New: Chai, my AI assistant, can brief you on my work in 60 seconds.

← All work

Case study · PepsiCo · AI & Data

Dialog with Data.

An AI chatbot that lets PepsiCo's commercial teams ask for sales data in plain language, instead of digging through dashboards or waiting on analysts.

Tablet mockup of the Dialog with Data 'Chat With AI' web app: icon sidebar, a drafted email reply in the chat thread, feedback icons, suggested follow-up topics and a message input

Some data in the screens is blurred or replaced to respect PepsiCo confidentiality.

At a glance

PepsiCo · via Tech Mahindra · 2023–25

The challenge

PepsiCo's commercial teams kept their sales data in Power BI dashboards and static reports. A basic question meant hours of navigation or a request to an analyst, and non-technical staff avoided the tools.

What I led

  • Design lead for DWD, owning UX/UI end to end
  • Planned and ran stakeholder interviews and workflow observation
  • Defined the intent categories, IA and conversation model
  • Designed and tested the Figma prototypes

Worked with product managers, data analysts and developers.

The outcome

70%+less time to retrieve data
60%+of routine queries handled by the bot
5user groups served from one interface
Stakeholder interviewsWorkflow observationUsage analysisBenchmarkingPersonas & journeysIA & conversation designFigma prototypingUsability testing
Dialog with Data
My role
Senior UX/UI Designer, design lead for DWD
Client
PepsiCo, via Tech Mahindra
Team
Product managers, data analysts, developers
Platform
Internal web app
Scope
Research, IA, conversation design, UI, prototyping, testing

Impact

Routine sales questions now take one chat message instead of a dashboard search or a request to an analyst.

70%+

less time to retrieve data

60%+

of routine queries handled by the bot

5

user groups served from one interface

Analyst workload dropped as the bot took over routine queries, and non-technical users who had avoided data tools started using DWD.

Context

An internal tool for PepsiCo's commercial teams.

Dialog with Data (DWD) is an internal web app built for PepsiCo through Tech Mahindra. Sales executives, category managers, data analysts, marketing teams and Revenue Management all depend on the same commercial data: weekly sales, market share and growth comparisons. Until DWD, that data lived in Power BI dashboards and static reports.

DWD replaces that with a conversation. People ask for reports, summaries and insights without needing analytics skills, for example tracking dollar and volume trends for Frito-Lay’s Salty Snacks week by week.

The problem

PepsiCo's commercial teams had the data, but getting to it took hours.

Weekly sales numbers, market share shifts and growth comparisons sat inside Power BI dashboards and static reports. Answering a basic question meant hours of navigation or a request to an analyst.

01

Slow access

Business users spent hours navigating complex Power BI dashboards to find the right numbers.

02

No real-time view

Reports were built by hand or on scheduled data pulls, which delayed decisions.

03

Analyst bottleneck

Many people needed an analyst to read the metrics, which slowed day-to-day work.

04

Low adoption

Non-technical stakeholders found the tools too complex, so usage was uneven across departments.

Business goal

Speed up decisions, reduce dependence on analysts, raise adoption of data tools and give every team consistent access to current commercial data.

My role

I led the end-to-end UX/UI design, from the first stakeholder interview to the final prototype.

I was the design lead for DWD on the Tech Mahindra team working for PepsiCo. I owned the research plan, the conversation structure and the UI, ran the interviews and usability tests, and worked with product managers, data analysts and developers so the interface matched how the data actually worked.

Research lead

Planned and ran stakeholder interviews with commercial leads, sales managers and analysts, plus workflow observation.

Product direction

Turned research into personas, journey maps and the problem framing the team built against.

Conversation & IA

Defined the intent categories, data structure and the guided Q&A model the chatbot follows.

UI & prototyping

Designed wireframes and high-fidelity Figma prototypes in line with PepsiCo’s design system.

Cross-team partner

Worked with data analysts, product managers and developers to map backend data logic to the interface.

Validation

Ran usability tests and added in-product feedback so the team could keep iterating after launch.

Process

The process, in six steps.

01

Discover

Interviews, observation, usage analysis

02

Define

Personas, journeys, goals

03

Structure

Intents, IA, conversation model

04

Design

Wireframes to hi-fi prototypes

05

Validate

Usability tests, in-product feedback

06

Measure

Retrieval time, query load, adoption

01  Discover

Before designing a chatbot, I needed to know which questions people ask most.

Discovery combined stakeholder interviews, workflow observation and analysis of how people used the existing data tools. The aim was to find the repeat questions, the workarounds, and where time was lost.

People told us the same three things: too many repetitive tasks, data in inconsistent formats, and no way to serve themselves.

Stakeholder interviews

Commercial leads, sales managers and data analysts.

Workflow observation

How people moved between dashboards, reports and analysts.

Usage analysis

Which queries came up most often, and how often.

Benchmarking

Enterprise chatbot experiences, to set success metrics.

Key findings

60%+

Most questions were repeats.

Over 60% of queries were repetitive: weekly sales, market share shifts and growth comparisons.

Speed

Speed mattered more than depth.

People rarely needed a full dashboard. They wanted a quick answer in context that they could trust.

Share

Answers had to be easy to share.

Data was pulled for presentations, updates and meetings, so people needed short summaries they could reuse.

Ease

BI tools were hard to learn.

Sales and marketing staff found the learning curve steep and wanted to ask the way they use search or a voice assistant.

Benchmark

How DWD compared with Microsoft Power BI Chat.

I benchmarked enterprise chatbot experiences to set success metrics. Power BI Chat was the closest alternative the teams already had access to.

CapabilityDWDPower BI Chat
Natural language queriesContextual NLPBasic NLP
Auto-generated summariesYes, shareableManual export
Live sales dataIntegrated with IRI and internal dashboardsManual or scheduled refresh
Easy for non-technical usersConversational UINeeds dashboard navigation
Analyst query load60%+ of queries automatedLimited impact
Fit to PepsiCo KPIsBuilt around PepsiCo KPIsPossible through the Power BI model
Scale across teamsDesigned for enterprise-wide useScalable with setup

02  Define · Users

Three primary personas, and two more groups who would use DWD on the side.

Sales Executive

Needs

Needs quick access to KPIs: weekly numbers, distribution changes, promotion results.

“Just tell me how last week went.”

Category Manager

Needs

Needs detailed, comparative views of brand and category performance, share trends and summaries.

“How did Salty Snacks do against a year ago?”

Data Analyst

Needs

Wants to validate findings and dig into trends, and spend less time on repeat data pulls.

“I want to stop answering the same question.”

Also served: Marketing teams (campaign impact, volume trends and consumer behaviour data) and Revenue Management (pricing, volume and profitability, modelling performance across markets and checking trade spend efficiency). Category managers included trade managers, who needed competitive insights.

Persona quotes are illustrative summaries of interview themes, not verbatim.

Journey

The weekly sales check-in, before and after.

I mapped journeys for the most common use cases, including the “Weekly Sales Check-in” and the “Market Share Summary”. The check-in shows the gap most clearly.

Before
  1. Open the Power BI dashboard
  2. Hunt for the right report
  3. Set filters by hand
  4. Export to slides
  5. Ask an analyst when stuck
After, with DWD
  1. Ask in plain language
  2. DWD clarifies the intent
  3. Filters applied in context
  4. Summary returned with the SQL
  5. Share or ask a follow-up

03  Structure

Organising the bot around the questions people ask.

I mapped common intents, such as checking weekly sales or comparing brand performance, and grouped them into five categories. Each category breaks down into the same four data building blocks, so the bot can handle new questions without new screens.

Intent categories

Sales Overview
Category Insights
Promotions
Market Share
Performance Summary

Data building blocks

MetricsDollars, volume, share
Timeframes1 week, 4 weeks, YTD
Comparisonsvs. previous period, vs. year ago
FiltersChannel, region, brand, product line

Conversation model

01

Clarify the intent

“Would you like sales for a specific product or total performance?”

02

Apply filters in context

Ask about “chips” and DWD applies the right product family filter.

03

Return a usable answer

“Sales for Lay’s increased by 8% vs. last week,” with options to expand, export or summarise.

Conversation flow

Every conversation was built to end in an action, not just a number.

Start

Natural-language input or dropdown selections

Mid-flow

Contextual prompts, such as “Would you like a summary report?”

End

An answer with buttons to expand, export or summarise

Under the hood this is a hybrid: a natural-language layer for flexibility, decision-tree logic for speed on known paths, and prompts that invite the next question. The structure was set up to take on new data sources, KPIs and business units later.

Key decisions

Five design decisions that shaped the product.

01

Show the SQL behind every answer

People were used to checking numbers with an analyst. Showing the query next to each response lets them see where an answer came from.

02

Guided prompts alongside free text

A blank chat box is hard for new users. Starter cards (Commercial, Consumer) and suggested follow-ups give a way in, while power users can type anything.

03

Summaries first, tables on demand

Research showed people wanted short answers they could share. DWD leads with a summary and lets users expand the full report card when they need detail.

04

Recent queries and quick reports

Over 60% of questions were repeats, so Recent Queries and Quick Reports let people rerun a common question instead of typing it again.

05

Comparisons as a built-in view

Top products, sales volume, channel performance and display metrics are the comparisons people asked for most, so each has its own view.

04  Design

From low-fidelity flows to a high-fidelity prototype in PepsiCo’s design system.

Wireframes focused on three parts: the chat layout with message input and AI prompts, the embedded report preview card (for example Frito-Lay Sales, 1 Week) and the action panel (Generate Summary Report, Expand). High-fidelity prototypes were built in Figma.

Getting started screen with two starter cards, Commercial and Consumer, each with a View Summary button, and a message input Commercial chat showing an embedded Frito-Lay report card with a Salty Snacks 1 Week table, an Expand button and follow-up options Generate Summary Report and Show more suggestions Chat With AI screen with feedback icons under a response and a list of suggested follow-up topics
01 — Getting started

A way in for new users.

Starter cards give new users a way in; anyone can still type a question.

Getting started screen with two starter cards, Commercial and Consumer, each with a View Summary button, and a message input
02 — Report card

Answers inside the conversation.

A weekly Salty Snacks report returned in the conversation. DWD leads with a summary and lets users expand the full report card when they need detail.

Commercial chat showing an embedded Frito-Lay report card with a Salty Snacks 1 Week table, an Expand button and follow-up options Generate Summary Report and Show more suggestions
03 — Follow-ups & feedback

The next question, one tap away.

Suggested follow-ups give a way in, while power users can type anything. Each response can be copied, saved, rated or regenerated, and the same chat also helps with everyday tasks such as drafting an email.

Chat With AI screen with feedback icons under a response and a list of suggested follow-up topics

Engagement

Seven features for repeat use.

Research and behaviour data shaped seven features meant to bring people back.

01

Smart prompts

Suggests next steps from past use. Someone who often checks Salty Snacks sales is offered “Would you like to see last week’s update again?”

02

Weekly digest

A conversational summary of key changes at the start of each week. It became a popular way back into the product.

03

Trending queries

Shows what colleagues are asking, which helps with the blank-page problem.

04

Guided discovery

A light tutorial for new users. After a report, DWD might ask “Did you know you can also compare this to the previous 4 weeks?”

05

Personal memory

Remembers preferred categories, timeframes and metrics.

06

In-context feedback

Thumbs up or down and “Was this helpful?” on every report.

07

Built-in help

A context-aware FAQ and quick tips inside the chat.

05  Validate

Usability testing, then feedback from inside the product.

I ran usability tests on the prototype to check the conversation flow and the report cards. After launch, the thumbs up and down buttons on each report gave the team a steady signal on which answers were useful, so we could iterate quickly.

06  Outcomes

What changed for PepsiCo’s commercial teams.

70%+

reduction in data retrieval time

60%+

of routine queries automated

Analyst workload dropped as the bot handled most routine queries, freeing analysts for deeper analysis and modelling.

Adoption grew among non-technical users who had avoided data tools before.

Teams made decisions faster and from the same, current commercial data.

Learnings

What I took away from designing an AI product for business users.

Show the source

With AI answers, people want to see where a number came from before they act on it. That is why every DWD answer shows its SQL.

Repetition is a signal

Repeat questions, over 60% of all queries, set the priorities: recent queries, quick reports and the weekly digest.

Design the structure

Intent categories and shared data blocks meant new questions and KPIs fit without new screens.

Next: the architecture was set up to add new data sources and more advanced analytics without changing the conversation model.

Let's build something people love to use.

Open to UX/UI leadership roles, design advice and friendly chats.

Email meLinkedIn ↗Resume (PDF)