AI Series Part 2 - AI in action: Case Studies from the Indian ecommerce and digital-native companies

#redbus #meesho #razorpay #Agentic-AI #AI #e-commerce #retail #Consumertech

Hello Hello,

In the previous post, AI series part 1 - Marketplace and digital-native companies 3.0: A Primer on AI-Powered commerce, I discussed likelihood of agentic-driven commerce and built upon scenarios based on market trends, and rapid development in the AI field. I hope you enjoy reading it.

A few years ago, during a Disney Strategic Foresight workshop, Richard Ramsey—then a VP of Latin America HR—shared something very profound with a group of about 25 of us that has stayed with me ever since: “The future is not evenly distributed.” Nearly a decade later, those words still seems so true and relevant to me.

Whether AI will level the playing field or further widen the gap remains up for debate. But one thing is clear: e-commerce and digital companies are racing to embrace AI—compelled to act and stay relevant so that they are not on the left tail side of the distribution.

According to an IBM study,

Indian Companies are Investing in AI for the Long Term, Intend to Use Open-Source Tools to Drive ROI and Innovation.

So In this post, I intend to share some of the use-cases that are currently implemented by ecommerce and Internet companies - Meesho, Razorpay, and Redbus. I selected these three from different domains to offer a broader view of the current AI landscape.

Some of the use cases I’m sharing are drawn from my own consulting experience with SMBs, MSMEs, and early-stage startups. As an advisor, I work closely with founders on driving revenue growth and exploring practical applications of AI across e-commerce and Martech workflows.

If you're looking to automate workflows or leverage AI in your business, feel free to DM me on Linkedin—always happy to help.

So let’s start without further ado.

---

Case study -1 Redbus leveraging AI for analytics, CX and to improve operational efficiency

Redbus is currently embedding AI into their existing workflows. Last week, I got a chance to connect with their analytics team and learnt about some of the work-flows where team is leveraging AI and agentic-AI to improve CX and gain operational efficiency.

Below are a few of the use-cases that the team is currently working on:

A. To improve customer experience, Redbus shares AI-driven insights and recommendations with bus operators for focused interventions:

Redbus has millions of reviews from customers who travel on buses operated by bus operators through the Redbus platform.

The team uses GPT-4 enterprise model to categorize feedback into key parameters – Vehicle quality (Speed, rash driving), service quality (Cleanliness, amenities), driver performance etc and extracts granular insights on customer experience.

The model not only considers high-level trends, but also brings out sub-issues such as “AC malfunction” or about “rest stop poor hygiene”. This enables a deeper understanding of customer pain-points.
Once GPT-4 model produces the insights, Redbus can flag low-performing buses or operators ((e.g., “Sangam Travels Bus #5 has 12 complaints about broken charging ports”) and can also share recommendations for targeted interventions with their bus partners. This help in improving service standards across the customer journey.

This kind of granular-level analysis and extraction without AI would have been expensive, manual, error-prone, difficult to scale, and lacks real-time capabilities.

B. AI-powered competition and market gap analysis can help product team to improve product features.

Another use-case where Redbus is using AI is competition analysis. Another use-case where Redbus is using AI is competition analysis. Through APIs, The aggregates app store reviews and benchmarks specific features against competition’s features i.e. competitors Y’s real-time tracking rated 4.8 vs. our 4.2”

By comparing feature performance, the analysis reveals gaps—for example, Redbus lags behind Competitor A in refund speed, indicating a key area for improvement.

C. AI-Generated Summaries & Smart Tagging to Enhance UX

Redbus is working on leveraging AI to make concise, meaningful summaries out of complex customer reviews, helping users make faster, and more informed decision. This is a very common use-case across ecommerce. Amazon executes this exceptionally well. While I don’t see this particular feature on Myntra or Flipkart.

D. text-to-SQL tool - Agentic-AI for analytics function and SQL for internal team

The category team often needs to run SQL queries to evaluate various performance metrics — and at times, these can be complex, multi-line queries that require advanced coding.

Redbus has made it easier for the internal teams to run queries like, “What’s the order share of the top 10 Bangalore-Hyderabad buses?” which is basically Number of orders for a bus divided by Total number of orders for that route. First you have to filter data by route and a timeframe, then you identify top 10 buses, then you calculate total orders for the route. Phew! You got my point. It’s too much work.

With agentic-AI workflows, you can run query database using natural language. Behind the scenes, orchestrator agents break down the task into intent parsing, metadata retrieval, SQL generation, and dataset merging. It can do complex metric calculation such as calculating order share or conversion rates in a minute.

And all of that, without writing a single line of SQL :)

(Marketers can use AI to get insights - like “What was the sales impact of our last campaign?” or track KPIs such as cart abandonment rates by region— without relying on data scientists, accelerating decision-making across teams.)

E. Smart Filters

The team is currently working upon intelligent, dynamic filtering systems that go beyond traditional dropdowns or checkboxes by leveraging machine learning, user behaviour, and contextual data to surface more relevant results with minimal manual input.

Imagine having a natural conversation with an AI tomorrow, where you casually share pain-points that you've experienced. For example, you might tell Redbus’s chatbot that during your last trip, the berth was too short for your height, and since you're tall, you'd prefer a bus with a longer berth next time. And next time you book the Redbus, it remembers that you are tall and you would need a longer berth, even when you don’t. Wouldn’t that be a great way to delight customers? :) Making them feel truly valued and understood.

Key implementation insights that can be learned from Redbus Playbook :

1.	LLM Selection: Test different models (GPT, Gemini, Claude) for fit; Redbus selected GPT-4 for its accuracy in travel-specific contexts.
2.	Data Privacy: Use synthetic data during prototyping to avoid privacy risks, and switch to enterprise APIs for scalability.
3.	Knowledge Base Tuning: Feed LLMs with structured metadata such as product specs, refund policies, or category taxonomies to improve output accuracy.
4.	Token Management: For large-scale datasets (e.g., >1M tokens), use enterprise-grade solutions like GPT-4 Enterprise.
5.	Agentic Workflows: Break complex processes into modular AI agent tasks (e.g., parsing → retrieval → analysis → output), especially in data-heavy tasks like text-to-SQL.

Companies can use actionable insights and AI-driven feedback analysis to ship new product features and product development. For instance, in case of Redbus -

Customer feedbacks that Redbus receive are highly specific and contextual. It can be used by Product and brand team for specific targeted improvement.

For instance, if there are getting recurring complaints about “App is difficult for non-English speaker” and there is a significant customer demand at regional-level, the product team can roll-out platform experience in regional language or voice-based search or booking.

AI can be used to close the gap between customer pain-points and product solutioning.

Another example — if customers are consistently commenting on difficulty with live bus tracking when reviewing competitors, Redbus can recognize it as a key differentiator. The brand can then double down by highlighting real-time tracking as a core value proposition in its campaigns — turning a product feature into a strategic lever to win market share.

Case study 2 : Razorpay is using Agentic-AI to automate Customer (Merchant) Support to bring cost efficiency and CX

Razorpay serves over 1 million customers and handles a high volume of merchant queries. It needs a robust support model to ensure that these queries are addressed promptly and at scale.

Manual handling of L1 merchant queries (Level 1 support) is slow and inefficient. Customer satisfaction scores is low due to delayed responses. There are complex internal workflows with ~500 conditions based on severity, priority, type of query that needed to route correctly. Then there are multiple team handoff required, creating delays and SLA management.

To solve this problem statement, Razorpay deploy agents that handles L1 merchant support queries automatically. It pulls answers from the internal documents and Freshdesk knowledge base. AI scans incoming tickets against 500+ predefined conditions to determine severity, priority, and the right team.AI scans incoming tickets against 500+ predefined conditions to determine severity, priority, and the right team. Dynamically sets and updates “promise time” (expected turnaround time) based on issue type. It understands the customer tone to frame appropriate response. It then prioritize and intelligently route.

Escalation : Agents can also manage the escalations. If SLA is breached, AI triggers the correct escalation workflow (who to escalate to, which team, etc.). And then track SLAs for each ticket.

If AI can’t resolve the query or the customer is dissatisfied, it escalates to an L2 human agent.

Case study -3 Meesho is working on leveraging AI across customer touch-points.

I had the opportunity to connect with leaders at Meesho who are spearheading AI initiatives. Across the board, they emphasized that AI is not just a tool but a central pillar of Meesho’s strategy. From customer support and seller tools to search, operations, and marketing, AI is deeply embedded in every critical business function. Notably, Meesho has also built an in-house AI Services team with the vision of offering its capabilities beyond the company—transforming internal excellence into a revenue stream.

A. Gen-AI powered Bot for customer support as well as seller support

Similar to Razorpay, Meesho is leveraging AI to drive customer support automation and cost optimization at scale.

Recently, It has launched GenAI-powered voice bot which can handle 60K calls per day in Hindi and English. It is planning to expand this to more regional languages.

We all have experienced frustration talking to Chatbots that keep us asking the same set of questions that feels like waking up in a time loop. Now, imagine the frustration of trying to communicate with a voice bot who doesn’t understand you. It becomes more important to have a voice bot with minimum lag and context-aware and have a real-time conversational capabilities.

The moment it sounds robotic or delayed — like there's clearly no human on the other side — it can instantly break trust for the seller.

This real-time conversational AI that Meesho has set up, it is claimed to have significantly reduced cost per call—currently at just 25% of what human agents cost, with expectations to lower it further to 15% as the system scales.

B. Vernacular search and Discovery

Meesho is building for Bharat. Hence, It becomes important to take India-specific context and languages into consideration.

In case of Meesho, you get search queries like - "lehenga dupatta for sister wedding pinkish red." To run this query, LLM requires high-quality, diverse dataset in Hindi, Hinglish, and regional dialects.

Traditional systems might struggle to understand this . To get the result, Base models (like GPT or LLaMA) need to be fine-tuned specifically on Indian queries, with reinforcement based on correctness and intent understanding.

Once implemented, It can understand what the user really wants and shows better results. Over time, AI also learns from how millions of people search, click, and buy—just like how language tools learn from reading lots of text. This helps Meesho build better systems for showing the most relevant products, improving what you see when you search or browse—making shopping feel easier, more personal, and spot-on.

C. Seller operation (Product listing, images, analysing market trends, and price competitiveness
In a recent talk that I attended at SPC, Meesho’s founder Vidit shared an insightful story about their early days. He spoke about spending significant time visiting local merchant shops that they aimed to on-board. A key insight that he shared: these merchants wanted to expand their business reach beyond the usual 5-kilometer radius but lacked the time, resources, or digital know-how to manage operations online.

To bridge this gap, Meesho initially empowered women running households—those eager to contribute financially from home—to handle these online operations.

Today, that role can increasingly be taken over by AI. AI now can play a central role by automating key functions for the sellers: from cataloguing products and generating images or videos to analysing price competitiveness. AI-powered Chatbots guide new sellers through the onboarding journey in real-time, making it easier for first-time entrepreneurs to set up shop. Advanced AI tool has potential to help sellers optimize pricing based on real-time market trends, while vernacular translation tools convert local-language product descriptions into standardized, platform-ready formats.

D. Operation and supply-chain

Meesho relies on predictive AI models for demand forecasting, fraud detection, and logistics optimization.

E. Marketing

Generative AI is used to automatically create product descriptions, email campaigns, and social media content, significantly reducing time-to-market for new campaigns.

Case study 4: Automating complex task across Martech, D2C

Best way to learn is to do it.

I built AI agents to automate marketing workflows - One of the agents that I created can plan, write, and schedule social media posts from a podcast in just minutes. It understands the brand tone, picks up on trends, and even generates images or hashtags.

If you’re curious about the details, I’ve shared more here :

Another agent helps with ad campaigns. It looks at past performance data (Previous campaigns), suggests what kind of ad to run, writes the copy, and even recommends how much budget to spend where.

These are some more agents that I am currently building for clients:

• Working on deploying a 24/7 intelligent AI assistant for a D2C brand website that can handle FAQ, order tracking, return processing, and product recommendations autonomously — without human support (Tools - Whisper API + Gumloop + Zapier + GPT-4)

• Creating an end-to-end AI-powered email system for the D2C brands – Sign In, Trigger – cart abandonment, order placement, Discount campaigns etc.

• Automating entire catalogue operation – Image, Video, Campaign design, content

Key Implementation Insights from my hands-on Experience with AI Agents

1. AI Improves Through Feedback Loops:

   One of the most critical aspects of working with AI agents is enabling continuous learning through feedback. For example, if you're using an agentic-AI framework to manage social media posts, the agents should learn and optimize based on audience engagement metrics like impressions, views, likes, and shares.

   On the surface, the idea that "AI improves with feedback" might sound obvious or even cliché—especially to those familiar with machine learning or AI operations. However, in practice, many AI implementations fail precisely because teams don't build robust feedback loops into their systems. Teams may gather insights but don't update prompts, fine-tune behavior, or retrain components regularly.

   A true agentic-AI setup should autonomously interpret results and iteratively improve—without manual intervention. That’s hard to build.

   Many companies deploy AI agents without clear access to post-action metrics (e.g., engagement per post, click-through rates, etc.).

   When I was working on AI agents to convert YouTube podcasts into social media content, the initial version didn’t include a feedback loop. I incorporated it in a later iteration.

2. Establish Clear Guardrails:

   
   It's essential to put guardrails in place to ensure AI agents operate within defined ethical, brand, and performance boundaries. This helps avoid unexpected outcomes and ensures reliability. For examples, if you are building an agent that can automate campaigns, you might need to incorporate brand guidelines.

   But more on this in the part 3 of this series in which : I will lay out the AI safety framework and non-negotiable guardrails that e-commerce and marketplace companies must establish to build responsibly.

   
   

   My suggestion to build an AI agents or agentic-AI - start with repetitive tasks:
   

   Identifying repetitive, time-consuming tasks within your organization. AI agents are highly effective at taking over such tasks, which often take humans hours or even days to complete, enabling teams to focus on strategy and creativity. Define the goal, break down those tasks into clear, structured workflows with defined inputs, outputs, and decision points, identify the right tool (APIs, LLM model), deploy and test iteratively.

The decreasing cost of computing and storage is making possible for companies to deploy AI across functions. AI is not a moonshot anymore, it’s a table stake. Companies that act early—experimenting, learning, and iterating—will shape the next era of digital transformation.

With this final thought, I say adieu. I would love to connect with more professionals who are building and implementing agents and learn more about the use-cases in the Industry. Feel free to DM or comment below.

Keep learning, Have a great weekend🎉

Thanks,
Neetu

© 2025 [Neetu Barmecha]. All rights reserved.
This content is the intellectual property of the author. Do not reproduce without permission.