AI Assistant for Food Supply Chain Management, Food Safety and Compliance
My role
User research, prototyping, UI design, Product design, Usability testing, Service Mapping
Results
Significant improvement on customer independence and satisfaction for early adopters
Foods Connected is an End-to-End Supply Chain Management Software for Food Safety, Compliance, Procurement and CSR platform used by teams across the food industry. The platform holds large volumes of data across multiple silos — specifications, questionnaires, and supplier documents — but finding a specific piece of information meant manually navigating between sections, requesting reports from support teams, or waiting for scheduled BI exports that could be up to 24 hours old.
An early version of the chatbot existed as an innovation project, but it returned raw data rows that users had to interpret themselves, and could only search one silo at a time. It was not ready for production use
The highest risk assumptions centred on trust specifically whether users would rely on AI-generated answers for compliance sensitive queries like allergens, and whether they understood that a missing answer was not the same as a negative one.
One of the most important pieces of work was mapping what happens across the full service
Four distinct failure states emerged, each with a different cause, a different owner, and a different resolution path.
Each had a different owner and a different resolution, only one of them was actually a chatbot problem.
The four failure states shaped the service design: the chatbot not finding information, data that was never entered, restricted access with no clear escalation path, and wrong data being returned confidently.
I produced and prototyped user flows for the core scenarios, covering the full range of states
Key design decisions included showing sources alongside every response by default, with links back to the originating document and a future path toward inline highlighting.
Empty states were designed to distinguish clearly between no data existing and a retrieval failure. For allergen and compliance fields specifically, the wording was defined, returning “no information found” rather than implying a safe negative result.
I also advocated against the early plan to open the chatbot in a new browser window. The shipped solution uses a full-screen sheet with collapsible panels, keeping users in context while querying across data sources
A demo mode, where the AI generates example conversations using the customer’s own dataset, was added to reduce the learning curve and support onboarding and adoption
Conversation logging is in place with plans to run sentiment analysis and identify friction clusters to drive ongoing improvements.
The demo mode has been adopted as part of the customer onboarding process, improving early engagement and reducing time to first value.
Beyond the measurable outcomes, one of the most significant changes was a shift in user mental models.
Early in the project, users approached information retrieval as a navigation task, they expected to move through the system, locate the right section, and find what they needed themselves.
Initial testing showed that users defaulted to short, keyword style queries, treating the chatbot like a search bar rather than a conversational tool, and became frustrated when vague inputs returned broad or unhelpful results. The demo mode was a direct response to this, giving users worked examples from their own data to demonstrate what a well formed prompt looked like and what the chatbot was capable of returning
By the time the feature shipped, a measurable preference had emerged. Users who had moved past the initial learning curve consistently preferred asking the chatbot over navigating the system manually, citing the reduction in steps and the ability to ask follow-up questions in the same thread without starting over.
This also surfaced the ongoing design assumption the chatbot must continue to reward natural, conversational input rather than train users into rigid query patterns. As the feature evolves, keeping that interaction model intuitive will be as important as expanding what the chatbot can do
