Causal AI in Telecom: AT&T’s Journey Beyond Predictive Analytics

AT&T, a titan in the telecommunications industry, has long been at the forefront of technological innovation. With millions of customers and an ever-expanding network, AT&T faces unique data challenges that require cutting-edge solutions. In recent years, the company has turned to advanced data science and artificial intelligence to tackle these challenges, with a particular focus on causal AI.

Causal AI has emerged as a critical tool for telecom giants like AT&T, especially when addressing complex issues such as customer churn. By understanding not just what happens, but why it happens, causal AI promises to revolutionize how companies make decisions and allocate resources.

In a recent presentation at The Causal AI Conference 2024, Sanchin Raj, Principal of Data Science and Advanced Analytics at AT&T, shared valuable insights on the synergies between predictive and causal models in customer churn reduction.

Speaker: Sanchin Raj, Principal, Data Science

This transcript has been lightly edited for length and clarity.

---

My name is Sanchin Raj. I work for AT&T as a practitioner of data science solutions for many of the business challenges we face. I have moved from small businesses to mid-sized areas, then to consumer, and returned to the enterprise unit, working under marketing analytics.

Current State of AI in Enterprise

These are independent numbers published by different institutions, but when you put them together, something emerges: the impact on the market is literally less than 1%. AI has not stood up to its hype when it comes to market impact.

We can all agree on this: all of the models and analytics that we develop and deploy are centered around decisions, and those decisions are tightly tied with return on investments. So why the failure? Why hasn’t AI stood up to its hype?

I’m convinced that the most important challenges come from a lack of cohesiveness, a disconnect among these three things: models, model outcomes, and ROI. If you dig deeper to understand why, you’ll see these issues keep coming up:

We keep using vanity metrics, talking about accuracy rates and F1 scores. These can take us directionally to a large degree, but by the end of the day, when it comes to returns, they do not have any assumptions baked into them.

These metrics don’t have any connection with the actual business outcomes.

We have methods and procedures for model development, structured thinking for everything, but when it comes to stakeholders – business partners and data scientist teams – there isn’t a cohesive, structured framework for how a problem is developed, stated, and taken to market or production.

4-Step Causal Thinking Framework

What I’m going to propose is practical suggestions and lessons I learned over the years. When I saw these issues, I tried to figure out why. I identified four issues and their solutions.

1. If you do not have a KPI that ties with your model outcome, engineer one.
2. Recognize that the value of insights is more important than your predictions.
3. Consider the costs of misclassification, not just the benefits of correct classification.
4. Take advantage of the power of persuasion that causal inference offers.

Detailed Explanation of Each Step:

1. Identify or Engineer a Business KPI

For customer churn modeling, we have predictive outcomes that clearly give us a propensity score, but when we look at volume, there’s almost no connection. It’s very important that we have a KPI with an expected non-linear outcome. For example, use Customer Lifetime Value (CLV) as a primary, leading indicator, with model scores fed directly into CLV. Use revenues and volumes as secondary indicators.

2. Value of Insights > Value of Predictions

Partner with stakeholders early on, unpack and reveal insights. For example, in our churn prediction model, we found Simpson’s paradox – long-tail tenures but opposing patterns between month-to-month and 2-year contracts. We also discovered that 50% of month-to-month customers leave within 6 months, but more than 50% of 2-year contract customers stay for 6+ years.

When you share this with your stakeholders, you create these aha moments. You create trust, you partner with them very closely, and then you create an anchor in the culture where they trust your models. Slowly, you start seeing the usage of the models go up.

3. Integrate Misclassification Costs

Business problem: We deployed a well-performing model, pulled the top two deciles, and ran a campaign. But in the end, stakeholders were disappointed. Everyone knows we lost money, but no one can pinpoint what, where, and how things went wrong.

We only considered the benefits of conversion, but not the costs of false positives and false negatives. Go beyond F1 score and accuracy rates. Incorporate not only the benefits of correct classifications but also the costs and risks of misclassifications.

Here’s a model outcome example. You have a classification matrix. You put together the number that is about 80% accuracy. A cost of, you know, the average cost of the false positives and false negatives that you keep in mind. Keep in mind that costs are not symmetric. I simplified it just by taking this symmetry cost.

For example, with an 80% accuracy rate and an average cost of $130 for misclassifications, you might barely make your case. But if the cost increases slightly, you could be underwater. Conversely, improving your model to 85% accuracy could significantly improve your returns. The point is that misclassification costs are vital to understand and dissect.

4. Leverage Causal Inference for Persuasion

Business problem: We deployed a great model with 84% accuracy rate, but only 3% converted. Stakeholders ask, “Why only 3%? Shouldn’t all the true positives convert, resulting in 14% conversion? We just ran a campaign promoting to all true positives.”

The issue is that we lost capital in promotional costs because not all true positives are persuadable. Implement a control-treatment plan, consider uplift modeling, develop Structured Causal Models, estimate Individual Treatment Effects, and use causal inference to target the persuadable customers.

Final Thoughts

I often get called on to solve complex problems, and the thinking is often, “Can we add more data?” The lesson I’ve learned is that there’s a decremental or declining value for adding more data at some point. It all goes back to playing some fundamental roles, going back to the basics, and thinking from the stakeholders’ point of view.

It’s equally important to avoid jargon, build trust, and think from the dollar value point of view. Then when you take it all to production, you’ll start moving the needle.

Are we out of the woods yet? No. If you think about what pundits would say, the challenges in the field vary. Some point out infrastructure issues, data quality problems, skill gaps, and now we’re talking about causality. It’s a journey, and if we as practitioners think causally not just from the outcome alone but as a framework, as a sort of power principles to keep in mind, we can start moving the needle. When you keep moving the needle here and there, eventually we’ll end up moving mountains.

The Role of Causal AI in Telecom 

Sanchin Raj’s presentation provides valuable insights into AT&T’s approach to causal thinking in data science. For the broader telecom industry, the application of causal AI presents significant opportunities for enhancing analytical capabilities and driving business value.

Enhanced Churn Prediction and Prevention: As discussed in Sanchin’s talk, Causal AI allows for a more nuanced understanding of customer churn. By identifying the causal factors behind customer decisions, telecom companies can develop more effective, targeted retention strategies. 

Personalised Marketing: Causal AI enables more sophisticated customer segmentation and targeting. By understanding the causal relationships between customer attributes, behaviors, and responses to marketing stimuli, telecom companies can design more effective, personalized marketing campaigns.

Improved Customer Service: By identifying the root causes of customer issues, causal AI can help in developing more efficient and effective customer service protocols. This can lead to faster resolution times and improved customer satisfaction.

Network Performance Optimization: Through causal analysis, telecom companies can better understand the complex interplay of factors affecting network performance. This can lead to more efficient resource allocation, proactive maintenance, and improved overall network quality.

As demonstrated by AT&T’s approach, a thoughtful integration of causal thinking into data science practices can lead to more actionable insights and improved business outcomes.
For telecom companies looking to advance their data science capabilities, partnering with advanced causal AI platforms such as decisionOS can provide access to cutting-edge tools and methodologies. These partnerships can accelerate the adoption of causal inference techniques and help unlock new avenues for trustworthy AI-powered decision-making.