Designing an AI manager copilot so managers can lead, not administrate.
I led the product and UX strategy for ARVO, an AI manager copilot. Before any design work started, our team needed to know: was this even a design problem?
ARVO's vision was right, but adoption was low. Before opening Figma, the team needed to understand whether it was entirely a design flaw at all.
10 of 12 managers spent 70–90% of their week on execution. A dashboard that required reading before it let you act wasn't helping.
Critical team alerts sat below paragraphs of passive copy with no visual hierarchy. 8 of 12 managers skimmed past them entirely.
"Micro-experiment," "Emotional Labor," "Privacy Threshold" confused 11 of 12 managers. The platform was speaking HR textbook. Its users were thinking in deadlines.
Before jumping right into tackling a UI redesign, we needed to figure out how ARVO is suppose to effectively supplement managers' workflow without adding more admin.
Round 1 tested the original platform. Round 2 tested our first attempt at fixing it. Each round told us what was broken, then what we'd got right, and what we'd missed.
In round 1, we asked 12 managers across 8 industries how managers handle people-development today, and where does their trust around anonymous data sit? Then we gave them the original ARVO platform and three tasks. The most revealing was a timed 3-minute check: extract for the most critical team insight before your next "hypothetical" meeting. Most ran out of time.
Five issues resulted and while two were easily resolved by design the rest aren't on the screen.
Research told us what managers didn't like. Still it pointed us in the direction of what good design could possibly look like. These five principles came from what managers said and became the rules every design decision was tested against.
What managers wanted
01 3 minutes read or it'll be ignored
Manager's Paradox — 10 of 12 managers care about their team but spend 70–80% of their week on execution. So this report cannot take more than 3 minutes or it'll be pointless to use.
02 Scan first. Read never.
Cognitive method — 8 of 12 managers skimmed past alerts entirely because it was buried below dense rows of passive reporting.
03 No jargon. Manager language only
Jargon & management 101 — 11 of 12 managers were confused by complex terms like 'micro-experiment', 'emotional labour' and 'privacy threshold'. For a quick read it left managers with more questions.
04 Earn trust before asking for it
Heavenly Gap — 9 of 12 managers were confident about being able to identify who wrote comments/survey responses; anonymous data is only useful if people believed it is.
05 Visible and actionable, or neither
Structural Resignation — 7 of 12 managers were frustrated to be shown systemic problems where they had no authority to fix. Less control, more stress, no improvement to visibility.
Instead of returning with wireframes, I ran a co-creation workshop. We mapped each friction point against business impact and design feasibility together with our client. Involving the client mid-process helped build the diagnosis and the final recommendations don't land as a surprise. As Team Lead, my goal wasn't sign-off. It was shared ownership of the problem space.
The first iteration applied what Round 1 had taught us — single-page architecture, traffic-light signals, energisers elevated. We weren't done just yet, we needed to test it again to iron out any unforeseen gaps.
First iteration — Improved structure, unresolved friction
Version one tested
13 managers. 4 evaluative scenario tasks. The Manager's Paradox and Cognitive Overload were resolved. The timed 3-minute baseline — our hardest task — passed for 12 of 13 managers. However, two new issues came through clearly that hadn't surfaced in Round 1.
7 of 13 managers couldn't read the profile page. Light-gray text on white failed WCAG AA. UI fix applied in the final iteration — full product audit still needed.
"It's gray, so it's not so visible for me... unless I have a bigger screen." — Manager, Retail
9 of 13 found the weekly checklist a chore regardless of content quality. Hence we applied a calendar integration in the final iteration where the underlying cadence model is a product decision.
"Take away the mark as done, please. It's crazy." — Manager, Energy
The 3-minute baseline. Every design decision was tested against one scenario: can a manager extract their most critical team insight and take action before they're rushing to their next meeting? The final iteration fixed what Round 2 confirmed was broken. Three decisions — each one earned by a finding, each one protecting what Round 2 had already validated.
