This is how I do Agentic Design at Mercury

By Sneak Peek Design

Categories: Design, Product

Summary

Mercury's design team uses high-level prototyping and deterministic confirmation patterns to make agentic banking experiences intuitive—requiring explicit user confirmation buttons instead of conversational interactions to ensure intentional action-taking, a critical safety pattern for fintech AI products.

Key Takeaways

  1. Start agentic design with loose, high-level prototyping using simple diagrams to ground understanding of how LLMs will process context, think through problems, and output legible results without requiring customers to do manual calculations.
  2. Implement deterministic confirmation via explicit buttons rather than conversational flows when designing agentic financial transactions—this prevents casual interactions and ensures users intentionally authorize actions like money transfers.
  3. Ship all product features live in a public demo environment accessible to customers, non-customers, and prospective employees—this accelerates design feedback loops and serves as a powerful recruiting tool for attracting designers who can evaluate the product before interviewing.
  4. Design for dual mental models in fintech: combine search and chat capabilities into unified interfaces while maintaining separate workflows for critical actions (transactions) and exploratory analysis (insights) to match user intent and cognitive load.
  5. Ground executive stakeholder reviews with step-by-step Figma walkthroughs rather than whiteboard prototypes or Slack links—this ensures shared mental models for complex agentic workflows before development begins.

Related topics

Transcript Excerpt

When you first started working on this project for designing agentic experiences, like what was the first few designs look like? What were your first few concepts like? >> So, I think at that time it was a lot of what is this new mental model that we want to work on? And just I think doing a lot more high-level prototyping and a lot the concepts were a lot looser. So, one of the things that I was doing was like oh, what if we combined search and chat into one? And I drew this little diagram. And I don't know if anyone else looked at this or thought that it was helpful, but it was really helpful for me to kind of understand what the system was doing. This was a grounding moment for me where I was like, okay. What the LLM can do is it can just get context, right? And then take that all in an…

More from Sneak Peek Design