Ankit Rawat, a Senior Software Engineer at Meta with over a decade of experience in e-commerce and payment systems, explains how the tools he built at Wayfair cut agent discovery time by 85% and call abandonment by 14% without replacing a single human operator.
Gartner’s January 2026 forecast landed like a cold shower on every company betting on AI-driven customer service. First: by 2030, generative AI cost per resolution will exceed $3, more than many offshore human agents charge today. Rising data center expenses and AI vendors pivoting from subsidized growth to profitability will drive that price up, not down. Second, and less discussed: by 2028, regulatory changes mandating easy access to human agents will increase assisted-service volume by 30%. Patrick Quinlan, Senior Director Analyst at Gartner, put it bluntly: full automation will be prohibitively expensive for most organizations.
Read those two predictions side by side. AI is getting more expensive, and the number of calls that require a live human is about to surge. For any e-commerce company that spent the last two years betting on chatbots as a cost-cutting play, this is a problem without an obvious exit.
Unless someone already built the tools that make live agents radically more effective, which is exactly what Ankit Rawat did at Wayfair before most companies realized they would need them.
Currently at Meta Platforms in Seattle, where he focuses on performance and reliability for WhatsApp Business flows, Rawat spent the previous two years at Wayfair AI-assisted support tools that don’t replace agents but make them dramatically faster and more accurate. His path ran through financial data management at his first job out of university, then e-commerce at Snapdeal, then five years inside Amazon’s logistics and retail checkout teams before he landed at Wayfair and eventually Meta.
What has to work before AI can help
Not every support call is created equal. A customer asking “where is my package?” costs relatively little, and a chatbot can handle it. A professional B2B buyer calling because a 90-item checkout flow just failed mid-transaction is an entirely different situation. Without deep context, the agent cannot diagnose the problem. Meanwhile, the buyer is already frustrated. A cart worth thousands of dollars is hanging in the balance. No chatbot resolves this call, because the issue is not a question – it is an infrastructure failure that created a crisis only a live agent with the right information can untangle.
At Wayfair, where Rawat joined as a Staff Software Engineer in 2021, these were exactly the calls that cost the most and escalated the fastest. Professional B2B customers needed shopping carts capable of holding far more items than the platform allowed. Thirty items were the ceiling. Bulk buyers found that limit absurd, and when the system buckled under real-world usage, the resulting support calls were the most complex and expensive in the entire operation.
Together with senior engineers across multiple organizations, Rawat led performance work that pushed cart capacity from roughly 30 to over 90 items, enabling approximately $150 million per year in B2B business at about 7% annual growth. Larger carts meant more complex checkout flows, more failure points, and more edge cases where a transaction could quietly break. But fixing the checkout was only half the job – the other half was ensuring that when something did break, the agent picking up the phone could actually help, quickly and with full context. And that is what Rawat built next.
Making agents faster without replacing them
Most of the current AI hype assumes a world where companies keep trying to automate entire conversations. Rawat’s work at Wayfair took a different path – one that looks increasingly prescient.
Rather than building a bot to talk to customers, he built a low-latency tool that reduced the customer discovery phase for support agents by about 85%. When a customer called, the system surfaced relevant context before the agent even picked up, integrating directly with telephony to identify the specific caller and pull their history instantly. Context that previously required manual searching across multiple systems now appeared on screen within seconds.
Think about what that means economically. A support agent handling B2B clients with 90-item carts and complex payment questions needs deep context fast. An agent who spends the first chunk of a call digging through order history loses the customer’s patience, and the cost per interaction balloons. Cutting that discovery time by 85% did not just speed things up – it changed the economics of every call.
Separately, Rawat led the design of a GenAI-powered conversation summarizer. Long customer histories across chat, SMS, and email were condensed into short, actionable summaries highlighting the key issue, actions already taken, and unresolved items. Call abandonment dropped by approximately 14%. Not because a bot handled the call, but because a human agent could finally do the job without drowning in context.
“The biggest challenges were always operating with incomplete information and aligning teams with different priorities,” Rawat says. “You address that by iterating quickly, setting clear success metrics, and building tight feedback loops with stakeholders.”
Why scale kills shortcuts
AI in support does not happen in a vacuum. Every tool Rawat built at Wayfair had to survive in an environment where he also drove operational and engineering excellence practices across teams totalling over 50 engineers and more than 10 active projects.
One example captures the difficulty. In initial testing, the conversation summarizer worked well: clean inputs, predictable conversation patterns, manageable volume. Once dozens of agents started using it simultaneously across different workflows, the edge cases multiplied. Interactions spanned chat, SMS, and email, often for the same issue, and the summarizer had to collapse all of that into a useful brief without losing critical details or confusing one case with another. Rawat’s team built consistent templates and validation loops that checked summary quality automatically, catching errors before agents ever saw them. Getting this right was not a model problem – it was a data discipline problem that only surfaced at scale.
“The key principle is prioritizing what the agent needs right now: recent interactions, current status, next-best actions – instead of showing raw data,” Rawat explains. “That is how you help agents ramp quickly and improve continuity across handoffs.”
What the support budget really buys
Companies keep asking the wrong question. They ask which AI model to buy. But the real question is what their agents need to be effective and what infrastructure has to exist before any model can help.
None of these solutions depends on having the newest or most expensive model. All of them depend on years of solving messy problems in environments where a broken system meant lost revenue and angry customers, not a postmortem slide nobody reads.
E-commerce support is not an AI problem. It is an engineering problem that AI can help with, if the foundation is already there.
