Scaling Sales Inside WhatsApp - Without Losing Control

AI-Assisted SaaS

B2B SaaS

WhatsApp CRM

Chrome Extension

RAG-Based AI

Workflow Automation

Product Strategy

Kraya is a WhatsApp-native CRM that lets sales teams manage leads, follow-ups, and conversations directly inside WhatsApp Web – with AI assistance designed to scale outreach safely, without breaking trust or workflows.

TEAM
1 Product Manager (me)
3 UI/UX Designers
5 Developers
5 QAs 

ROLE
Product Strategy & Roadmapping
AI System Design (RAG-based, human-in-the-loop)
UX/UI Ownership (Extension + Dashboard)
QA Lead (Edge cases, safety, fallbacks)
GTM & Stakeholder Alignment

TIMELINE
11 Months
(strategy → MVP → launch → adoption → iteration)

TL;DR – Executive Summary

Kraya AI is a WhatsApp-native CRM built to help high-velocity sales teams manage leads, follow-ups, and responses directly inside WhatsApp Web.

I led Kraya from 0→1, identifying the core workflow breakdown, defining system architecture, designing safe AI usage, and shipping a Chrome extension–based product adopted by paying teams.

Early impact across pilot teams:

  • ~2.5× increase in leads processed per rep

  • ~40% faster median time-to-close

  • ~78% reduction in leads lost due to no response

This case study focuses on product judgment, system design, and responsible AI adoption in a customer-facing workflow — not just features shipped.

Kraya embeds CRM workflows directly inside WhatsApp Web.

Why Kraya Exists

Kraya was not built to “add AI to WhatsApp.”

It was built to fix a structural breakdown we repeatedly observed in sales-driven businesses.

WhatsApp had become the primary sales channel — but teams were operating without any real system around it.

Sales teams were:

  • Managing hundreds of conversations manually

  • Tracking follow-ups in spreadsheets or memory

  • Responding inconsistently based on individual judgment

  • Missing leads due to delayed responses

  • Lacking visibility into pipeline health or accountability

Traditional CRMs existed, but they lived outside WhatsApp.
Every context switch reduced speed, accuracy, and adoption.

Core Insight
If sales happens inside WhatsApp, the CRM must live inside WhatsApp – not beside it.

Kraya embeds CRM workflows directly inside WhatsApp Web.

The Real Problem (Beyond “WhatsApp Is Not a CRM”)

This was not a tooling problem.
It was a control and trust problem.

Business Reality

  • Every inbound message required manual effort

  • Lead qualification varied by sales rep

  • Slow responses caused drop-offs

  • No reliable way to prioritize high-intent leads

  • No system to scale without breaking trust

Why This Was Hard

  • WhatsApp conversations are unstructured and context-heavy

  • Errors are customer-visible

  • Fully autonomous AI carried reputational risk

  • Over-engineered CRMs failed adoption tests

Key PM Question
Can we reduce manual effort and improve conversion without sacrificing trust or control?

Early Hypotheses That Shaped the Product

Before designing solutions, I framed the problem into testable hypotheses:

  • If CRM workflows live inside WhatsApp, follow-ups will become consistent

  • If lead context is visible alongside conversations, reps won’t rely on memory

  • If AI assists with language-heavy work (not decisions), teams can scale safely

Every feature shipped was evaluated against these hypotheses.

Product Strategy: System First, AI Second

One of my earliest decisions was not to start with AI.

AI only works when:

  • There is structure

  • There is context

  • There are clear boundaries and fallbacks

Kraya was designed as a multi-layer system, with AI layered on top – not bolted on.

The Four Core Layers

  1. WhatsApp-Native Execution

  2. CRM & Analytics Backbone

  3. Automation & Follow-Up Engine

  4. AI & Knowledge Systems

Kraya layers
Kraya was intentionally built bottom-up - establishing execution, visibility, and automation before layering AI to ensure trust, control, and predictable outcomes.

Layer 1: WhatsApp-Native Execution (Chrome Extension)

This layer determined adoption.
Instead of forcing sales teams into a new tool, I embedded CRM functionality directly into WhatsApp Web.

What I Led

  • Persistent CRM sidebar inside WhatsApp
  • Lead details, stages, notes, calls, follow-ups
  • Zero tab-switching workflows

Key Product Decisions

  • CRM actions live inside the chat context
  • Notes auto-save to remove friction
  • Inline stage updates reduce cognitive load
  • UX parity with dashboard prevents relearning

Explicit Tradeoff (Deliberate)

Choosing a Chrome extension was intentional.

Pros

  • Zero learning curve
  • Immediate access to active conversations
  • Faster validation and iteration

Constraints

  • Requires WhatsApp Web to be open
  • Limited background automation
  • Browser-dependent performance

I accepted these constraints to validate workflow adoption first, before scaling infrastructure.

Layer 2: CRM & Analytics Backbone

The dashboard acted as the source of truth.

Capabilities I Owned

  • Custom pipelines and stages

  • Days-in-stage tracking

  • Call logs and follow-ups

  • Bulk lead actions

  • Filters by owner, stage, overdue

  • Lead imports (CSV, Google Sheets, API)

Every AI action reflected CRM state – never guesses.

Layer 3: Automation Before Intelligence

Before AI replies, I designed rule-based automation.

What Came First

  • Stage-based follow-up sequences

  • Time-delayed reminders

  • Pause-on-reply logic

  • Manual overrides at every step

  • Full visibility into active automation

AI should enhance predictable systems – not replace them.

This reduced risk and built confidence in later AI adoption.

Layer 4: AI Systems

RAG-Based Knowledge Grounding

I led the design of a Retrieval-Augmented Generation (RAG) system to ground AI responses in company-specific knowledge.

Teams could upload:

  • PDFs

  • CSVs

  • Website links

  • Internal documents

AI responses were generated only from this knowledge base, avoiding hallucinations and generic replies.

Knowledge base upload screen

Custom, Organization-Specific AI Agents

Each company effectively operated with a custom, constrained AI agent.

Agent behavior was shaped by:

  • Company knowledge base

  • Qualification rules

  • Automation logic

  • Confidence thresholds

The agent could:

  • Engage new inbound leads automatically

  • Ask qualifying questions

  • Draft responses

  • Flag intent

  • Update lead stage

All actions were visible, reversible, and override-able.

AI-Assisted, Not Autonomous (Critical Judgment Call)

I explicitly rejected full autonomy early.

AI was allowed to:

  • Interpret language ambiguity

  • Draft responses

  • Assist qualification

AI was not allowed to:

  • Take irreversible actions

  • Send messages blindly

  • Make payment or routing decisions

Principle
AI handles ambiguity. Rules handle responsibility.

Teams wanted AI support without risking trust, WhatsApp policy issues, or accidental customer messages, reinforcing human-in-the-loop design.

Operational Control & Business Hours

AI behavior was intentionally tied to active user context and business rules.

AI replies and automation ran only when the WhatsApp extension was active and within configured business hours, ensuring messages were not sent blindly or at inappropriate times.

This gave teams full control while still benefiting from AI assistance during active work hours.

QA, Safety & Evaluation

I personally led QA strategy across AI and non-AI flows.

What Was Tested

  • Edge cases across extension and dashboard

  • Logged-out and failure states

  • Automation conflicts

  • AI fallback behavior

Metrics I Defined

  • Manual effort reduction

  • First response time

  • Qualified lead conversion

  • AI override rate (trust signal)

AI success was defined as helping humans perform better, not replacing them.

Results & Impact

Across pilot teams and early adoption:

0 x

More Leads Processed

~ %

Faster Deal Closing

0 %

Fewer Lost Leads

~ 0 %

Reduction in Manual CRM Work

What Didn’t Work (And Why It Mattered)

  • Some users expected full automation → trust friction

  • Extension limits restricted background workflows

  • Early AI prompts needed tuning

These failures informed onboarding, positioning, and roadmap priorities.

Final Reflection

Kraya reinforced a belief I now apply to every AI product I lead:

Great AI products are not about intelligence – they are about judgment.

Judgment in:

  • Where to apply AI

  • Where not to

  • How fast to let it act

  • How to measure trust

That judgment is what I bring as a Product Manager.

What This Case Study Demonstrates

  • System-level product thinking

  • Responsible AI design using RAG and constrained agents

  • Strong UX and adoption instincts

  • Ownership across Product, Design, QA, and GTM

  • Ability to ship AI in real, high-risk workflows