Mamo, Web dashboard, 2026

AI Insights

Mamo’s first AI feature is a chat inside the merchant dashboard. Merchants ask questions about their money, like revenue so far this month, why payments are declining, or how this month compares, and get answers back.

It started in a design sprint with Joharah, Head of Design, and the product squad. Joharah led the sprint and owns the main UI that came out of it. My side was the system underneath: the Blaze components, every flow working on mobile, every interaction state, and the annotations engineers build from.

11
components, built as one Blaze section
Apr to Jul 2026
from the team design sprint to the rollout to merchants
+61%
more merchants using it, July to August 2026 (33 to 53)
Team
Mujtaba M (me)Product Designer: design system, states and handoffJoharah AlomairHead of DesignAsma AlyamaniChief Product OfficerFrank StodulskiLead Front-End EngineerPhu TruongBack-End Engineer

The challenge

An AI chat interface is a bigger scope than a standard screen. It needed voice input, several response formats, responsive behavior in every flow, and components that would hold up for the next AI feature. Most of it had to be built from scratch.

The sprint and the design review gave us the UI. Someone still had to build the system layer underneath it.

How I approached it

01

The design sprint

The work started with a design sprint in April 2026, led by Joharah. Everyone put ideas on the table to solve the problem, and I put mine in with the rest. We warmed up with predictions: each of us wrote down one thing we wished our analytics could tell us. Then an audio-only exercise, giving one specific, actionable insight to someone who could only hear it. After that we looked at analytics in products outside fintech, then sketched Crazy Eights under constraints: analytics without a dashboard, AI only, and a user who never opens analytics. I was in the AI only group.

The ideas that got votes fell into themes: talk to the dashboard, proactive AI, and summaries. Then everyone wrote a concept pitch. Mine was Mavis, a financial co-pilot you can type to or talk to, with quick actions so you don’t have to write a question. I vibe-coded a quick mock to show the concept at a high level.

My eight Crazy Eights sketches for the AI only constraint, one idea a minute: daily texts from a bot, a voice call about the metrics, investor docs, a Vision Pro version. The colored dots are votes.
My eight Crazy Eights sketches for the AI only constraint, one idea a minute: daily texts from a bot, a voice call about the metrics, investor docs, a Vision Pro version. The colored dots are votes.
The Mavis mock, vibe-coded during the sprint to show the concept at a high level: a chat you can type or speak to, quick actions you can tap, and an input that says ask anything about your business.
The Mavis mock, vibe-coded during the sprint to show the concept at a high level: a chat you can type or speak to, quick actions you can tap, and an input that says ask anything about your business.
02

Starting with the system

After the first internal review, when the UI and UX were close, I started leaving comments on the file as to-dos for myself: what the design system layer would need, the component architecture, the build plan, and what would let it grow with Mamo’s AI roadmap.

A feature like this only scales if it starts with a proper system, before more designs and flows get built on top of it. So I handled the system first and made sure everything underneath it was in place.

03

Component suite

I built the AI components as their own section in Blaze: 11 components from scratch. They were the prompt input (with voice recording and a live audio waveform), chat sidebar and its menu item, top bar, bottom bar, user bubble, response (text, table, chart and system messages), action bar, thinking state, assistant label, and loader.

The AI suite as one Blaze section: inputs, chat chrome, response types, and states, built so the next AI feature can reuse them.
The AI suite as one Blaze section: inputs, chat chrome, response types, and states, built so the next AI feature can reuse them.
04

The final designs

The components came together as Ask AI, a chat inside Insights. It opens from the Insights page, where a prompt input and three quick starts sit above the charts. The chat then takes over the screen: pinned and recent chats on the left, and answers as text, charts or tables, each with copy, like, dislike and retry underneath. These are the specs engineering built from, and every screen runs on the Blaze components above.

The phone versions had to carry the same answers, charts and tables included, so they scroll and fold instead of shrinking or dropping anything.

The entry point: Ask AI on the Insights page, with a prompt input and three quick starts above the charts.
The entry point: Ask AI on the Insights page, with a prompt input and three quick starts above the charts.
A chart response in the chat. Pinned and recent chats sit on the left, the answer comes with a CSV download, and the actions sit under it.
A chart response in the chat. Pinned and recent chats sit on the left, the answer comes with a CSV download, and the actions sit under it.
A table response: pagination and a CSV download, with the same actions under the answer.
A table response: pagination and a CSV download, with the same actions under the answer.
The same chat on a phone: the chart answer, and a wide table you scroll sideways.
The same chat on a phone: the chart answer, and a wide table you scroll sideways.
The same chat on a phone: the chart answer, and a wide table you scroll sideways.
05

Interaction states

  • Speech-to-text: the full state set, covering listening, transcribing, two-stage feedback, and confirm and cancel controls.
  • Chat response actions: copy, retry, like, dislike and regenerate, each end to end with snackbar feedback and navigation between retry versions.
  • Sidebar actions: pin, rename and delete, with modals, snackbar and redirect behavior.
  • Chat loading: a thinking state with copy tied to the type of query.
  • Table and chart loading: each with its own thinking state.

Responsive was the hardest of these. On a phone you still have to show what someone would see on a large screen, without compromising the charts and tables, and finance data is heavy by nature. It has to stay understandable. The base designs covered mobile, but not every flow, so I made sure all of them worked at mobile sizes, including the table treatment, horizontal scroll, stacked footer and download placement.

The speech-to-text state set: listening at first input, listening through a full take, transcribing, then the editable result.
The speech-to-text state set: listening at first input, listening through a full take, transcribing, then the editable result.
06

Handoff

Annotations are how a design gets across to the people building it. A developer can’t ask a screen what happens when the input grows to two lines, or while a table is loading, or on a small phone. The designer’s job in the handoff is to answer those questions before they’re asked: how each state behaves, what the thinking copy says, what changes on mobile, and what happens at the edges.

I annotated every scenario on the handoff screens, so engineering could build from them without guessing.

One handoff screen, annotated: thinking-state copy keyed to query type, input single vs multi-line, and the edge cases a developer would otherwise guess at.
One handoff screen, annotated: thinking-state copy keyed to query type, input single vs multi-line, and the edge cases a developer would otherwise guess at.
07

What emerged

The placeholder came from my side. Prompts cycle through the empty input to spark questions, things a merchant could ask about their business. They’re a way in for anyone who doesn’t pick one of the preset chips. I moved them with a fade and a small vertical drift, from Emil Kowalski’s animation principles, instead of a typing effect. Typing takes longer. A fade is quicker, eases in, and keeps it clean.

The placeholder animation: prompt suggestions cycle through the empty input so merchants always have a way in.
08

Motion

The motion work ran alongside the components. I used Emil Kowalski’s animation principles as a working reference, and wrote each animation up as a spec so engineering could build it as designed.

The placeholder cycle is one of them. The spec says when it plays (when the input is empty and unfocused, and it stops the moment you focus or type), what moves (a fade and an 8px vertical drift between prompts), how long (220ms ease-out, each prompt held for 2.8 seconds), and what not to touch: opacity and vertical movement only, no layout properties, so it stays cheap to run.

The spec for the placeholder cycle: trigger, motion, timing and easing, and the properties that are allowed to animate.
The spec for the placeholder cycle: trigger, motion, timing and easing, and the properties that are allowed to animate.

What came of it

A spec ready to build for Mamo’s first AI feature.

  • Components: 11 AI components in Blaze that future AI features can reuse.
  • States: every flow works on mobile, and every interaction state is designed and annotated, so engineering could build from the spec without guessing.
  • Pilot: the beta launched to learn what merchants actually ask for, with a feedback loop that tracks each request the feature couldn’t handle.
  • Rollout: it reached every merchant in July 2026.
  • Adoption: unique merchants using it went from 33 in July to 53 in August, up 61%. The base is small, since the feature was still a pilot, but it was growing.

Reflection

What I’m proudest of

  • Helping build Mamo’s first AI feature on the dashboard.
  • The edge cases: there were a lot of them, and each one has an answer.
  • The system layer: it’s properly defined, so the next AI feature starts with building blocks that already exist. Plug in what’s there, add what’s needed.

What I would do differently

  • Build for more than one surface from day one: the sprint’s strongest theme was AI that meets you where you work, in WhatsApp, in Slack, in a daily digest. We built the chat inside the dashboard first. I’d define the answer types (text, table, chart) as content any surface can render, so the same answer could show up anywhere.
  • Design for actions, not just answers: the sprint pointed at an assistant that does things, like drafting the message for an unpaid invoice or flagging a customer about to churn. Our action bar covers copy, retry and rating. I’d plan the confirm, undo and approval patterns early, because trust is the hard part once an assistant acts.
  • Remember the merchant’s goals: the prediction I wrote in the sprint was about setting a goal for the business and what it would take to get there. The chat answers one question at a time. I’d plan for it to hold a goal, track it and bring it up, so it works with the merchant over time, not only per question.