Designing Compass: UX for an agentic AI research assistant
Sole designer on a new category of AI tool, from the conversational interface and clarifying questions through to the branded deliverables it produces.
Compass is Voxpopme's agentic AI research assistant. It takes a question, searches a customer's research, analyses their video responses and returns an answer. As the only designer at Voxpopme I shaped the whole experience, from the conversation and onboarding to the microcopy and the deliverables it generates.
The core of the work was getting Compass to give people the answer they actually wanted on the first try. Answering fast was easy, however a fast answer often missed what the user was really after, so I dug into how good answers are structured as stories and designed clarifying questions that quickly check the user's goal, which means people get what they want straight away rather than a quick answer that misses the mark.
Customers drowning in their own data
A beta call with an enterprise customer summed it up in one phrase, a "cloud of selfie videos," hundreds of hours of footage with no fast way to find the insight inside it.
Compass existed to turn that cloud into an answer, so the design job was to make that feel effortless and trustworthy from the very first session.
Clarifying questions that check the goal
Early behaviour told me two things at once.
When Compass just answered as fast as it could, people got a quick response but often not the one they needed, so the output looked impressive without being useful. When it opened with a clumsy mandatory question before doing anything, people read it as the tool not being ready and backed out to plain AI Chat. Neither pure speed nor a blocking gate solved the real problem, which was understanding what someone actually wanted without making them work for it.
So I approached it as a storytelling problem rather than a speed problem, and I did the research to back it up properly. I built a research repo of secondary sources on how a genuinely useful answer is structured around the user's underlying goal, spoke directly to customers about what they were really trying to get out of Compass, and used Claude to search back through past customer calls for the same signal. From all of that I designed clarifying questions that check the goal efficiently and in the user's own language, so the very first answer is shaped around what they were actually trying to find out. The questions stopped being a gate and became the thing that makes the answer land, which means people get what they want straight away.
Designing the outputs, not just the conversation
An answer is only as good as the thing the user walks away with, so I also designed what Compass produces.
I wrote skills that generate properly branded deliverables built to tell a story rather than dump data, including a PDF findings report and presentation outputs, which means a user can take a Compass answer straight to their stakeholders without reformatting it first. The same storytelling thinking that shaped the clarifying questions runs through here too, since the goal throughout is turning research into something someone can act on and share.
Waiting and onboarding
Compass runs a multi-step process, understanding the question, searching projects, analysing video and preparing a recommendation, all of which takes time.
Understand
Interpret the question being asked.
Search
Find the relevant research projects.
Analyse
Work through the video responses.
Recommend
Prepare the answer and its evidence.
So I wrote copy for every loading and async state to reduce anxiety rather than just fill silence, including "come back later" messaging for longer tasks. I also rewrote onboarding after the behaviour data showed people getting a good first answer and then immediately creating a project, missing what made Compass valuable, so the new introduction conversation and product tour teach the value rather than walk through features.
A first month that made the case
Compass's first full month of enterprise usage showed the design was landing.
Most users produced a tangible deliverable, most came back without being prompted, and returning users came back repeatedly. Adoption settled into two healthy patterns, which gave the business a much clearer picture of who Compass was really for:
Across large accounts
Many people in an organisation, each using Compass lightly. Broad reach, lighter individual depth.
Individual power users
A smaller number using it intensively and often. Narrower reach, but deep engagement.
Compass first
Compass started as a way to search research, however the direction now is Compass first, where the assistant becomes the way users do everything.
From launching projects to making showreels, directly from whichever LLM they already work in, such as Claude or Copilot. That shifts the design problem from "answer a question well" to "let someone run the whole platform through conversation," so the goal-checking and storytelling groundwork here becomes the foundation for a much bigger surface rather than a single clever feature.
What I learned
Agentic AI has no established design patterns, so the interesting problems are new ones. The real tension isn't speed versus quality, it's that a fast answer and a useful answer aren't the same thing, and the way to close that gap is to understand the user's goal cheaply rather than either guessing or interrogating them.
It also reinforced that UX writing is product design, since on a solo team the words in a clarifying question, a loading state or a findings report carry as much weight as any layout decision.