Stagwell bets its challenger pitch on differential privacy

Almost a year after Stagwell struck its eyebrow-raising deal with Palantir, chief AI officer John Kahan explained the bet to Digiday: rather than strip data to protect privacy, Stagwell's platform adds noise to it via differential privacy, a technique Kahan donated from Microsoft to Harvard's Institute of Quantitative Social Science and then hired Harvard to implement inside the Palantir system. "We treat privacy quite different than our competitors. We believe it's a feature," Kahan said. The platform is deliberately not run by Stagwell: "It's not intended to be run by us. It's actually intended eventually to be run by each of our customers." Kahan framed the scale differently from Epsilon's behavioural-data model: Stagwell's US dataset spans over 1 billion email addresses, 260 million individuals, 160 million households and roughly 60 years of Harris Poll data, which he argued measures how people think, not just what they do. The US Supreme Court, he noted, ruled differential privacy the only data technology it would allow in the last census.

The crux

Stagwell is positioning differential privacy (adding calibrated noise rather than stripping data) as a competitive moat against Epsilon-style behavioural-data rivals, betting that privacy-as-a-feature plus model-agnostic infrastructure wins over brands worried about both compliance and lock-in.

Sell-side read

Differential privacy and clean rooms are becoming the sales infrastructure publishers transact through, but the pitch needs a caveat: noise-injected data is deliberately less precise than raw data, and clean rooms do not confer automatic compliance. Publishers should sell governed collaboration, not absolute accuracy, and keep their own first-party graph as the negotiating asset.

Watch

Whether Stagwell's customer-run platform actually lands with brands, and whether differential privacy becomes a stated requirement in publisher RFPs for data collaboration.