AI Products Are Data Products

Problem

I started off as a SQL developer designing and creating databases. I would pull data into a spreadsheet and analyze it trying to reduce the cost and increase the revenue for the company. Understanding the impact and significance of the smart use of data led me to my first AI job.

Below, I will share three different AI projects that I worked on and that reveal the immense potential AI products have. What I came to realize is that AI products are essentially data products. If you don’t have data in your product you can’t build an AI product. That is, if you have no means to create or input data, you can’t build AI into your product.

Actions taken

I decided to single out three AI products I was working on -- one on structured data, two on unstructured data -- to highlight the immense potential of AI products and the main difficulties that come along with building them.

My first AI product was a telecom product. I had data about every call a person made for the last two years. Based on their calling behavior I had to predict who would cancel their subscription and who wouldn’t and what we could do to keep them engaged. For example, it would cost the company 5 dollars to make them customers and it could make 50 cents from them every month. That meant that a customer had to be with the company for at least 10 months to make up for the costs of acquiring them. If the company couldn’t make customers stay for two years, it could perhaps increase the revenue from them. I had to understand if and why a person would pay 2 dollars instead of 50 cents and how could I build a product out of it. I built a product based on structured data, machine learning, and logistic regression. I used data to do AI and use the output of the AI to build an AI product. I helped the company understand what was the campaign that made a customer stay (100 more minutes a month or a new phone at 50 percent less price) and how it could incentivize and retain customers.

The second product I built was based on unstructured data and was using text data and natural language processing techniques. The company I was working at was helping attorneys who would go through financial statements and make an assessment if a company was worth acquiring. To make that assessment attorneys would have to read thousands and thousands of documents. If one person would do it alone it would take them eight months and six people could do it for two months. They would read the documents, highlight the import parts, summarize that in a separate document, and repeat the process all over again. We proposed to scan the documents and present them in a readable format. I automated the process using AI and writing the code in Python, but also created a new business model. We offered attorneys to pay us only 10 percent of the money our product would save them. Not only that I built a new product for the company but I generated a new revenue stream.

The third AI product that I want to single out is an app that allows customers to create facial expressions and voice messages that resemble them. We had to teach software to recognize human emotions and expressions (winking, nodding, etc) and had to record people doing funny expressions, recognize their faces, and differentiate between genuine and fake smiles, for example. The product had to be ethical and include minorities and we didn’t have much data on minorities. Not only that building AI part itself was hard but curtailing our biases in assuming one ethnicity, gender, or age added to the problem. That included also modeling for all possible combinations of hair color, mascara, nose piercing, etc. and collecting all those data was impossible. In addition, I had to expand my knowledge on race and ethnicity, bias, and inclusion and it was entirely a new product experience for me.

Lessons learned

• All AI projects start with data. If you don’t have data you can’t do AI.
• Structured data is the best form of data to start your AI journey with. It’s well understood, very well documented and can work with unsophisticated AI algorithms.
• Between lots of data and complicated algorithms choose the former. What happened in the past is a far better indicator of the future than complicated mathematical approximations.
• If you cant get your hands on structured data try text data. We read and write every day and understand and applying AI to this data is quite intuitive to us.