Keenpipe · AI Sales Agent UX/UI Case Study

Keenpipe reads a company's website, works out who buys what it sells, finds those companies and the people inside them, scores every lead and drafts outreach per lead and per channel. Then it stops and waits for approval. The redesign is about that hand-over from machine to person: making it fast, legible and safe.

The audit

I audited the live product screen by screen and logged 57 coded findings. Sorted by theme (control, clarity, language, recovery), they pointed to one missing quality: trust. A single "Execute" button sent real LinkedIn requests with no confirm, no undo and no sender shown; three counters on one screen disagreed; raw machine states reached the screen.

What I did

• Six principles, written before any screen, that became the developer's definition of done
• A new information architecture and a naming dictionary: one word per object
• A design system: ten colours, a contrast-checked text ladder, ten type sizes and 17 pass-zero components, plus nine additions and three accessibility fixes to existing tokens
• The approval card: person first, a score with four sourced checks, every claim checked against evidence, a 5-second undo instead of a confirm
• A recovery pattern: every stop names its cause, consequence and fix, in eight sizes
• Key flows: agent creation, company profile filled by AI, voice with live preview, reply to meeting
• Four states for every page and three breakpoints, 1440 · 1024 · 768
• Five open decisions drawn both ways with their cost
• The keenpipe.com landing page, where the hero runs the product live
• Developer handoff: specs, routes, data shapes and a 10-point definition of done

Success metrics (a daily batch cleared in under 20 minutes, a profile saved in under 5, zero silent failures) are post-launch targets, not results.