OneDrive photos agent

Problem

Large language models (LLMs), like OpenAI’s GPT-5, change product design. Instead of designing for specific user inputs, we need to account for a wider range of natural language interactions. The OneDrive Photos Agent’s instructions must match the user needs to help find and organize photos effectively.

Approach

Early collaboration with the engineering team helped me understand the technical constraints and opportunities. I worked with user research to identify common photo management tasks and pain points. This informed my design of the agent’s conversational flows and response patterns. I set the product naming and wrote and tested the prompt suggestions. I also structured, rewrote, and tested key instructions for the agent.

Content examples

View full width

Outcomes

While still in limited preview, our user research shows that even the limited initial scope of the agent helps users achieve a key user job in OneDrive, finding photos. Within 2 weeks of the limited release in October 2025, hundreds of people are using the agent to find photos every week.

We continue to add high-value features. User metrics like number of queries per user and retention rate drive the focus of improvement. Direct user feedback guides my design changes.