Explain Predictive Model to Chemists

Company Context

FormuSense is a B2B SaaS company that sells AI tools to industrial formulation teams in coatings, personal care, and specialty chemicals. The product helps chemists predict formulation outcomes such as stability, viscosity, and performance before running expensive lab experiments. The company has 120 enterprise customers and is expanding from pilot programs into multi-year platform contracts.

Problem

Adoption is stalling after initial sales because formulation chemists do not trust the model's recommendations. Internal data shows that while 78% of customer accounts upload historical lab data, only 24% of chemists use model predictions weekly. In user interviews, chemists repeatedly say: "The scores may be accurate, but I don't understand why the model is suggesting this formula." Sales wants a more technical explanation layer, while design wants a simpler confidence and rationale experience.

You are the product manager for the prediction workflow. Your challenge is not to improve model accuracy directly, but to define how the product should explain the model's core mechanism to a formulation chemist in a way that increases trust, usability, and decision quality.

Deliverables

1. Identify the primary user segment(s) and the job they are trying to get done when interpreting a model prediction.
2. Define what "explaining the model" should mean in the product experience, including what information to show and what to avoid.
3. Propose and prioritize an MVP set of explanation features for chemists using the prediction workflow.
4. Define success metrics and an experiment plan to validate whether the explanation experience improves trust and adoption.
5. Discuss key trade-offs, especially between scientific rigor, simplicity, speed, and legal/commercial risk.

Constraints

• MVP must launch within 10 weeks with 1 designer, 3 engineers, and shared support from 1 ML engineer.
• The underlying model is a gradient-boosted ensemble and cannot be replaced this quarter.
• Explanations must work across three prediction types: stability, viscosity, and drying time.
• Customer contracts prohibit exposing proprietary training data from other clients.
• The workflow cannot add more than 2 extra clicks or materially slow prediction generation.