How Predictions Work

Jun 9
4 min read

Pecan AI is a B2B predictive analytics platform that helps teams turn business data into model-based predictions. This project focused on helping users understand how predictions work before asking them to train a model.

Project overview

How Predictions Work was an interactive learning flow created for Pecan AI’s Learning Hub. The flow explained predictive analytics through a simple churn example, showing how raw historical data becomes training data and how training data supports a future prediction. The experience was shipped, tested live for roughly four weeks, and used as part of Pecan’s product-led growth strategy.

The challenge

Predictive analytics was difficult for new users to understand. Users were confused by training data, data requirements, model logic, and the difference between historical outcomes and future predictions. This created a learning barrier before users reached more complex product areas like SQL, data integrations, and model training.

Users did not understand what training data is.
Users did not know what data they needed.
Users did not trust the model logic.

The goal

The goal was to turn a technical concept into a short, visual, product-native learning experience. Instead of relying on long documentation, videos, or support-heavy onboarding, the flow needed to help users understand the value of their data and move closer to self-serve model training.

We wanted to teach predictive analytics the same way it is taught in the classroom, but make it simple enough to understand inside the product. Noam Brezis‏, CTO & Co-Founder at Pecan

My role

As the Senior Product Designer, I worked directly with Pecan’s founder, who was also the subject matter expert. He brought the teaching method and predictive analytics knowledge. I translated it into a guided learning flow, structured the experience, refined the UX copy, designed the UI, and aligned the solution with our design system and PLG goals.

Product Design & strategy
Learning flow structure
Copy refinement
UI & Design system alignment

Research

Because Pecan followed a lean design process, we focused on fast learning and quick validation. I reviewed Gong calls, customer conversations, recorded sessions, behavioral data, and customer success feedback to understand where users struggled and where the concept started to make sense. I also worked with the DPM, engineers, and data analysts to validate the flow from both product and technical perspectives.

Gong calls
Recorded sessions
Behavioural data
Customer success feedback
Internal data analyst review

Core insight

Users did not need a full technical explanation at the beginning. They needed to see the logic in motion: start with historical data, create samples, compare past activity with known outcomes, and use those patterns to estimate a future result.

The solution

We created an interactive walkthrough that taught predictive analytics through one clear business example. Churn was selected as the default use case because it is familiar and easy to explain, while the flow structure stayed generic enough to support other prediction use cases.

Flow structure

The flow followed a simple learning path: select a use case, review input data, move through training samples, and end with a prediction output. Each screen focused on one learning point and reused the same visual system, so users could connect the table, timeline, sample date, known outcome, and final prediction without learning a new layout each time.

Key design decisions

Every UI element was reviewed against the learning goal. The stepper showed progress through the walkthrough. The timeline explained time-based logic. The table exposed the data structure behind the model. We used known_outcome instead of label to avoid confusing historical truth with a model prediction. The copy stayed short because the visuals carried most of the explanation.

Another important decision was where this flow should appear. Most users do not enter a product to study, especially during onboarding. If this walkthrough was forced into a quick start flow, many users would skip it without building real understanding. Instead, we designed it as contextual learning: users could open it from the Learning Hub or from relevant points in the model training process, such as when they needed help understanding a use case or data requirements.

Why "Churn" as default

Churn worked well as the first example because it has a clear before and after structure. A customer is active, then the model checks whether they stay active or become inactive. This made the concept easier to understand while still representing the same logic used in other prediction use cases.

Prediction output

The final screen made the core distinction clear: training data uses known outcomes from the past, while prediction output estimates what is likely to happen next. This was important because users often confused the known historical outcome with the model’s future prediction.

Validation

The flow was shipped and tested live for roughly four weeks. We measured clicks from the Learning Hub, step progression, walkthrough completion, short feedback quiz responses, and whether users who completed the flow were more successful in training their own model afterward.

Outcome

The walkthrough created a clearer bridge between education and activation. Users could understand the value of their data before moving into technical setup, but only when the learning was relevant to their current intent. From a PLG perspective, the flow helped reduce the education gap, made the product feel more approachable, and supported a stronger path toward self-serve model training.

What I would improve next

The next step would be to connect the walkthrough more directly with the model training flow. After users understand the concept, the product could guide them into their own use case, help them identify the required data, and reduce the gap between learning and action.