Sagar Kumar

Jio

AI Chat Pattern

A chat framework for the Jio ecosystem: a standalone destination, an embedded module inside existing Jio apps, and a broker into partner services.

Industry

Consumer App · B2C · Conversational UI

Platform

Android & iOS

Duration

12 weeks

Context

Where it started

Jio runs a lot of apps. Telecom, retail, groceries, video, conferencing. Each one had built its own chat support, and none of them talked to each other. JioMart's looked nothing like JioMeet's. JioCinema's behaved differently again. Three of them ran on the same vendor and still had nothing in common.

The effect on users was simple. Figuring out how to get help in one Jio app taught you nothing about the next one.

Across 125,560 users over six months:

reached support

opened a query

got to an agent or bot

dropped at the product section

The drop-outs are the part I kept coming back to. Those people didn't stop needing help. They picked up the phone. Phone support is the most expensive channel Jio runs, so a discovery problem in the UI was showing up as a line item in operations.

JioMart

JioMeet

JioCinema

Ajio Retail

Jio products existing chat experience

Context

The Brief

This is the part that reframed the whole project for me. The brief wasn't "fix the chatbot." From the kickoff:

The Ask

"Enabling all the products with chat framework so that they don't have to do it themselves. Chat as a PaaS product. Configurable for each Jio product. A customer who goes into the ecosystem will find it familiar. Costs should also reduce."

Uniformity of experience - Low learnability cost - Reusability

How it was meant to evolve

Initially it will be generic, with time it will be flexible, pertaining to the need of business

Generic first, configurable later.

And the business case

Cost should also reduce

Why it matters to users

A customer who goes into the ecosystem will find it familiar

Who else it was for

An ecosystem which internal teams can look up - the best practices, best configuration for their business.

Other jio designers and enigneers were users too.

So the users I was designing for weren't only the 400M people using Jio apps. They were also the designers and engineers in other business units who would pick this up and ship it without me in the room. That changed what "good" meant.

One boundary I should be clear about: the conversational AI ran on Haptik, Gemini & custom GPT's. I didn't build the model. I built the tone of voice, the interface, and the pattern language on top of it.

Action

Process

Stakeholder research across the portfolio. An audit of every chat implementation already live. Interviews. A survey with 94 respondents. Five personas. A six-platform benchmark. And a pattern study comparing menu-based, keyword-based, contextual, and voice bots on cost, failure mode, and fit.

Understanding

Secondary research with business stakeholders across Jio's product portfolio. An audit of every existing chat implementation. Primary interviews. A 94-person survey. Five personas.

Framing

A six-platform competitive benchmark, and a pattern study comparing menu-based, keyword-based, contextual, and voice bots on cost, failure mode, and fit.

Building

Conversational architecture first, then component families, then high-fidelity states. Prototyping and usability testing throughout.

Three findings that changed the design

People came for status, not conversation: Order tracking, delivery updates, troubleshooting. Transactional things, badly served by a blank text box.

People came for status, not conversation: Order tracking, delivery updates, troubleshooting. Transactional things, badly served by a blank text box.

All five personas wanted visible choices: They differed on context and converged on behaviour. Text-heavy flows were the common complaint.

All five personas wanted visible choices: They differed on context and converged on behaviour. Text-heavy flows were the common complaint.

The market said the same thing: 59% of Indians still prefer speaking to a human for customer service, and average time-to-resolution in India runs 2h17m against 39min globally. Designing the exit mattered as much as designing the answer.

The market said the same thing: 59% of Indians still prefer speaking to a human for customer service, and average time-to-resolution in India runs 2h17m against 39min globally. Designing the exit mattered as much as designing the answer.

Action

Research: what shaped the direction

The main portfolio page keeps only the findings that changed a design decision. The detailed interview, survey, persona, and benchmark artifacts are linked rather than stacked into the reading path.

COMMON REASONS TO INTERACT

Suggestions and recommendations; registering complaints; bookings; payments and refunds; technical assistance; feedback; tracking orders and queries.

WHAT USERS VALUED

Order tracking and status updates, product recommendations, troubleshooting, and technical support.

ECOSYSTEM GAP

Most chatbot usage was concentrated in MyJio and Ajio; many users had not used Jio chatbots across businesses.

RESEARCH COVERAGE

Secondary research with business stakeholders; primary interviews; a 94-person survey; five personas; and synthesis into a reusable chat system

USER-STORY CONSTRAINTS

The supporting story inventory made the interaction requirements concrete: greet users on entry; support registration/login in chat; let users ask a new topic while shopping and return to the original task; and open chat from in-app notifications such as tracking, delivery, or agent assignment.

Action

Three months of research before I drew anything high- fidelity

Four steps. I've kept the ones that changed a decision and linked the rest.

01 Audit what already existed

Six Jio products, captured live. Six unrelated patterns, three of them on the same vendor. What it changed: it reframed the problem. Not one broken bot, but six teams solving the same thing separately.

02 Talk to users

Interviews, a 94-person survey, five personas. People came for status rather than conversation — tracking, delivery updates, troubleshooting. All five personas diverged on context and converged on wanting visible choices. What it changed: quick actions for the high-frequency intents instead of leading with a blank text box.

03 Study the patterns before choosing one

Menu, keyword, contextual and voice bots, compared on build cost, failure mode and fit. Guided patterns are reliable and rigid. Open patterns are flexible and fail badly. At this scale neither works alone. What it changed: it's the basis for the hybrid decision. Without it, hybrid is a preference.

04 Benchmark and market

Six platforms compared against Jio, plus India-specific support behaviour. 59% of Indians still prefer speaking to a person for customer service, and average resolution time here runs 2h17m against 39min globally. What it changed: it turned the escape hatch from a nicety into the second design decision.

Action

Four decisions, and what each one cost

Guided actions over pure conversation

Visible quick actions alongside open text and voice. Chips handle the high-frequency intents, open input catches the rest

What it cost: it reads less like an assistant and more like a guided tool. I took that trade for completion speed, and because chips work for the people who won't type at all.

The exit stays visible

Handoff to a person available at the points where people give up, rather than buried in a menu.

What it cost: more human handoffs, and more support spend, deliberately. Given 59% prefer a person and the drop-outs were already calling, we were paying for those users anyway. Just later, and by then they were annoyed

One interaction language across every host app

Same chat items, same states, whether it sits in the assistant or inside JioMart.

What it cost: no host app got a treatment tuned to its own conventions.

The component library came before the visual polish

A shared repository inside the Jio Design System instead of a bespoke build per business.

What it cost: no host app got a treatment tuned to its own conventions.

Action

Principles That Guided Every Decision

The core strategic shift: move from a decision-tree navigator to an intent-aware assistant. The bot should feel like a knowledgeable Jio representative — not a branching FAQ.

Principles set for this project:

Progressive disclosure — show only what’s needed, when it’s needed
Escape hatch always visible — “Talk to agent” accessible at every step, not buried
Human tone first — warm, plain language; no jargon or robotic phrasing
Resolution over engagement — success = task completed, not time in session
Accessibility by default — large tap targets, high contrast, voice input support
Feedback as infrastructure — every AI response gets a thumbs up/down

Principles set for this project:

Progressive disclosure — show only what’s needed, when it’s needed
Escape hatch always visible — “Talk to agent” accessible at every step, not buried
Human tone first — warm, plain language; no jargon or robotic phrasing
Resolution over engagement — success = task completed, not time in session
Accessibility by default — large tap targets, high contrast, voice input support
Feedback as infrastructure — every AI response gets a thumbs up/down

Progressive disclosure — show only what’s needed, when it’s needed
Escape hatch always visible — “Talk to agent” accessible at every step, not buried
Human tone first — warm, plain language; no jargon or robotic phrasing
Resolution over engagement — success = task completed, not time in session
Accessibility by default — large tap targets, high contrast, voice input support
Feedback as infrastructure — every AI response gets a thumbs up/down

Action

Key Decisions & Trade-offs

Decision 1

Hybrid conversation instead of AI-only

We considered a fully conversational experience but found that users preferred quick actions for common support tasks. Rather than forcing natural language for every interaction, we combined quick replies with open text input.

Trade-off

Slightly less conversational.

Much faster task completion.

Decision 1

Hybrid conversation instead of AI-only

Trade-off

Slightly less conversational.

Much faster task completion.

Decision 2

Persistent "Talk to Agent"

Research showed users often wanted reassurance that human help was always available. Instead of hiding escalation inside menus, we exposed it throughout the journey.

Trade-off

Slight increase in agent visibility.

Large increase in user confidence.

Decision 2

Persistent "Talk to Agent"

Research showed users often wanted reassurance that human help was always available. Instead of hiding escalation inside menus, we exposed it throughout the journey.

Trade-off

Slight increase in agent visibility.

Large increase in user confidence.

Decision 3

Memory over personalization

Rather than introducing highly personalized AI responses, we prioritized conversation continuity. Users cared more about not repeating themselves than receiving personalized recommendations.

Decision 3

Memory over personalization

Rather than introducing highly personalized AI responses, we prioritized conversation continuity. Users cared more about not repeating themselves than receiving personalized recommendations.

Decision 4

Design system before visual polish

Instead of creating unique interfaces for each Jio app, we invested in reusable conversation components that could scale consistently across products.

Decision 4

Design system before visual polish

Instead of creating unique interfaces for each Jio app, we invested in reusable conversation components that could scale consistently across products.

Results

Bringing the Experience to Life

Designing the input system

Every conversation starts at the input bar. We designed it to handle text, voice, files, images, and complex multi part prompts with equal ease — meeting users at their comfort level instead of forcing them to type.

Text & Voice Responses

Text and voice share the same response system — thinking states, reasoning transparency, and per-message actions for feedback, copy, share, and read-aloud.

Voice adds visible listening and speaking states with a persistent Stop control, so whether users type or talk, they always know what the assistant is doing.

Text and voice share the same response system — thinking states, reasoning transparency, and per-message actions for feedback, copy, share, and read-aloud.

Voice adds visible listening and speaking states with a persistent Stop control, so whether users type or talk, they always know what the assistant is doing.

Loading & Failure States

Every wait and every failure is accounted for — thinking indicators, reasoning-time transparency, and streaming states set expectations while the AI works; failed responses, timeouts, and connectivity drops recover with clear messaging and one-tap retry. The assistant never leaves users staring at silence.

Results

One pattern, three deployments

The framework had to survive three quite different situations

On its own — a dedicated destination with its own home, assistants and tools.

Inside another Jio app — a module wearing the host's brand, inheriting its account context, escalating into that business's own support team.

Brokering outside Jio — the same conversation reaching into food delivery, flights, buses and cabs, resolving into partner systems.

Results

What shipped, and what I can stand behind

What shipped was a chat language, not a screen. Four component families across light and dark. A shared pattern library in the Jio Design System. Documented behaviour for loading, failure, voice, attachments and handoff, so another team could implement it without asking me questions.

On the numbers

CSAT (from 2.6)

4.2/5

Call volume drop

55–60%

Avg resolution time

<90s

Users reached

400M+

Where it goes next

Configurability, not just consistency. v1 shipped generic on purpose. Per-business configuration is the step that makes it a platform rather than a pattern library.
Voice, properly. Demand was highest among users over 45, exactly the group that struggled most with text. Timeline killed it in v1 and I think that was the wrong call.
Regional languages past Hindi. Script and edge-case handling surfaced late. For a 400M-user base that can't be a later phase.
A measurement layer the host teams share. Every deployment reports into its own funnel. Nobody can see whether the pattern is improving across the ecosystem or just locally.
A contract partners implement against. Each brokered service is its own build today. • • • • •

Configurability, not just consistency. v1 shipped generic on purpose. Per-business configuration is the step that makes it a platform rather than a pattern library.
Voice, properly. Demand was highest among users over 45, exactly the group that struggled most with text. Timeline killed it in v1 and I think that was the wrong call.
Regional languages past Hindi. Script and edge-case handling surfaced late. For a 400M-user base that can't be a later phase.
A measurement layer the host teams share. Every deployment reports into its own funnel. Nobody can see whether the pattern is improving across the ecosystem or just locally.
A contract partners implement against. Each brokered service is its own build today. • • • • •

Results

Looking back

What worked

Doing the pattern study before touching UI, so the hybrid decision had a basis instead of a preference behind it. Pulling support leads in early — they knew operational constraints no amount of user research would have surfaced. And treating feedback as infrastructure rather than a UI element, which turned out to be the thing with the longest life.

What I'd do differently

Push harder on voice in v1. Run the regional language beta longer. Build the shared analytics view before launch instead of chasing it after. And get the AI team in the room from week one — some patterns had to change once model limitations showed up, and that was avoidable.

The thing I actually took from it

Designing a framework is a different job from designing a product. A product works if users finish tasks. A framework works if other teams choose it over building their own. That flipped what I optimised for. Fewer clever decisions. More decisions I could explain to someone I'd never meet.