937: How to Design AI-First Products — with Marc Dupuis
Marc Dupuis, CEO of fabi.ai, argues that AI tools will elevate, not eliminate, data scientists by automating routine coding and enabling a focus on higher-value tasks. He advocates for building automated workflows that meet stakeholders in their native tools like Slack, delivering far greater ROI than traditional dashboards, and shares core principles for designing effective AI-first products.
Marc Dupuis, co-founder of fabi.ai, believes the current wave of AI is the “single most exciting time for data scientists” in the last decade. Rather than making data professionals obsolete, AI-native platforms are poised to augment their capabilities, automating lower-value tasks and allowing them to focus on more strategic work.
Elevating, Not Eliminating, the Data Scientist
The value of a data scientist is shifting away from raw coding ability and memorization. In an AI-driven world, knowing the exact syntax for a Plotly chart off the cuff becomes less valuable. Instead, the focus shifts to two critical areas:
- Deep Business Acumen: The most impactful thing a data practitioner can do is deeply understand how the business makes money—the P&L, core drivers, and strategic goals.
- Fundamental Skills: With AI handling much of the coding, data scientists can dedicate more time to experiment design, hypothesis formulation, and result interpretation. A strong, fundamental understanding of statistics and machine learning becomes paramount for supervising the AI and ensuring its outputs are sound.
AI tools also empower non-technical counterparts, like product managers, to perform initial data exploration. This leads to a more data-literate organization where business users can ask more intelligent, better-formulated questions when they do engage with the data team.
The Power of Workflows Over Dashboards
A core philosophy at fabi.ai is that data teams can achieve a much higher impact by meeting stakeholders where they already work. Data scientists often spend countless hours building dashboards that are either viewed infrequently or immediately exported to a spreadsheet.
“If you have a choice between creating an automated workflow and creating a dashboard, you should always pick the workflow.”
The return on investment (ROI) for dashboards is notoriously hard to prove, and driving behavior change to get users to adopt them is even harder. In contrast, automated workflows that push AI-generated insights directly into tools like Slack, Microsoft Teams, or email ensure the work is consumed. This approach removes the friction of a user having to remember to visit a dashboard, interpret it, and then take action. The highest ROI comes from workflows that can skip the insight-sharing step entirely and trigger the necessary action directly.
Principles of Building AI-First Platforms
Building a true AI-first product requires more than just adding an AI chatbot on top of an existing application. It involves a fundamental rethinking of the user experience.
- Variable Autonomy: Drawing on Andre Karpathy’s “autonomy scale,” a well-designed platform should allow users to choose their level of AI involvement. A product manager might use a fully autonomous mode to build a data app via conversation, while a data scientist might prefer a semi-autonomous mode where they can manually intervene, review, and edit the AI-generated code. This flexibility makes the tool accessible and powerful for all skill levels.
- Total Reversibility: When working with AI, users will inevitably take a “wrong turn.” Every action the AI takes must be fully reversible. This means implementing features like version snapshotting, allowing a user to view a diff and revert to a previous state if an analysis goes astray.
- Essential Guardrails: To prevent issues like runaway data warehouse costs or inaccurate conclusions, several guardrails are necessary. This includes a caching layer that stores query results as a data frame for subsequent analysis, preventing repeated hits to the database. It is also crucial for data teams to curate and limit the AI’s access to a few core, well-understood data tables, which both bounds the AI and improves the accuracy of its results.
Navigating Data Quality and Metrics
There’s a necessary balance between data preparation and speed. You don’t need a “perfect” data model across the entire organization before leveraging AI. However, for any specific analysis, it’s critical to spend the majority of the time cleaning and preparing the immediate data source. Whether it’s a messy spreadsheet or raw database tables, cleaning the data before feeding it to the AI will yield far more effective and reliable results.
AI also provides a powerful environment for exploring and validating new metrics. Before a new metric is formally adopted and built into a core dashboard, teams can use AI-driven tools to experiment in a “messy phase.” A product manager and data scientist can collaborate to see if a hypothesized metric is actually useful. Once it has proven its value over a quarter, it can be formalized and integrated into the official data models and dashboards.