Synthetic Research Only Works If the Humans Underneath Are Real
At Cannes Lions, NewtonX CEO Sascha Eder made the case that simulated buyers are only as good as the verified data they are trained on.
Whatever else Cannes Lions 2026 was about, it was about AI. “I asked someone what they thought was the main theme. Obviously, this year you hear AI everywhere. I feel like it’s much more prevalent than last year,” Sascha Eder told me when we sat down at the Palais. It was NewtonX’s second time attending the Cannes Lions Festival and a more organised one, he said: a team of four, a diary of client meetings, and a session on synthetic B2B buyers that put him on stage alongside Jamie Cleghorn, Global Head of Customer Practice at Bain & Company and Zoom’s CMO, Kimberly Storin.
Eder is the CEO and co-founder of NewtonX, the New York-based B2B research intelligence company he started after a spell at McKinsey, where the frustration of advising clients without access to decent data became the founding thesis. The company now works with more than 600 of the world’s largest organisations, including Google, TikTok, Salesforce, Stripe, Microsoft, and Coinbase, supplying what he calls hard-to-get information: the verified professional insight that lets an enterprise, a consultancy or a hedge fund make decisions it would otherwise be guessing at.
The bots reached the research panels before the buyers did
The reason NewtonX exists, Eder argues, is that the market research industry has never fixed its supply chain. “For the last 25 years the industry has been rampant with fraud, and that’s being exacerbated by AI. More bots, even more sophisticated fraud.”
The mechanics of that fraud have evolved fast. “These days you have to deal with sophisticated people impersonating on LinkedIn, you have to deal with bots, you have to deal with AI agents that can simulate a real response in a survey that you almost can’t distinguish from a human.”
His answer is a verification stack rather than a single check. The NewtonX Graph indexes over 1.1 billion professionals, enriched over the years with signals such as IT and advertising budgets by industry and company size. When a client wants advertisers with retail budgets above $20 million, the company pinpoints them, recruits them across multiple channels, then verifies them individually: corporate email, phone, social profile age and authenticity, and topic-specific testing on every project. “There’s not one piece of verification that is bulletproof. It’s really the combination of multiple layers that makes it safe.”
The panel model, in which respondents self-report, is where the rot concentrates. On a panel, he notes, people designate themselves as a CFO one day and a CMO the next to qualify for more surveys. Pre-targeting from verified data closes that door.
Cheaper and faster are conditional claims
Verification is the plumbing. Synthetic personas, the subject of Eder’s Cannes session, are what the industry queued up to hear about, and the appetite is easy to understand. They promise research without recruitment: an always-available simulated buyer that answers immediately and at scale.
Eder’s definition carries a condition that most of the market skips. “A synthetic persona is an AI model that is built, especially in the B2B space, on real verified buyers. If you try to sell to an IT buyer that sits in mid-market finance companies, the model learns from those. The premise of synthetic is it’s cheaper, it’s faster, it’s immediate, and you can do much more at scale, but people tend to forget these things only hold true if the model actually works.”
He is equally deliberate about what the technology is for. “It’s not about whether synthetic research is better than real research. It’s about the use cases where teams don’t use any research at all, where they guess the answer, or wait weeks or months to run a real study.”
On familiar territory, the pitch is arithmetic. “Instead of launching 100 ads, assessing them and paying $50,000, you can launch 1,000 ads, have the synthetic personas pre-filter them down to 20, launch those, spend $10,000 and have as much or more conviction than the old method.”
He tells clients to walk away when the concept is genuinely new. “If you say, here’s the new hardware device OpenAI is launching that’s supposed to replace the iPhone that no one has ever seen, and you ask a synthetic audience, that’s not going to work, because it’s a brand-new concept.” The same logic applies to money. A new product line with $100 million to $500 million behind it gets primary research, no argument.
Agencies may be the buyers most exposed to that trade-off. Demand from them has exploded in two years from a standing start, and for independents priced out of enterprise-grade B2B data, he is blunt about what synthetic offers. “B2B data is expensive, high quality is expensive, but this is one of the areas where synthetic can provide a lifeline. Is it the same as primary research? No. But is it better than these agencies not having access to any data? Yes.”
Nobody wants to talk about stale personas
Verified data ages, and a persona trained on last year’s market is quietly answering last year’s questions. Eder concedes this is the part of the pitch the industry avoids. “Nobody talks about it, either because they don’t want to do it or because they’d rather not talk about it. The reality is you’ve got to update the personas every few months.”
NewtonX builds proprietary personas for clients and refreshes them roughly quarterly, using what it calls persona shells: behavioural and attitudinal segmentations for specific buyer types, enriched with the client’s own data. Refreshing every three months, he argues, still costs less than running a heavyweight brand tracker.
Validation is done the hard way: the company splits a real audience in two, builds the model on one half, runs live research on the other, and compares the results. “Confidence, depending on the topic and the audience, can vary anywhere from 85% to above 95%, which is extremely high in some cases. I would argue even 55% or 60% would be more than sufficient, because the trade-off is not making a decision based on any data at all, and there your confidence gets close to zero.” The effort is led by a VP of Research and Innovation who previously ran research at Salesforce-owned Tableau.
Half the buying committee never shows up in the CRM
Push on the hardest objection, that B2B purchases are shaped by politics, budget cycles and relationships that never appear in survey data, and Eder does not retreat into the model. “People are irrational. People say one thing and do another.”
His response draws on the multi-year buy-ability research NewtonX has fielded for LinkedIn and Bain, which mapped how enterprise buying committees behave. “You have the buyers everyone thinks of, and then you have hidden buyers, people in the background nobody is aware of, but they influence the decisions. If you’re not aware of it, you’re at risk of losing out.” The published findings put roughly half of decision-making influence with those hidden buyers in procurement, finance and legal. Personas built on that foundation can model committee dynamics to a degree, he says, but only because the primary research identified the dynamics first.
That insistence on primary data is also what separates B2B from consumer work. Consumer synthetic tools benefit from an ocean of training data; B2B does not. “People say, well, synthetic worked for B2C, so if I train my model on the web, I’m sure I can ask it to be a procurement specialist. That’s where the model doesn’t have the data, but makes up the audience nonetheless, and gives you a confident-sounding answer.”
The replacement narrative did not survive contact with the data
NewtonX’s study with The Wall Street Journal, which surveyed 82 brand marketing VPs on generative AI, went in carrying three hypotheses: that marketers expected AI to replace people, that it was already generating strong returns, and that cost-cutting was the point. “All three were debunked.” Marketers wanted efficiency and better campaigns; the returns were prospective; and the head-count story kept collapsing on inspection. “In many cases the hiring actually goes up when you deploy AI, because you now need new roles. You go from individual contributors to agent orchestrators. The skill sets change, which is also an argument for people not being replaced but needing to be reskilled.”
The business behind the argument is in credit. NewtonX raised a $32 million Series B at the end of 2021, taking total funding to $47 million, reached profitability within two years and is funding growth organically, with around 200 staff, a London office building out its European presence, and a fresh citation from Gartner as a frontrunner among synthetic population and behavioural simulation startups.
Everything Eder sells depends on buyers who ask what the model was trained on, and he is candid that only the sophisticated ones do. The rest of the market is being offered an ever-cheaper way to hear a confident answer from tools trained on the same polluted panel data that created the crisis in the first place. Twenty-five years of fraud sits underneath them, and nothing at this year’s festival suggested the industry is in a mood to slow down and check.
