The AI in the Room — Privoce

Problem

Remote work created two gaps: individuals lost a clean memory of what they accomplished, while teams still needed progress visibility. AI could help with both — but only by observing work.

Role

Lead Product Designer · defined the consent model · designed the AI worklog end-to-end

Team

Founder + engineers

Timeline

Oct–Dec 2025

Platform

Web app · desktop

Outcome

Shipped Dec 2025 · live today

Key Contribution

Consent model · worklog anatomy · source controls · Screenshot Blur

Visibility was already a problem. Surveillance wasn't the answer.

Hybrid work created a visibility gap: employees felt productive, while leaders weren't always convinced. Microsoft's Work Trend Index captured the disconnect:

87%

of employees said they were productive

12%

of leaders had full confidence their teams were

But the opportunity wasn't only managerial visibility. For an individual worker, a passive worklog can become memory infrastructure: what did I work on, where did my time go, what did I finish, and where should I pick back up tomorrow?

AI offered a new answer for both sides: observe selected work signals and turn them into a useful record automatically.

The same mechanic that makes the feature useful also makes it dangerous: it watches your screen.

That created the product tension:

Can ambient AI work for the individual first — then create team awareness without turning into bossware?

Try it — the live demo

The AI Analysis demo mid-run — work-log entries generating beside the shared screen

Start the demo Opens full screen · runs in your browser · no real AI, no capture

A working recreation of the shipped feature — opt in, watch the log write itself, toggle a data source and see the next entry change, open a teammate's report. The numbered dashed pins annotate the design decisions.

The design decision

The easiest version of this feature would have minimized friction: turn it on automatically, capture everything, generate a perfect report.

I designed the opposite defaults:

Off by defaultnothing is captured before the user agrees

Permission per sourceusers control what contributes to the worklog

Visible capture stateobservation is never silent

Source-cited outputusers can trace what the AI believes back to what it saw

Screenshot Blurteammates get enough evidence to understand the work without receiving the raw screen

AI Analysis Settings dialog — what the AI will do, written above an enable switch — Nothing is captured until this dialog is answered — the switch sits under a plain description of what turning it on means.

The AI Work Log Source Data row — Share Screen, To-do List and Time Spent as separate checkboxes — Share Screen, To-do List, Time Spent — three permissions, not one switch.

The shared screen while analysis runs — a Snapshot analyzed toast at the top and a persistent AI capturing badge at the bottom — A badge that stays up the whole time it watches, and a toast the moment each snapshot is taken.

One worklog entry expanded — source chips, then a panel naming the snapshot, the sources permitted at capture time, and who can see it — Every entry opens to name its snapshot, the sources permitted when it was taken, and what the AI could not see.

A teammate's worklog opened read-only — the summary is legible and a line reads Screenshot Blur: on — A teammate reads your summary read-only. With blur on, any stored screenshot reaches them blurred.

That adds friction to a feature whose promise is effortless automation.

But here, friction is part of the trust model.

The product still creates shared awareness; it just makes the boundary between "my team can see my progress" and "my team can inspect my screen" explicit.

Outcome

Shipped December 2025 and remains live in VoceSpace.

–40%

time spent on status reporting · 50 → 30 min per person per week

64%

of users still opted into AI Analysis after 30 days

60%

active team adoption

Together, these three metrics would test the full product thesis:

Does it save time? → Do individuals keep it on? → Does adoption become dense enough to work at the team level?

What the AI can — and can't — know

VoceSpace already combines screenshots, todos, time statistics, and historical context. That is enough to produce factual activity logs. But the harder product ambition is to connect those observations to intent, progress, blockers, and collaboration. The next frontier isn't better description. It's knowing when the AI has enough context to connect the dots — and when it should ask instead of guess.

Verified delivery signal · 12 weeks — concept to shipped

Craft notes · the VoceSpace UI library, extracted from source

Next project

Built for Everything — Napoz

→