What do people actually think? Knowing that is the difference between a product that finds a market and one that misses it entirely. Market research is the $140 billion industry built to answer that question. It sits beneath nearly every consumer brand, enterprise product, and government in the world. But for an industry this foundational, how it works is surprisingly unchanged. The year is 2026, and market research runs almost exactly the way it did in 1975.

The structural problem with the existing model

Every unit of research requires a distinct human at each step:

• A recruiter to find respondents
• A screener to qualify them, a moderator to run the session
• A translator to cross language barriers
• A coder to extract meaning from what came out

Scale is simply a function of headcount, with no way to run a study faster without hiring more people.

Time is the second constraint. A human moderator can run three to five interviews a day: a ceiling that’s shaped by scheduling friction, time zones, no-shows, and the cognitive weight of the work. These logistics alone consume weeks before a question is asked. Transcription, coding, and synthesis then take weeks further, only at the end of which are raw recordings turned into usable insight. End-to-end, a qualitative study takes four to ten weeks, and that is considered normal.

These structural constraints produce a predictable cost. A single participant through a qualitative interview runs between $500 and $5,000; a single-market study costs $25,000 or more: figures that narrow the addressable market to large incumbents with dedicated research functions. And even they cannot solve for pace. The calendar constraint is often more binding than the budget. For everyone below that tier, a two-person insights team at a startup or a single product line inside a larger company, the price is simply out of reach.

Research, as a result, happens once a quarter for those who can afford it. For everyone else, the default is shipping with incomplete information and hoping their instinct holds.

The incumbents' own economics confirm how entrenched this is. Each of them spend more on payroll than they generate in operating profit, and while the industry grew 35 - 40% between 2021 and 2025, almost none of that expansion reached the bottom line.

What AI changes

Think about how legal discovery worked before e-discovery software. Every unit of output required a proportional unit of human attention: armies of associates, warehouses of documents, billing rates that made litigation the exclusive domain of large corporations. E-discovery didn't make lawyers faster. It broke the linear relationship between document volume and human hours entirely.

Market research had the same constraints and was never able to escape them. AI breaks it the same way, and the TAM implication goes beyond expanding the existing market. Companies that couldn't afford research at $25,000 can afford it at $2,500. This structural unlocking with AI creates a market that didn’t exist before. It is happening across 2 main tracks:

The first is AI-moderated interviews. A cohort of companies, Listen Labs, Outset, Strella, Knit, Genway, Conveo, have built agents that run the full qualitative cycle. They recruit, screen, conduct interviews, transcribe, code themes, and deliver a finished report. Cycles that once ran ten weeks are compressed into two, and the costs come down 80–90%. The agents work across languages and time zones at a volume that human teams cannot match.

HubSpot's two-person insights team ran more than 100 AI-moderated interviews on how the AI era is reshaping their customers' needs, passing the findings directly to their marketing and product teams. Chubbies Shorts, a US D2C apparel brand, used AI interviews to reach children and parents for a new kids' line; a demographic that traditional focus group scheduling has always struggled to accommodate.

The second track is synthetic users, that is, AI-generated respondents built from the traces a population actually leaves behind through transaction histories, survey responses, social media behavior, and geospatial patterns. When enough of those signals are fused, the output is a simulated population you can query like a real one to pressure-test a product launch or a pricing decision without convening a single focus group. Companies like Aaru, Simile, and Vectorial are trying to solve this problem.

While exciting, the synthetic users as a category are still being stress-tested. The grounding problem is the central unsolved challenge. Most synthetic personas today are too close to LLM priors: coherent, plausible, and subtly wrong in ways that are hard to catch without a ground truth to compare against.

It’s similar to the early days of quantitative finance. Quant models in the 1980s reproduced historical market behavior with impressive accuracy, until novel conditions exposed the gap between what the model had seen and what was actually happening. Synthetic users face the same failure mode. Reconstructing how a known population responded to a known survey is the easy version of the problem. Predicting how people respond to something genuinely new, a product category that didn't exist, a price point outside historical range, is where the approach either proves out or breaks down.

Two related challenges sit underneath this. Demographic representation: the populations that leave the richest digital traces aren't always the ones that matter most for a given study. The companies that build transparent validation frameworks, showing clients not just where the model is accurate but specifically where it fails, will earn the kind of trust that makes them core infrastructure.

What we believe

The durable winners will be the companies solving the problems incumbents are choosing to leave alone. No one has cracked grounding yet, the work of anchoring synthetic personas to real behavioral signals, and the company that does will own the data layer underneath every downstream research tool.

India has a specific angle. Consumer tech platforms here sit on uniquely rich behavioral data ranging from purchase patterns and delivery frequency to price sensitivity. A panel-aggregation layer built on that data is a defensible India-first wedge.

Separately, the back-office of market research is already outsourced to India today through WNS, Genpact, EXL, Course5, and the captive centers of the incumbents themselves. An AI-native services firm could collapse those costs and take the margin.

Here's our full deep-dive on the space.