top of page

AI Predictive Inventory Forecast
MVP Foodservice Supply Chain 

TL:DR
The Problem

Foodservice forecasting happens quarterly and takes a ton of time; manually reviewing, including, and excluding historical data from several years 

The Solution

Create a Predictive AI model to create forecasts and embed in our current Inventory Alerts product

The Impact

Our model was 5-8% more acccurate depending on the brand and enabled a more granular view of the data

My Role

End-to-end Designer, User Researcher, and Cross-Team Facilitator

Context: Company, Users, & Project

Company

  • Arrowstream is a Foodservice B2B Supply Chain Management SaaS product

  • Multi-persona platform with Brands, Distributors, and Suppliers utilizing different products

  • Complex and data-heavy, ingesting and mapping data from the distribution center feeds, supplier feeds, and user uploads

  • Helps users make critical decisions in their supply chain with focused reporting 

Users

  • Users describe forecasting as "a science and an art"

  • They value high data visibility

  • Users are split on AI, with most leaning towards caution with a few very excited outliers

Project

  • Procurement is focused on making sure distribution centers and stores have the inventory they need

  • Items each have varying levels of priority for alerting, lead time to receive, usage, and base levels of stock expected

  • Forecasts are affected by lots of factors:

    • Geographic location

    • Seasonality

    • Item Attributes, like shelf life

    • Ability to be substituted 

  • Forecasting is extremely important with little margin for error

  • This project is an additional feature in our core product

Project Scope: Team, Timeline, My Role & Constraints

Team & Timeline

  • 2 teams I worked across

    • AI Team: AI Engineer, PM, and me

    • Central: PM, 2 engineers, and me

  • This is an MVP release and an ongoing project

    • 2 weeks on Discovery, and User Research

    • 1 week on Ideation, Scoping, Design, and User Validation

    • 1.5 months on Development split between teams

My Role

  • End-to-end design

  • Qualitative user research

  • Collaborative scoping of product requirements

  • Experience design

  • Prototyping and Usability Testing

  • Dev Handoff

Constraints

  • ​Leadership initiative with a fast timeline

  • Scope shifted from a standalone UI to integrating this model into an existing UI

  • Forecasting was limited to the Top 100 items per brand due to server and data storage constraints

Impact: Increased Forecast Accuracy, Easier Data Inputs, & Actionable Alerts

Forecast Accuracy Increased by 5-8%

  • Beat the average accuracy of 85-90% based on brand on the Top 100 Items

Anamolous Data was surfaced by the model for exclusion

  • Less time spent looking for outliers. Moved from a manual process to a one click approval process. 

Actionable Alerts Driven by AI

  • AI model powers our existing Inventory Alerts, surfacing issues and providing suggested actions

Discovery: User Needs & Model Design

3 External Brand Interviews

Goal: Understand brands' current forecasting methods and goals

Aha! Moment: Our biggest limitation is brands' wanting higher accuracy on items with low or no history

  • At what level are brands forecasting, and how are they generating that forecast?

  • What is their current forecasting accuracy?

  • What are their biggest forecasting pain points?

  • How would they track the effectiveness of a new model?

12 Internal Meetings

Goal: Check in on the progress of the AI model using customer data and deal with handoff between teams

Aha! Moment: We can release faster and still meet user needs by using the existing UI

  • What is the difference in accuracy we can get by using the model?

  • What are the limitations we have in user inputs? 

  • How are we dealing with handoffs between teams?

  • What are each internal team's responsibilities? 

Results: MVP Scope, Future State, & Not Feasible

MVP Scope

  • Beat the average accuracy of 85-90% based on brand

  • Model ingesting historical data and creating a daily usage forecast, this is limited to the top 100 items due to server and data constraints

  • Surfacing anomalous data for the user to act on, reducing the need for manual review

  • Provide actionable insights in our existing UI

Future State

  • Allow users more control over which items are forecasted on

  • Integrate APIs around weather or event information

  • Show users more insight into the logic behind the forecast

  • Side-by-side comparison between AI & traditional models 

  • Allow users to set more custom alerting thresholds

Not Feasible/Potential Workarounds

  • Forecast completely new promotions, items, and stores  

    • Could have the user find a comparable if available​

  • Forecast on Tail Skews (Low volume items)

    • This one is tricky with so little data, needs more research​

Design: Integrated into an existing UI

Screenshot 2026-05-01 at 1.25.43 PM.png
MVP Scope

  • ​The AI model is available on the Configuration Settings, and "Powered by AI" is displayed when in use​​​​

  • The notification bell surfaces anomalous data to users to exclude from the model

  • Potential Issues: Icons that link to the relevant record in the system are colored based on severity. An actionable suggestion is listed. 

  • All Confidential Information has been redacted.

MVP Release:
Our model accuracy is currently 5-8% higher, depending on the brand. We launched our MVP for all customers. We have 6 early adopter brands and are working on gathering feedback for the next iteration. 

Key Learnings: What I Would Do Differently

1. Have users create an Item Group before building the model

We focused heavily on higher accuracy for Proof of Concept, so we selected top usage items that we had lots of data for. Users often already have their most crucial/core items forecasts down.

2. Be more explicit in communication with users on AI

Our users are not all clear on the differences between generative and predictive AI. Beginning discussions with level-setting information and providing more information in the product would help users feel more comfortable utilizing the model. ​

bottom of page