How LLMs Cracked the Voice-of-Customer Code to Build Billion-Dollar Products

Here's the dirty secret about product development: customers rarely tell you what they actually need. They'll complain about app navigation when what they really want is faster discovery. They'll rant about delivery delays when they're craving hyper-local restaurant suggestions. They'll request "easier ordering" when they need conversational, intention-aware commerce. The gap between what customers say and what they need has always been the difference between market success and spectacular failure.

Enter large language models—and suddenly, this age-old challenge has a surprisingly elegant solution.

The Stadium Seating Revelation

Professional analysts have long decoded customer complaints into underlying needs: “clean, unobstructed view; spacious, well-cushioned seats; room to stretch legs; easy access”. These insights—called "customer needs" in Voice of Customer (VoC) research—historically sparked industry transformations. But recent research from MIT and Northwestern reveals something remarkable: fine-tuned large language models (LLMs) can now extract customer needs as well as—or better than—professional analysts.

From Brazil to China: How Global Companies Decoded Hidden Needs

Brazil's iFood, handling 80 million orders monthly, faced an existential problem: personalization at scale. The company discovered that customers didn't just want "food delivery"—they wanted conversations about their intentions. In June 2025, iFood launched Ailo, an AI conversational assistant that understands contextual needs. Instead of browsing menus, users say things like "I want a romantic dinner tonight" or "I'd like a light meal that arrives in 30 minutes," and the AI delivers tailored restaurant suggestions.

Early results? 48% higher conversion from search to shopping cart, and 33% faster order completion on WhatsApp than the traditional app. Behind the scenes, iFood's system runs over 100 AI models on every single order, analyzing recommendation patterns, fraud risks, and delivery optimization—all powered by LLMs integrated with AWS infrastructure.

Meanwhile, in India, Swiggy discovered a critical customer need: restaurants were showing up for people who'd never order from them. Using machine learning to analyze 17 million monthly transactions and 3.5 billion cumulative orders, Swiggy trained AI models to show hyper-local availability. Bachelor-heavy neighborhoods see snacks and soft drinks recommended; family areas see groceries and staples. What customers actually needed wasn't more restaurants—it was the “right” restaurants at the *right* time. The result? 20% increase in order frequency and 24 million monthly active users by 2024. To solve the merchant side, Swiggy created agentic AI for stores that analyzes sales patterns and give restaurant owners personalized growth strategies using natural language analysis.

Singapore-based Grab operates across eight Southeast Asian markets with 42 million monthly transactional users. The company's hidden customer need discovery came from analyzing driver frustration: partners were spending time in low-demand areas. Using LLMs and machine learning,

 Grab launched GrabRideGuide in 2024, an AI tool that predicts high-demand zones in real-time, helping drivers optimize routes and maximize earnings while ensuring passengers get rides faster. Another innovation addressed merchant pain: Grab's photo-to-text AI allows restaurant partners to upload menu photos, and generative AI automatically categorizes items, saving hours of manual data entry. For delivery customers, Grab's NLP engine interprets custom delivery notes ("leave it on the blue gate"), stores them, and makes them available for future orders.

TikTok's Global Content Intelligence

ByteDance's TikTok demonstrates LLM-powered personalization at a planetary scale: 170 million US users, 700 million on Chinese Douyin, plus 300 million on Toutiao news app. The company's core discovery: customers don't have a single preference—they have contextual, evolving preferences. TikTok's AI recommendation engine analyzes implicit signals—video watch time, scroll speed, rewatch frequency, time of day—to identify what users actually want, not what they claim to want. This wasn't built on surveys asking "what content interests you?" It was built on LLM-powered behavior analysis. The result? International revenue surged 95% year-over-year to $39B in 2024, with TikTok Shop integration showing the model extends beyond content into commerce. 

iFood's Ailo: Conversational AI as Customer Need Discovery

Beyond recommendations, iFood recognized a meta-need: customers wanted AI that understood intent. Ailo, built on a hybrid architecture combining Anthropic, OpenAI, and AWS models, interprets subjective desires—"I'm craving something filling but healthy" or "romantic dinner tonight"—and translates them into actionable restaurant suggestions. Since June 2025, early pilot data show Ailo increases shopping cart conversion by 48% and reduces order completion time by 33%.

The Technical Breakthrough

Research from MIT and Northwestern examined whether LLMs could automate the extraction of customer needs from unstructured data. A supervised fine-tuned (SFT) LLM trained on just over 1,000 examples outperformed professional analysts in blind evaluations. For product categories, the LLM identified 100% of primary and secondary customer need categories while human analysts missed 12.5% of primary groups. The LLM maintained 92% accuracy without hallucinations 

The Playbook for Market Research

For management and data science students conducting market research:

Start with unstructured data: Online reviews, support tickets, user behavior patterns—these contain goldmines of insight.

Fine-tune LLMs for your context: Base models without fine-tuning underperform; companies like iFood and Swiggy achieved breakthroughs by training models on regional, behavior-specific data.

Scale intelligently: Swiggy's 3.5 billion order dataset and iFood's 80 million monthly transactions generate signals that train increasingly accurate models.

Balance breadth and specificity: Customer needs aren't "ease of use" (too generic) or "checkout in 2.3 seconds" (too specific). They capture jobs-to-be-done at the right abstraction level—like Grab's "drivers want to maximize earnings" or iFood's "users want conversational ordering".