GenAI assistant

Jun 6
4 min read

Pecan AI is a predictive analytics platform that helps teams train machine learning models and turn business data into predictions. This project focused on using GenAI to help users move from a vague business question to a clear model training starting point.

Project overview

The GenAI assistant was shipped across several product iterations. It first appeared on the homepage to help new users start their first session, then supported the quick-start model training flow, and later evolved into a copilot inside a notebook-style predictive analytics experience. This case study focuses on the chat itself, its role in reducing activation friction, and the value it created for users, product learning, and PLG growth.

The challenge

Users came to Pecan with a business outcome they wanted to predict, but often did not know how to translate that goal into a use case, template, data structure, or SQL-based model setup. Static templates helped with common use cases like churn, but many customer needs were more specific, sensitive, or did not fit cleanly into a predefined template. The challenge was to make GenAI useful inside a complex ML workflow without turning it into an open-ended chatbot that created vague answers, wrong expectations, or unsupported model paths.

Users did not know where to start.
Templates did not always match the real use case.
The SQL editor created a high learning barrier.
Users needed guidance without losing control.

My design challenge

The hardest part was designing the assistant as both flexible and safe. It needed to understand messy business language, support specific prediction needs, and create a useful starting point, but it also needed enough structure to keep users inside a valid model setup flow. I had to balance user freedom, model logic, technical constraints, and trust in the generated output.

Why GenAI

GenAI gave us a way to shift the experience from template-first to intent-first. Instead of forcing users to choose from a fixed list of use cases, the assistant let them describe what they wanted to predict in their own words. This created a more natural entry point into model training and gave the product a better way to understand user goals, business context, and data needs.

My role

As the Senior Product Designer, I led the UX strategy, chat experience, user journey mapping, UI design, and design system alignment. I worked closely with data scientists, engineers, PMs, subject matter experts, and customer success to shape the assistant’s behavior, map edge cases, validate technical logic, and make sure the chat supported the model training flow instead of becoming a separate AI feature.

Research

The project was shaped by Mixpanel, PostHog, LogRocket recordings, Gong calls, customer success feedback, user interviews, and internal reviews with data analysts and data scientists. I looked at how users interacted with the chat across each product phase, what they typed, where they got stuck, which templates they selected, and how their intent changed once they better understood what Pecan could do.

Product evolution

The assistant evolved in three shipped phases. In the homepage version, it helped users explain what they wanted to predict and matched them to an existing template. In the quick-start flow, it supported a guided path to train a first model without pushing users directly into the SQL editor. In the notebook experience, it became a copilot that helped users generate more specific templates, support data mapping, and understand the next steps in a linear model training flow.

The solution

The final version worked as a guided GenAI assistant inside the model training experience. Users could describe their prediction goal, and the assistant helped translate that intent into a relevant model starting point. It supported the user through template generation, data mapping, and next-step guidance, while the surrounding notebook flow kept the experience structured and understandable.

Key design decisions

The main design decision was to balance flexibility with control. A fully open chat could capture user intent, but it also risked vague prompts, unsupported requests, and unclear expectations. A fixed template list was safer, but too rigid for specific business needs. We designed the assistant as a guided copilot: flexible enough to understand user intent, but structured enough to support a reliable model setup path.

Innovation value

The assistant was not added as a novelty layer. It solved a real product problem: users needed a better way to start. It also created a new research channel for the team. By reviewing chat prompts and interactions, we learned what users actually wanted to predict, which use cases were missing from our template library, what data users had, and where the product needed stronger guidance.

Business value

From a PLG perspective, the assistant helped users reach value faster without relying on sales calls, long documentation, or customer success intervention. It supported self-serve activation by helping users start with their own business question, move into model setup, and return for additional sessions with a clearer understanding of Pecan’s value.

Results

We measured the assistant through user interaction, model setup progression, data upload or connection, return sessions, team creation, and reduced customer success questions around how to start. The strongest signals came from users engaging with the chat, moving further into the training flow, returning after the first session, and using Pecan to build models for their teams.

Chat interaction
Model starts
Data upload
Return sessions
CS questions

Outcome

The GenAI assistant helped move Pecan from a static template experience toward an intent-driven model setup. It gave users a simpler way to begin, helped the team understand real prediction needs, and created a stronger connection between user intent, data requirements, and model training. For the business, it supported product-led growth by reducing early friction and making the product easier to adopt.

Next step

The next direction was to expand the assistant beyond template generation. The long-term vision was to help users inspect data quality, identify missing fields, map data correctly, and improve model accuracy before training.