AI-assisted formulation workspace for 4,500+ scientists at L’Oréal
Challenge overview
L’Oréal Research & Innovation develops thousands of cosmetic formulations every year across skincare, haircare, and makeup. Scientists explore complex ingredient combinations, evaluate predicted properties, and validate results through laboratory experimentation.
To accelerate this process, L’Oréal introduced an AI-assisted platform capable of predicting formulation outcomes based on historical research data and scientific models.
The challenge was not only to integrate AI into the workflow, but to make predictions understandable, trustworthy, and usable by scientists who needed to remain fully in control of their decisions.
Role
Scope
Senior Product Designer
AI-assisted formulation workflows, data visualization, and decision support interfaces.
I collaborated closely with scientists, product managers, engineers, and data scientists to design a workspace that allows researchers to explore formulations, interpret predictions, and evaluate options with clarity.
Impact
Platform adopted by 4,500+ scientists globally
Up to 28% productivity improvement in formulation exploration
Faster comparison and evaluation
of formulation options
Product & System Constraints
Designing this platform required balancing scientific workflows, AI model capabilities, and enterprise system constraints.
Scientists were working with highly complex datasets including ingredient properties, historical formulations, and predicted product performance. Any interface needed to support detailed analysis without overwhelming users.
Several factors shaped the product design:
AI predictions were generated from complex scientific models that needed to remain interpretable to researchers
The platform had to integrate with existing research workflows and laboratory validation processes
Scientists required full visibility into data sources and assumptions before trusting model outputs
The interface needed to display large volumes of formulation data while remaining readable and navigable
These constraints influenced how predictions were surfaced, how information was structured, and how much control scientists retained over the workflow.
Anatomy of the AI tool's core page: Suggestions and predictions of raw material effects for substituting ingredients in a formula
Predictive data
Relying solely on the color proximity criterion is not enough.
It is necessary to have predictions on other key criteria that can help users make their choice, such as:
Fading: the gradual loss of intensity or vibrancy of hair color over time
Selectivity: the ability of a product to specifically target certain parts of the hair while preserving others.
The proximity to the source
The user has an overview of the projection of a new blend.
They can even compare the impact on the product's color that they 'clean' at the time of application (T0) and an estimate of the normal color degradation 15 days later (T15).
The other important attributes.
For any chemical formula, there is a list of attributes that are also important for making a choice, notably including a prediction of the cost of the new formula with the new ingredient and its eco SPOT score.
The Solution
The final product is a formulation workspace designed to help scientists explore multiple formulation options and understand predicted outcomes before moving to laboratory testing.
Instead of positioning AI as a black-box recommendation engine, the interface focuses on transparency and exploration. Scientists can review predictions, examine underlying assumptions, and compare alternative formulations within a single environment.
The platform enables researchers to:
• explore different ingredient combinations
• compare predicted formulation performance side by side
• understand how predictions were generated
• validate or adjust AI suggestions before experimentation
The goal was not to automate scientific work, but to augment it by helping scientists evaluate more possibilities in less time.
When I joined the project, an early proof of concept had already been built by engineers in partnership with IBM. In parallel, UX research had been conducted to understand how scientists actually create new cosmetic formulations.
The research revealed an important reality: formulation work is deeply rooted in the physical world. Scientists spend most of their time experimenting at the laboratory bench, and digital tools were only lightly integrated into the process.
The challenge was therefore not to redesign the workflow, but to augment it carefully with AI-assisted insights.
My work started with reviewing the research findings and mapping the existing formulation journey to identify where predictive models could support scientists without disrupting their established practices.
Introducing AI into an existing scientific workflow
Structuring the formulation journey
Through collaborative workshops with product, research, and data science teams, we translated research insights into simple jobs-to-be-done personas and mapped the key stages of the formulation process.
Focus AI where it creates real value
Rather than digitizing the entire process, we focused on specific steps where predictions could accelerate exploration without disrupting existing laboratory practices.
Making scientific data usable
Working closely with data scientists, we reviewed the scientific datasets behind the predictive models to determine what information actually helps scientists evaluate a formulation.
Early prototypes surfaced large volumes of model outputs and parameters, which quickly became overwhelming. Through iterative testing with researchers, we refined the interface to focus on the key indicators needed to assess a formulation, while keeping deeper scientific data accessible when needed.
Show the right data at the moment of decision
Instead of exposing all model outputs, the interface prioritizes the signals scientists need to evaluate a formulation.
Additional parameters remain available when needed, but the primary workspace focuses on the information that supports faster and clearer decision-making.
Iterating with real users
The platform was redesigned in Figma based on the existing engineering implementation. Early prototypes were iterated closely with engineering and data science teams to ensure feasibility and data clarity.
I then conducted lean usability tests with scientists, validating small parts of the workflow while developers progressively integrated improvements into sprint cycles.
