From messy addresses to smarter decisions: How Snapdeal is building AI for Bharat

Achint Setia, CEO, Snapdeal.For most e-commerce businesses, an address is supposed to be a straightforward piece of data: a street, a locality, a PIN code and perhaps a landmark.In India, it can be something very different.It could be “near Hanuman Mandir, opposite Sharma Medical, Main Chowk” — an address that may make perfect sense to a local delivery executive but is difficult for a machine to interpret.

For Snapdeal, where more than 80 per cent of orders come from non-metro cities and smaller towns, this is not an edge case.

It is part of the operating reality of the marketplace.The same complexity extends to the seller ecosystem.

Snapdeal works with thousands of MSMEs with very different levels of digital maturity.

Some operate sophisticated digital catalogues, while others still rely on spreadsheets and manual processes.That has shaped the company's approach to AI.Rather than simply deploying larger models, Snapdeal is building systems that can understand India's real-world complexity, learn continuously from operational data and work despite imperfect information.“AI for Bharat is fundamentally different.

The challenge is less about building larger models and more about building systems that understand India's real-world complexity, learn continuously from operational data, and work reliably despite imperfect information.”The approach is visible in how Snapdeal is applying AI to some of the less glamorous — but more consequential — parts of its business: getting an order delivered, understanding where a customer actually lives, helping sellers come onboard and making information easier for employees to access.When delivery intelligence becomes a business capabilityOne of Snapdeal's most significant AI applications sits behind a simple customer outcome: whether an order gets delivered successfully.Retina AI has improved successful deliveries by around 3 per cent.

Snapdeal does not disclose the underlying delivery-success base or the specific cohort against which the improvement is measured, so the figure is best understood as the reported improvement rather than converted into an absolute number.At marketplace scale, however, even a relatively small improvement can have a wider operational effect.

Every additional successful delivery can reduce customer disappointment, lower reverse-logistics costs, improve courier efficiency and increase seller profitability.

Fewer failed deliveries can also mean fewer customer-support interactions and fewer Return-to-Origin shipments.What makes Retina AI different from a conventional fraud or risk model is where it sits in the workflow.It operates as a real-time decision engine before an order enters fulfilment, estimating the probability of different outcomes using hundreds of signals generated across the customer journey.

These include ordering and fulfilment behaviour, payment patterns, product characteristics, geography, courier performance, seller behaviour and transaction-level signals.The resulting risk score can influence whether additional verification is required, whether another courier could improve the probability of successful delivery or whether an order needs additional review.The objective is not to reject orders.

It is to increase the probability that a shipment reaches the customer while keeping unnecessary operational costs down.Learning from where the models got it wrongLike most production AI systems, Retina AI evolved through trial and error.Earlier versions relied more heavily on historical fraud indicators and static, rule-based thresholds.

They identified risky transactions, but also generated more false positives, occasionally flagging genuine customers whose behaviour differed from historical patterns.That forced the team to rethink the model.India's e-commerce behaviour varies across cities, regions, payment preferences, shopping occasions and seasonal events.

A rigid set of assumptions could therefore become outdated quickly.Snapdeal shifted towards continuously learning models that combine behavioural, transactional, geographical and logistics signals.

Just as importantly, it began treating operational outcomes as training data — allowing delivery results to feed back into the system and improve predictions over time.That learning is central to the company's broader AI approach: the best intelligence often comes not from a generic model, but from understanding what actually happens when a decision meets the real world.Rethinking what a “bad” address meansThat thinking becomes even clearer with Snapdeal Pinpoint AI.When the company started working on address intelligence, the obvious assumption was that the problem was “bad addresses”.

The team arrived at a different conclusion: Indian addresses were never really designed to be machine-readable.For a local delivery executive, “near the temple, behind the blue gate, opposite Sharma ji's shop” can be perfectly useful.

It is only a bad address if you are a computer.So instead of trying to make customers enter more structured addresses, Snapdeal chose to learn from addresses where deliveries had already succeeded.Pinpoint AI takes the customer's original input and enhances it using delivery location, patterns from nearby successful deliveries and other internal signals.

The system can also use customer-confirmed information when an address is ambiguous.It took close to two months of iteration before the team began seeing the numbers move in the right direction.Today, the system processes close to 1 million addresses every month, including across Tier 3 towns and beyond.The company says address-resolution accuracy has improved, although it is still in the early stages of measuring the full impact.Perhaps more interesting were the things the system taught Snapdeal.Address completeness, for example, was not the strongest predictor of whether a delivery would succeed.

Contactability mattered.

A customer could provide a messy address and still receive the package if the courier could reach them by phone.The team also found that metro addresses without a house number could form a high-RTO cohort, challenging the assumption that address problems would be concentrated primarily in smaller towns.Even precise geo-coordinates did not automatically solve the problem.

Addresses built purely from latitude and longitude were less deliverable than manually entered addresses, because human descriptions often contain contextual information that coordinates cannot — which gate, which floor or which shop a property is behind.The lesson is simple: AI works better when it understands the way people actually operate, rather than forcing people to operate the way the system expects.From AI tools to changed workflowsThe same principle is playing out inside Snapdeal.Its internal GenAI platform, Snapstride AI, gives employees a conversational interface to organisational knowledge.

Instead of searching across multiple systems and documents, an employee can ask a natural-language question and receive an answer grounded in approved internal documentation within seconds.The platform is now used by approximately 90 per cent of employees, with strong adoption across Customer Support, Product, Engineering, Marketing, Analytics and Operations.Snapdeal does not view usage alone as the measure of success.

It looks at time saved, reduced search effort, faster issue resolution, consistency of customer-support responses, employee satisfaction and repeat usage.

The fact that employees voluntarily return to the platform is itself considered an important signal of value.AI has also changed the seller onboarding journey.Two years ago, onboarding a seller could take a couple of weeks, with the process moving through multiple teams and manual checkpoints.Today, a seller can complete the process in a couple of days through a single self-serve flow with zero manual intervention.

It has become the intelligence layer running through our commerce platform, from how customers discover fashion and sellers create catalogues to how orders are fulfilled and employees access knowledge.

The biggest shift is that decisions across the organisation are increasingly informed by continuously learning systems that improve with every interaction.

Building AI for Bharat has taught us that solving local problems requires local context.

Models become truly valuable only when they learn from the realities of Indian consumers, sellers and logistics networks.”That philosophy has also influenced how Snapdeal thinks about its AI architecture.

The company builds proprietary systems where its own commerce, logistics and marketplace data create a meaningful advantage, while using external foundation models for more horizontal capabilities such as language, content generation and conversational interfaces.Looking back, Snapdeal says its most important decision was to embed AI directly into operational workflows instead of building isolated AI products.

If starting again, the company would invest even earlier in unified data infrastructure, recognising that continuously refreshed operational data is critical to production AI systems that improve over time.For Snapdeal, then, the AI journey is not really about making a marketplace more automated.It is about making it more intelligent about the reality in which it operates — messy addresses, diverse customers, fragmented seller networks and all.And that may be the more interesting story of AI for Bharat: not how well the technology understands the ideal world, but how well it learns to work with the real one.