SAILES

What Inspired Us

In the world of enterprise sales, predictions are only half the battle. Sales teams at Schneider Electric were facing a critical challenge: while machine learning models could predict whether a deal would be won or lost, they couldn't explain why! This "black box" problem created a trust gap, how could consultants take action on predictions they didn't understand? How could managers refine their strategies without knowing what actually drives success?

We were inspired by a simple but powerful question: "What if AI could not only predict outcomes but also teach us what makes them happen?"

This challenge resonated with us because it sits at the intersection of cutting-edge AI and real-world business impact. We wanted to build something that would empower sales teams to work smarter, not just harder.

How We Built It

The Foundation: Data Understanding

We started by diving deep into Schneider Electric's CRM data—millions of data points representing sales opportunities, customer histories, product mixes, competitive landscapes, and relationship dynamics. Understanding this data was crucial before building anything.

A Three-Layered Architecture

Rather than building a single monolithic model, we developed a three-layered explainability framework:

Global Insights Layer - We implemented Feature Importance Analysis and TabCBM (Tabular Concept-Based Models) to understand what matters most across all opportunities. This revealed that customer relationship quality dominates success, accounting for 65% of prediction power.
Local Explanations Layer - Using Decision Path Analysis with Random Forest models, we could explain individual predictions. By analyzing which sub-models acted as "tie-breakers," we identified which specific trends were most important for each unique opportunity.
What-If Analysis Layer - We implemented a Tabular Variational AutoEncoder (VAE) to generate counterfactual explanations. This cutting-edge approach finds the minimal changes needed to flip a prediction, giving sales teams actionable next steps.

The LLM Integration

We powered everything with an LLM-based assistant that translates complex model outputs into natural language. The assistant has access to all three layers, allowing it to provide comprehensive, human-friendly explanations like: "This opportunity has a high probability of success because the customer has a 95% historical success rate and we've had 12 previous positive interactions with them."

What We Learned

Technical Insights

Interpretability ≠ Simplicity: We learned that truly explainable AI requires multiple complementary approaches. No single technique—whether SHAP, LIME, or decision trees—tells the complete story.
Counterfactuals are Hard: Generating realistic "what-if" scenarios for tabular data is significantly more challenging than for images or text. We had to carefully balance validity, proximity, and sparsity.
Concept-Based Models are Powerful: TabCBM automatically extracted meaningful business concepts from raw features, creating a natural bridge between data science and business language.

Business Insights

The model revealed surprising patterns:

Competitive threats are overestimated—they reduce win probability by only 8-12%
Stalled opportunities (90+ days) rarely recover, with 60% lower win rates
Past product success is a 3.5x multiplier for future wins
Relationship quality matters more than product features or pricing

Team Dynamics

We learned the importance of balancing technical sophistication with business usability. Early versions were technically impressive but too complex for end users. We had to constantly ask: "Would a sales consultant actually use this?"

Challenges We Faced

1. The Explainability Paradox

Challenge: The most accurate models (deep neural networks) are the least explainable, while simple models (linear regression) are explainable but less accurate.

Solution: We found a middle ground with Random Forests for prediction (83% accuracy) combined with post-hoc explainability techniques. TabCBM provided the best of both worlds—accuracy and interpretability by design.

2. Data Quality and Completeness

Challenge: Real-world CRM data is messy. Missing values, inconsistent formats, and data entry errors were everywhere.

Solution: We invested significant time in data cleaning and feature engineering. We also designed our models to handle missing data gracefully rather than requiring perfect inputs.

3. Making Counterfactuals Realistic

Challenge: Early versions of our VAE generated counterfactuals that were technically valid but practically impossible (e.g., "increase customer interactions by 500").

Solution: We added constraints and proximity metrics to ensure counterfactuals were both plausible and achievable. We optimized for minimal feature changes (average 1.5 features) to make recommendations actionable.

4. The Causation vs. Correlation Problem

Challenge: Our model identifies patterns, but correlation doesn't imply causation. We couldn't definitively say that increasing interactions causes wins.

Solution: We were transparent about this limitation in our documentation. We framed insights as "patterns observed in successful opportunities" rather than guaranteed cause-and-effect relationships.

5. Bridging the Technical-Business Gap

Challenge: Data scientists and sales teams speak different languages. Terms like "feature importance" and "counterfactual generation" meant nothing to our end users.

Solution: We built the LLM assistant as a translation layer, converting technical outputs into business-friendly language. We also created visual dashboards and focused on actionable insights rather than technical metrics.

What We're Proud Of

Real Business Impact: Our insights are already changing how sales teams prioritize their efforts
Technical Innovation: We combined multiple state-of-the-art techniques (TabCBM, VAEs, decision path analysis) in a novel way
User-Centric Design: We never lost sight of who would actually use this system
Transparency: We were honest about limitations and designed the system to be auditable

What's Next

This project opened our eyes to the possibilities of interpretable AI in business contexts. Future enhancements include real-time updates, automated alerts for at-risk opportunities, and integration with email and calendar systems for automatic interaction tracking.

But beyond features, we learned that the future of AI in business isn't just about accuracy—it's about trust, understanding, and empowerment. SAILES represents our vision of AI as a collaborative partner, not a mysterious oracle.