AGILITAS

Designing the MVP for an AI-driven food reformulation product

I led the design and research for a 0-1 MVP launch for Agilitas, an early stage startup building a B2B web product for food & beverage companies.

AGILITAS

Designing the MVP for an AI-driven food reformulation product

I led the design and research for a 0-1 MVP launch for Agilitas, an early stage startup building a B2B web product for food & beverage companies.

Role

Founding Product Designer

User Research

Timeline

January - June 2024

Team

3 Founders, 1 Product Manager, 1 Designer, 4 Developers

Team

3 Founders, 1 Product Manager, 1 Designer, 4 Developers

Role

Founding Product Designer

User Research

Timeline

January - June 2024

Team

3 Founders, 1 Product Manager, 1 Designer, 4 Developers

OUTCOMES

Agilitas successfully launched, reducing the food reformulation process from months to weeks with AI-driven features.

3

paid multi-year customer contracts

10x

growth in active users, from 10 to 100s

75%

of users completed primary task of uploading and modifying a formula

PROBLEM

Food scientists are underserved by current tools, forced to manually upload or sift through paper documents.

PROBLEM

Food scientists are underserved by current tools, forced to manually upload or sift through paper documents.

Right now we have to manually input data into spreadsheets, which takes us hours to get started.

Chris A. Head of R&D

Right now we have to manually input data into spreadsheets, which takes us hours to get started.

Chris A. Head of R&D

I have a binder full of ingredient papers that I look through when I need to find alternative ingredients

Beverly M., Sr Food Scientist

I have a binder full of ingredient papers that I look through when I need to find alternative ingredients

Beverly M., Sr Food Scientist

The user journey in the food development process:
The user journey in the food development process:

Define project brief

Analyze formula, ingredients, or process

Make adjustments

Test

Finalize formula

Define project brief

Analyze formula, ingredients, or process

Make adjustments

Test

Finalize formula

FIRST HYPOTHESIS

We hypothesized that an ingredient comparison tool would save food scientists significant time in their workflow, modernizing the ingredient analysis process.

The ingredient comparison tool is comprised of two components:

Summary

An overview of the table, including a summary of what to watch out for when choosing ingredients to substitute in the formula.

Comparison Table

Compare attributes of up to 4 ingredients, specific details of each are extracted from uploaded ingredient specification documents.

After conducting 8 user research interviews, a bigger pain point was uncovered - the trial and error process in perfecting the formula.

After conducting 8 user research interviews, a bigger pain point was uncovered - the trial and error process in perfecting the formula.

Depending on the product, it can take me over 100 tries to get the ingredients and ratios right.

Skyler T., Food Scientist

Depending on the product, it can take me over 100 tries to get the ingredients and ratios right.

Skyler T., Food Scientist

What I care about most is how these ingredient substitutions are going to impact my formula"

Mackenzie M., R&D Engineer

What I care about most is how these ingredient substitutions are going to impact my formula"

Mackenzie M., R&D Engineer

PIVOT TO SECOND HYPOTHESIS

Based on the research insights, we pivoted to a hypothesis that an outcomes-based formula modification tool would expedite the food development process.

The first iteration post-pivot.

Jumpstart prompts

Food scientists can make AI-assisted changes using common formula modifications.

Manually modify formula

Providing the option to make manual changes to the formula, including the ingredients list or proportions

Nutrition Label

View how the changes to the formula impact the end-result nutrition label.

After a few rounds of usability testing, there were 4 key changes to the product experience.

Prioritize project goals

As part of the project brief, food scientists have to achieve nutritional or regulatory targets in their formula.

The formula targets is at the very top so that food scientists know what their targets are, and whether that specific formula version meets those targets or not.

Version control

Each formula goes through multiple versions, with tweaks in the ingredients, proportions, or preparation steps.

Food scientists like to toggle between different versions to view what has been changed from version to version, so that they know whether they are closer to their formula targets.

Open-ended chat with AI

In the first version, users were completely skipping over the jumpstart prompts. 

Rather than constraining to specific types of modifications, an open-ended chat aligns with current AI design principles and supports a wider range of modifications.

Support for multiple units

Although grams is the most common unit of measurement, food scientists who work with beverages work in mL.

Separating the quantity and the unit supports formulas that have more than one unit of measurement.

The final iteration post-pivot.

Reflections from designing with AI

Reflections from designing with AI

Transparency is key for building trust

Let users know when content has been generated by AI, so that they know to fact check it or review it. Trust in AI output is still very much a work in progress.

Transparency is key for building trust

Let users know when content has been generated by AI, so that they know to fact check it or review it. Trust in AI output is still very much a work in progress.

Design for AI feedback loops

While working with AI engineers, I learned that the best way to improve AI output is through feedback loops. A common design pattern in today's AI tools is the opportunity to regenerate. Rather than allowing users to regenerate endlessly, disabling the button until a reason for regenerating was given helped improve regenerated output.

Design for AI feedback loops

While working with AI engineers, I learned that the best way to improve AI output is through feedback loops. A common design pattern in today's AI tools is the opportunity to regenerate. Rather than allowing users to regenerate endlessly, disabling the button until a reason for regenerating was given helped improve regenerated output.

High quality output requires high quality input

Clear, intentional guidance through detailed prompts or specific objectives helps ensure AI delivers results aligned with user expectations.

High quality output requires high quality input

Clear, intentional guidance through detailed prompts or specific objectives helps ensure AI delivers results aligned with user expectations.