KTS - [SF8] AI-based Customer Behavior Prediction (New Cus)

Inspiration

New customers are the hardest segment to understand because they have little or no historical interaction with SVFC. Traditional recommendation and targeting models work best when there is rich first-party behavioral data, but for new customers that signal is weak, delayed, and often incomplete. We wanted to solve this cold-start problem by using broader consumption behavior from external ecosystems such as telco, social media, e-wallet, and e-commerce, then combining those signals with each customer’s earliest reactions to SVFC offers. The goal was to help the business move from broad acquisition campaigns to more precise, personalized engagement from day one.

## What it does

KTS - [SF8] AI-based Customer Behavior Prediction (New Cus) predicts what a new customer is most likely to be interested in next. It analyzes cross-domain behavioral indicators, such as spending patterns, digital activity, product affinity, engagement timing, and channel preferences, then compares those signals with patterns learned from similar past customers.

The solution generates a ranked prediction of likely interests, product categories, or next-best offers for each new customer. It also adapts as more early interaction data becomes available, such as clicks, views, responses, or conversions from SVFC campaigns. This allows the system to improve targeting quality quickly, reduce wasted outreach, and support more relevant customer journeys early in the relationship lifecycle.

## How we built it

We built the solution as a behavior prediction pipeline that combines external consumption signals with SVFC interaction data.

First, we defined a feature layer from multiple data domains, including telco usage behavior, social engagement tendencies, e-wallet transaction habits, and e-commerce purchase patterns. These inputs were transformed into standardized behavioral features such as spending intensity, category preferences, time-of-day activity, responsiveness, and digital maturity.

Next, we trained machine learning models on historical customer cohorts to identify patterns between early-stage signals and later product interest or conversion behavior. We used similarity-based matching and classification logic to compare new customers against known behavior profiles from past customers.

We also designed the solution to continuously update predictions when new response signals arrive from SVFC offers. This gives the model a feedback loop, allowing it to refine recommendations over time instead of relying only on static onboarding data.

Finally, we structured the output for business use: predicted interest segments, next-best-offer suggestions, confidence scoring, and explainable drivers that help marketing and sales teams understand why a customer is being targeted in a certain way.

## Challenges we ran into

One of the biggest challenges was the cold-start problem itself. New customers naturally have limited direct data with SVFC, so the model needed to make useful predictions from weak and indirect signals.

Another challenge was data heterogeneity. Telco, social, e-wallet, and e-commerce data all have different formats, frequencies, and behavioral meanings. Turning these into a shared representation without losing important context required careful feature engineering.

We also had to address signal quality and bias. Not all external behaviors are equally predictive, and some patterns can overfit to historical segments if not handled carefully. Balancing model accuracy with fairness and generalization was a key design challenge.

Interpretability was another issue. Business teams need to trust the predictions, so we had to think beyond accuracy and provide reasoning that explains which signals influenced the model’s output.

## Accomplishments that we're proud of

We are proud that the solution directly addresses a real business pain point: understanding and engaging new customers before enough internal history exists.

We are also proud of building a cross-source intelligence layer that does not rely on just one data stream. By combining behavioral signals from multiple ecosystems, the model produces a richer and more adaptive view of each customer.

Another accomplishment is the practical business framing of the output. Instead of generating abstract scores only, the system translates predictions into usable actions such as next-best offers, priority segments, and engagement guidance for downstream teams.

Most importantly, we demonstrated how AI can turn weak early signals into actionable predictions, making personalization possible much earlier in the customer journey.

## What we learned

We learned that predicting new customer behavior is less about having perfect data and more about combining small signals intelligently. Even limited early actions can become highly informative when they are interpreted in the context of broader behavioral patterns.

We also learned that feature design matters as much as model choice. The quality of behavioral abstraction, such as how we represent intent, readiness, and affinity, strongly influences prediction usefulness.

Another key learning was that explainability is essential for adoption. A model that performs well but cannot justify its predictions is harder for business users to trust and operationalize.

## What's next for KTS - [SF8] AI-based Customer Behavior Prediction (New Cus)

Next, we want to improve the model with real-time learning so predictions can update instantly as customers interact with new campaigns and channels.

We also plan to expand explainability features so business users can see not only the prediction but also the top behavioral drivers and recommended actions behind it.

In the longer term, we want to integrate the solution more deeply into campaign orchestration, CRM workflows, and next-best-action engines so that predictions can automatically trigger personalized journeys, product offers, and retention strategies for new customers at scale.

Built With

• dart
• flutter
• javascript
• qwen