to quickly make your product offerings accessible in other languages. The power of generative AI is that you can do much more by also changing the context:

• Reading level - Translate and write for reading level B1 to make it more readable. Replace difficult words with synonyms and plain language.
• Length - Write the text very sec (dry) so that the description fits into this field, or make it longer.
• Style - Make the text less dry, or rewrite it in your own distinctive tone of voice.
• Language variation - Adapt the text for the Flemish market or the Dutch market.
• Thematic - Have the text respond to the theme (such as "Christmas"), for example, adding that it can be a nice Christmas gift.

You can ask a generative AI to do anything. However, the quality of the output depends heavily on how the input questions are formulated. The output is only as good as your question.

3. Filling empty fields and descriptions

Supplying detailed, complete product information in a market-compliant file format is still far from being the standard in every industry. For example, essential fields are not filled in or the description is very brief. Turning this into good product data takes time and attention: one of our customers reports that one person writes between twenty and fourty complete descriptions per day.

This is another area where you can deploy generative AI by filling fields based on the description, and vice versa. All you must do is review this data, thus productivity and creativity skyrocket. Interestingly, you can also use images as input: write descriptions or create new product records based on these images. By formulating your question intelligently - such as by including the eight possible colors, for example - you can quickly have many fields filled automatically. When this involves hundreds or thousands of products, every field you can fill this way makes a noticeable difference.

Generative AI such as ChatGPT or Gemini have access to the entire Internet and therefore knowledge of all standards, such as ETIM and 2BA. As a result, you can classify products and populate fields based on these standards without major modifications or additional training.

4. Making product data accessible with a chatbot

A common idea is to make product data accessible with a chatbot, for example, to give customers inspiration. The problem with that is that there is no longer a human filter in between that prevents incorrect information from being given. A freely interpretable question, such as, for example, "Which coat goes with a wedding?", may yield a strange answer that does not match customer expectations. Or you might have preferred to point out the promotions or just a product with more margin. As a brand or provider, you then lose control.

However, if you deploy a chatbot for your own employees with accompanying training, you give them a practical tool with which they can better help customers. They can keep asking questions, verify which direction the customer is already thinking and then put the chatbot to work in a focused way.

Less mistakes, less work

Anyone who has worked with ChatGPT knows that generative AI is anything but flawless: the disclaimer is there for a good reason. So, you will have to check the output carefully for the time being. However, that's considerably less work than having to track and fill in everything completely manually. You can then shift the use of your available manpower to correcting errors and making considerations or devote it to other important tasks. Over time, when the generative AI is well aligned with your company's requirements, you can gradually reduce control.

By formulating the prompts better and better and thanks to the self-learning ability of generative AI the error rate of the output will decline over time. Combining generative AI with humans is a gradual learning process. The challenge here is to be keen on the design of that combination and not to view generative AI as a "black box" from which all sorts of things automatically roll out: it is an extremely useful tool with advantages and disadvantages that you must learn to deploy effectively.