Building AI Products For Indian Users: The Challenge Of Scale, Access, And Cost

Building for Indian scale means designing AI for millions of people with very different devices, languages, and connectivity. It is not enough for a system to work well on a fast phone and stable broadband. It must remain useful on mid-range phones and patchy networks while supporting many ways of speaking and searching.

These constraints interact. Limited bandwidth can slow responses. Budget hardware limits memory and processing power. Multiple languages make understanding and recommendations more difficult. Lower monetisation also means companies cannot simply spend heavily on every request. ShareChat and Moj therefore invest in data processing and recommendation systems that serve large audiences efficiently.

The challenge is broader than technical reach. A product must still feel competitive and affordable. Companies are exploring voice interfaces, open-weight models, and on-device processing, while building infrastructure that can adapt as models change. Indian scale therefore requires efficiency and flexibility from the start.

On-device AI means processing some AI requests on the phone itself rather than sending every request to a remote server. This can make features more responsive when connectivity is slow or unreliable. It can also reduce the amount of data that must travel across the network and lower server workload.

The trade-off is that phones have limited resources. A model or feature may need substantial memory and processing power. It can also consume battery and compete with other apps. Meesho is testing on-device AI specifically by measuring its effects on memory, battery life, and user experience, especially on lower-end devices.

On-device processing is therefore not a universal replacement for cloud AI. Companies must decide which tasks phones can handle well and which need remote infrastructure. If the balance works, users can receive faster, more dependable assistance despite patchy networks.

The reported figures show that AI can affect both growth and operating economics. Meesho reported 22% higher conversions among Vaani users than non-users in some Tier III and IV cohorts. That suggests the tool helped some shoppers move from expressing interest to completing purchases. Vaani was also activated by 1.5 million users in its first month.

Rapido’s figure highlights a different scale question: cost per transaction. Cofounder Rishikesh SR said the combined cost of acquiring models and employing people for AI work was less than four paise per ride. This gives the company a concrete measure for keeping experimentation financially accountable as usage grows.

These results are reported company figures, not a universal guarantee for every user or product. Still, they show two important outcomes: better discovery can support conversions, while careful infrastructure and governance can keep AI costs small enough for high-volume services.

When AI misunderstands a shopper’s intent, product discovery becomes less useful. The shopper may see irrelevant results, struggle to describe what they want, or abandon the journey. If the system cannot process a local language, people with limited digital experience may face an additional barrier to using the service.

Logistics errors create a more direct operational cost. Shadowfax deals with incomplete, inaccurately written addresses and incorrect pincodes. If AI maps such an address to the wrong location, a delivery can be delayed or misrouted. The company uses historical delivery data and AI-based address matching to improve location identification.

These problems affect both sides of a transaction. Customers lose time, convenience, or access. Businesses may lose conversions and spend more on delivery operations. Shadowfax puts the margin impact of unclear addresses at around 1%, showing why accuracy matters alongside speed and automation.

Companies are building switchable AI systems because models evolve rapidly. If an application is tightly tied to one provider, changing models may require rebuilding important parts of the product. That can slow improvement, raise costs, and create dependence on a single vendor or technical approach.

ShareChat and Moj are developing infrastructure that lets teams replace models, reuse contextual information, and compare performance after a switch. Rapido has also built software layers for working with different models, alongside rules governing their use. These layers separate the application from the model, making experimentation more manageable.

Open-weight models provide an alternative to relying only on hosted providers. They can give companies more control over deployment and adaptation, although the article does not claim they are always cheaper or better. Meesho is exploring them, along with on-device AI, to manage the cost of serving a large user base.

ShareChat and Moj use AI infrastructure, data processing, and recommendation systems for a large multilingual audience using mid-range phones and limited bandwidth. Their goal is to deliver relevant content efficiently while keeping costs under control. They also want the freedom to change models without rebuilding their applications.

Meesho applies AI to product discovery through Vaani, a voice-led shopping assistant. It helps shoppers express what they want, especially people with limited digital experience. Rapido takes a governance and cost approach. It provides employees access to advanced models through software layers and sets rules for their use, keeping reported model and people costs below four paise per ride.

Shadowfax focuses on logistics. It combines historical delivery data with AI-based address matching to identify precise locations from incomplete or inaccurate addresses. These examples show AI solving different business problems, from relevance and accessibility to affordability and delivery accuracy.

An AI model is a trained computational system that identifies patterns and produces an output from an input. Depending on its design, it might interpret a voice request, recommend content, understand product intent, or match an address to a location. Models differ in capability, size, speed, and resource needs.

At millions of requests, every technical choice multiplies. Larger or more capable models may improve accuracy but require more computing power and memory. Remote processing can provide stronger hardware, yet depends on network quality and creates serving costs. On-device processing can improve responsiveness during poor connectivity, but may drain battery or exceed a phone’s limits.

Companies therefore balance accuracy, speed, reliability, memory, battery life, and cost. Meesho is testing these trade-offs on lower-end devices. ShareChat and Moj focus on efficient infrastructure, while Rapido tracks AI spending per ride. The best model is not always the biggest one; it must fit the product’s users and economics.