The idea
Ask, instead of digging
One number meant picking a dashboard, a store, and dates. The assistant takes the question in plain words and answers with text, a chart, or a table.
Quick look
I designed an AI assistant that answers leaders' questions in plain words.
A chat assistant for store data, and two changes after leaders stopped using it.
Who it is for
Three kinds of leader, grouped by what they did, not by title.
How the project went
- BuildOne box that answers with text, a chart, or a table.
- LaunchIt shipped inside the dashboards, with three ways in.
- Find out whyLeaders tried it once and drifted back, so we asked why.
- SimplifyTwo changes: ready questions, and summaries that arrive on their own.
Change one
A place to start
Prompt library in the header opens ready questions beside the chat, sorted by topic. Hover a card to copy it, or pin it to your home.
Built for our models
A rough question, made clear
To keep costs down, we did not use the largest models. Smaller models need a clear question, so Improve prompt rewrites a rough one before it is sent.
Change two
Answers that arrive on their own
A morning summary waits on the home screen. On any dashboard, Summarize page reads the page and says what changed.
About the project
Getting one number took three choices in a dashboard
Where are leaders spending effort just to get one number? Jarvis had detailed dashboards for stores, cameras, and queues. Getting one number meant picking a dashboard, a store, and dates.
Which dashboard, which store, which dates: every question started with three choices.
Who I designed for
Three kinds of leader I designed for
Not by title. Drawn from how leaders used the dashboards, and from what the team heard after launch.
The quick checker
"I just need the number. I do not have time to find the right dashboard."
- The pattern
- Wants one number between meetings, and has no time to pick a dashboard.
- What they needed
- One number, in plain words, without picking a dashboard.
The blank page
"I opened it and did not know what to type."
- The pattern
- Opens the box and does not know what to ask.
- What they needed
- Somewhere to start that is not an empty box.
The occasional visitor
"I only look when something is already wrong."
- The pattern
- Opens the dashboards only when something already looks wrong.
- What they needed
- A reason to look before something goes wrong.
Where it lives
The assistant shows up in three places inside the dashboards
Leaders already lived in the dashboards, so the assistant had to show up there.
The welcome card shown the first time, the floating button, and the header item next to search. All three are in the final design.
Version one
Version one: a single box where you ask in plain words
I designed the first screen as a greeting and a single box. Ask in plain words, and the assistant answers with text, a chart, or a table.
Ask in plain words, and the answer comes back in the right shape.
Show the work
Each step is visible, so leaders trust the answer
While the assistant works, the thinking panel lists each step, like finding stores in Haryana and Delhi. Seeing the steps is what lets a leader trust a number they cannot check themselves.
First step
All steps
Step one doneThe answer
The right shape
A chart for counts, a table for events
A count comes back as a chart with the total in the middle. A list of camera events comes back as a table with photos.
Feedback
One tap to say an answer missed
Thumbs up or down is one tap. A thumbs down asks what went wrong, so the team knows what to fix. The reasons came back to the team as a list of misses to work through.
After launch
It shipped. But adoption was low.
Leaders tried it once, then drifted back to the dashboards. We went looking for why.
Nobody knew what to ask.
An empty box asks you to invent a question. Most leaders did not.
This became the prompt library.
Nobody came back to ask.
The assistant only spoke when spoken to. If you did not open it, it gave you nothing.
This became the morning summary.
Change one
Change one: a prompt library of ready questions
I designed a set of ready questions that open from Prompt library in the header, beside the chat and sorted by topic. Leaders can copy one or pin it to their home.
Hover a card to copy the question into the box, or pin it to your home.
Improve prompt
A rough question becomes a clear one
To keep costs down, we did not use the largest AI models. Smaller models give good answers only when the question is clear. A loose question like mumbai west footfall n occupancy last 2 months got a loose answer. So Improve prompt rewrites it into a full question first, and Undo brings the original back.
Version one
With Improve prompt
Change two
Change two: a morning summary that arrives on its own
I designed short summaries that the assistant writes on its own. They show what changed, what needs a look, and questions to ask next.
Every morning
A morning summary on the home screen
Three short points from yesterday's data, like footfall, queues, and camera health. Each point has Ask about this, for a quick follow-up.
On any dashboard
Summarize the page you are looking at
A leader already looking at a dashboard should not have to leave it to understand it. So the summary lives on the dashboard itself, one tap away.
- Tap Summarize pageThe button sits next to the page title.
- It reads the pageIt uses the filters you already set, and shows each step.
- The summaryWhat changed, what needs a look, and what to ask next.
Your choice
Leaders decide when summaries arrive
A simple schedule: how often, and at what time. Then the topics to cover, the stores, and whether to get it by email too.
What happened
What each change was meant to fix, and what I did not measure
Prompt library
This is what it was designed to change. We did not measure it. What I watched for was whether a leader who had frozen at the empty box tapped a ready question instead.
Improve prompt
Not measured. I watched whether loose questions, like the Mumbai footfall one, came back as full sentences before reaching the model.
Morning summary
Not measured. I watched whether leaders who only opened Jarvis when something was wrong now saw it every morning, without opening anything.
Prototype
Try Jarvis GPT
I used Claude to turn my Figma screens into this working prototype. The screens and sample data come from my final design.
Ask a question, open the prompt library, or ask about the morning summary. Press Esc to close.
