How much data do you need to get a positive ROI from AI?

I have been asked this question so many times in my 30+ years of working in the AI world. Instead of attempting to give a general answer (which would be impossible), I would like to share the answers within the context of two specific solutions. I will presume that you are looking for ROI (Return on Investment) as the success metric.

Voice of Customer (VOC AI) Solution:
Predicting what your customers and your competitor’s customers are talking about, or complaining about, as accurately and rapidly as possible, can help increase positive sentiment, improve customer experience, prevent churn, and increase ideal customer acquisition. Improving positive sentiment by 1% could increase your revenue by 0.3 to 0.5%, that’s millions.

Generative AI works really well here, and best of all you actually need NO PAST DATA. Of course using the past data could help, but you don’t need to label the data, but just describe in a prompt what you want to predict as sentiment, topic, subtopic, complaints. You can hook up the solution to your live data and see the results right away. At TAZI we have been able to curb output format problems and hallucinations through well-designed prompts, and you can keep updating the solution based on the accuracy of its predictions. You observe the results in TAZI and give feedback directly to your solution, using TAZI’s UI.

You may be also assuming that you need to be a data scientist to do all this but you don’t! I can’t believe that in the past we needed a ton of NLP tools and models to get this done, and it used to be so much effort. Today, you just need a local LLM (7B parameters will do) and a consumer grade GPU. And you don’t need labeled data!

Client Attrition (Churn) Prediction:
Predicting and preventing attrition is something every institution should do even if the churn ratio is in single digits. You worked hard to acquire those customers and you shouldn’t lose them, any of them. To put this in the context of value, reducing churn by 1% could have a value of 6 times the revenue you get from each customer (Customer Life Time in years + CAC). It is even more if new customers bring more risk, unpredictability, or they cost more than the existing customers.

We found that predicting attrition in advance and having the details on why the attrition is happening, are key in preventing churn because you have time to take the right actions, for each client. Even if you can’t take actions now (maybe your teams are too busy), knowing who will be churning helps you plan for the future, as opposed to relying only on totals from the past year.

For churn prediction, 2 years of past data is usually a good starting point (at TAZI we have also succeeded with a lot less data than that, with the help of LLMs).

As the success metric, we found the ROI to be quite useful. ROI (Return on Investment) is defined as ((Net Profit-InvestmentCost)/InvestmentCost) * 100. Usually you are able to see if there is a positive ROI in predicting attrition with about 2 years of data. Just last week I saw an 800% ROI for a customer, and now I can’t stop thinking about who else in the world is not using AI to predict churn because they think they don’t have enough data. Don’t be that person!

Zehra Cataltepe, CEO & Co-Founder, TAZI.AI