Arnabot

Arnabot

Arnabot

AI Project Management Assistant

AI Project Management Assistant

AI Project Management Assistant

AI assistant that turns unstructured team conversations into project management actions—eliminating 4 hours of weekly overhead per PM

AI assistant that turns unstructured team conversations into project management actions—eliminating 4 hours of weekly overhead per PM

My Role

My Role

My Role

Design Lead

Design Lead

Design Lead

Timeline

Timeline

Timeline

Q2 2025

Q2 2025

Q2 2025

Team

Team

Team

1 Designer, 1 AI engineer , 2 DevOps

1 Designer, 1 AI engineer , 2 DevOps

1 Designer, 1 AI engineer , 2 DevOps

Overview

Overview

Overview

Background

I joined this project after connecting with an AI engineer who was frustrated by how much time his team spent updating project management tools instead of building. We assembled a small team (1 designer, 1 AI engineer, 2 DevOps) to solve a problem we all experienced firsthand: teams were drowning in PM overhead, spending 4+ hours weekly organizing work across scattered tools rather than executing it.


Challenge

  • Small teams and individual contributors were drowning in project management overhead. Traditional PM tools required too much upfront structure, while async collaboration left work scattered across docs, Slack threads, and personal notes.



The core problem: teams spent more time organizing work than executing it.

Background

I joined this project after connecting with an AI engineer who was frustrated by how much time his team spent updating project management tools instead of building. We assembled a small team (1 designer, 1 AI engineer, 2 DevOps) to solve a problem we all experienced firsthand: teams were drowning in PM overhead, spending 4+ hours weekly organizing work across scattered tools rather than executing it.


Challenge

  • Small teams and individual contributors were drowning in project management overhead. Traditional PM tools required too much upfront structure, while async collaboration left work scattered across docs, Slack threads, and personal notes.



The core problem: teams spent more time organizing work than executing it.

Research
& Discovery

Research & Discovery

Research
& Discovery

What I did

  • Surveyed 47 PMs managing AI development projects at tech companies (Series A-C stage)

  • Analyzed 4 leading PM tools (Jira, Monday, Trello, Agile)

  • Conducted secondary research analyzing industry reports, articles, and blogs on project management trends and challenges

What I did

  • Surveyed 47 PMs managing AI development projects at tech companies (Series A-C stage)

  • Analyzed 4 leading PM tools (Jira, Monday, Trello, Agile)

  • Conducted secondary research analyzing industry reports, articles, and blogs on project management trends and challenges

Key findings

  • Unstructured input is the norm

    83% of teams start with Slack threads, whiteboard or voice notes. Industry data confirms ~80% of organizational information is unstructured, yet tools like Jira force immediate structure, creating friction at project inception. 🔗


  • Context switching costs 4 hours/week

    My survey found teams lose 45min/day toggling between tools. Broader research shows workers switch apps 1,200+ times daily, losing ~4 hours weekly. 🔗


  • AI projects need different primitives

    Data science involves extended research and unknowns, unlike traditional dev work. Teams track experiments and model iterations. Organizations adapting agile for AI work must customize fields or use separate experiment tracking tools. 🔗


  • The "planning theater" problem

    PMs invest 2+ hours crafting detailed plans to satisfy stakeholders, yet reality forces immediate updates. In my interviews, teams consistently reported plans becoming outdated within days. They wanted lightweight, living plans — Kanban boards that evolve daily, not static Gantt charts that don't survive first contact with execution.

    "Responding to change over following a plan."


  • Collaboration happens in comments, not tasks

    most critical decisions like clarifications, implementation choices & status updates, lived in comment threads, not task fields. Teams tag each other, discuss, and decide within comments. Tools weren't designed for this conversational workflow.

Key findings

  • Unstructured input is the norm

    83% of teams start with Slack threads, whiteboard or voice notes. Industry data confirms ~80% of organizational information is unstructured, yet tools like Jira force immediate structure, creating friction at project inception. 🔗


  • Context switching costs 4 hours/week

    My survey found teams lose 45min/day toggling between tools. Broader research shows workers switch apps 1,200+ times daily, losing ~4 hours weekly. 🔗


  • AI projects need different primitives

    Data science involves extended research and unknowns, unlike traditional dev work. Teams track experiments and model iterations. Organizations adapting agile for AI work must customize fields or use separate experiment tracking tools. 🔗


  • The "planning theater" problem

    PMs invest 2+ hours crafting detailed plans to satisfy stakeholders, yet reality forces immediate updates. In my interviews, teams consistently reported plans becoming outdated within days. They wanted lightweight, living plans — Kanban boards that evolve daily, not static Gantt charts that don't survive first contact with execution.

    "Responding to change over following a plan."


  • Collaboration happens in comments, not tasks

    most critical decisions like clarifications, implementation choices & status updates, lived in comment threads, not task fields. Teams tag each other, discuss, and decide within comments. Tools weren't designed for this conversational workflow.

PMs don't need another PM tool. They need their existing conversations to become their project management system.

Ideation

Ideation

Ideation

The research pointed to a clear tension: teams needed less friction, but AI that acts autonomously creates its own kind of friction — anxiety. These questions shaped the exploration:


  • Teams don't start in tools — they start in conversation

    How might we meet them there, without adding yet another place to manage?

  • The overhead isn't the work — it's the translation between where decisions happen and where they're supposed to live.

    How might we reduce the gap between "we decided" and "it's documented"?

  • AI that acts without asking doesn't feel helpful — it feels like a liability.

    How might we design suggestions that feel auditable, not authoritative?


Three approaches were tested to answer that last question:

Real-time suggestions

AI surfaces updates continuously as conversations happen, with constant notification stream.

  • "Constant pings broke focus. Users started ignoring all notifications."

Active collaborator

AI is always visible in the interface, ready to assist at any point during work.

  • "Having AI always present felt intrusive. Users wanted to control when to engage."

Passive Listener

AI listens silently in the background, surfacing a single summary only when a natural break occurs — meeting ends, thread closes.

"It felt like a colleague who had been in the room the whole time, not a bot that kept interrupting."

The passive listener model shaped every design decision that followed. If AI speaks only at natural breaks, the interface needs a place for those summaries to land — and a way for users to act on them quickly, without switching context.

Solutions

Solutions

Solutions

Side Panel —
Work Where Teams Already Are



Side Panel — Work Where Teams Already Are

Side Panel —
Work Where Teams Already Are



Matching Product to AI Readiness

Once I knew AI should only surface suggestions at natural breaks, the next question was: where do they land? PMs are already in five tools. Adding another screen wasn't the answer — the summary needed to come to them. That's what the side panel does.

How it solves the problem

Instead of PMs manually updating 3-5 tools after each meeting, AI proposes cross-tool updates in one preview screen. User reviews and approves in 30 seconds, without leaving the tool they're already in.

How it solves the problem

Instead of PMs manually updating 3-5 tools after each meeting, AI proposes crosstool updates in one preview screen. User reviews and approves in 30 seconds.

How it solves the problem

Instead of PMs manually updating 3-5 tools after each meeting, AI proposes cross-tool updates in one preview screen. User reviews and approves in 30 seconds, without leaving the tool they're already in.

Mobile —
Approve Anywhere

Mobile —
Approve Anywhere

Mobile —
Approve Anywhere

PMs don't sit at their desks between meetings — they move. By the time they're back at a laptop, five suggestions have piled up and the context is gone. The mobile flow lets them approve in 15 seconds during a break, without losing the thread.

Unexpected value

During user interviews, a PM said: "Our auditors ask 'Why did you change the timeline?' I usually get confused before I answer."

It turned out the activity log wasn't just for users, it was for organizational memory. A feature I designed for speed became infrastructure for accountability.

Unexpected value

During user interviews, a PM said: "Our auditors ask 'Why did you change the timeline?' I usually get confused before I answer."

It turned out the activity log wasn't just for users, it was for organizational memory. A feature I designed for speed became infrastructure for accountability.

What Did the AI Do?

What Did the AI Do?

What Did the AI Do?

Side panel and mobile handle real-time decisions. What they don't cover is everything AI did between sessions — every update, every change. The transparency page gives PMs a complete audit trail and undo controls for each action.

Impact
& Outcomes

Impact & Outcomes

Impact & Outcomes

  • Consolidated 3-5 tool updates into a single approval flow, targeting the 4 hours PMs lose weekly to coordination overhead

  • Designed mobile as a parallel channel so decisions don't wait for a desk

  • Launched MVP validating the core hypothesis: the real PM tool isn't a new app, it's the conversations teams are already having

  • Consolidated 3-5 tool updates into a single approval flow, targeting the 4 hours PMs lose weekly to coordination overhead

  • Designed mobile as a parallel channel so decisions don't wait for a desk

  • Launched MVP validating the core hypothesis: the real PM tool isn't a new app, it's the conversations teams are already having

What I Owned
& Learned

What I Owned & Learned

What I Owned & Learned

Product vision evolved through design

I started with a simple transcription dashboard. Through iterations, I realized the real need was a side panel that syncs across tools, but users needed a "home base." This led me to design the desktop activity page as the source of truth.


Mobile solved an unexpected use case

During user interviews, I discovered PMs miss context during back-to-back meetings. I added mobile quick-view so they can catch up in 15 seconds without interrupting workflow, turning a gap in attention into a design opportunity.


AI products need grounding mechanisms

I learned that ephemeral AI suggestions (side panels, notifications) create anxiety without a persistent place to review decisions. The activity page became that anchor, proving that AI tools need both real time assistance and retrospective visibility.


Product vision evolved through design

I started with a simple transcription dashboard. Through iterations, I realized the real need was a side panel that syncs across tools, but users needed a "home base." This led me to design the desktop activity page as the source of truth.


Mobile solved an unexpected use case

During user interviews, I discovered PMs miss context during back-to-back meetings. I added mobile quick-view so they can catch up in 15 seconds without interrupting workflow, turning a gap in attention into a design opportunity.


AI products need grounding mechanisms

I learned that ephemeral AI suggestions (side panels, notifications) create anxiety without a persistent place to review decisions. The activity page became that anchor, proving that AI tools need both real time assistance and retrospective visibility.


Hadarmzr@gmail.com

Hadarmzr@gmail.com

Hadarmzr@gmail.com

☕︎ Made with coffee

☕︎ Made with coffee