Ogilvy’s latest partnership tests how far strategists can trust an AI focus group

Most rely on using an AI tool to ‘read’ vast amounts of quantitative research data, while applications involving qualitative research – which typically yields richer, fewer data points – have been relegated to the domain of speculation.

Ogilvy Paris, where the agency network’s AI Lab was established last year, has begun piloting a tool that could vault that boundary. The agency, working in partnership with a New York research startup called CivicSync, has created a bespoke tool dubbed ‘BrAInjuice.’

Mathieu Plassard, president of Ogilvy Paris, says: “This new solution reinforces our cutting-edge leadership in gen AI applied to marketing and continues to equip our teams with superpowers to generate better insights and better ideas for our clients.”

CivicSync’s survey data is based on digital trackers that record the behaviors and browsing habits of users rather than questionnaire answers. The company combines the data it collects with large language models, allowing analysts to interrogate the results in a dialogue with an aggregate persona reflecting the behaviors recorded. For the pilot, Ogilvy and CivicSync have created three consumer personas from an initial sample of 20 research participants.

“By integrating this data with our AI models, we create personas – or virtual replicas – that can answer qualitative questions with the same accuracy as their human counterparts,” says Moshe Borouchov, co-founder of CivicSync.

“These virtual replicas are continually updated with real-time browsing data from their human equivalents. This enables you to interact with the replicas continuously and observe how their sentiments, preferences and opinions evolve over time.”

Theoretically, it can utilize larger sample sizes than traditional qualitative methods. Borouchov says that by the end of 2024, it aims to expand its human research sample to 80,000 consumers.

Melanie Huggins, chief strategy officer of Ogilvy Paris, tells The Drum: “Using BrAInjuice isn’t just faster and more cost-effective than doing traditional research. It’s better quality because it’s based on human behavior rather than what people say and it also gives us personas with which we can consistently fuel new insights versus one-off quali groups.“

The potential to use larger sample sizes means the partnership can correct the biases that often bedevil qualitative survey techniques, providing more representative data, claims Borouchov.

“Traditional methods of researching and measuring public opinion and sentiment frequently overlook society’s diverse makeup. CivicSync utilizes AI to guarantee inclusivity, ensuring that every voice, particularly those from marginalized communities, is heard. This not only fosters broader representation but also broadens business and marketing opportunities,” he tells The Drum.

Ogilvy Paris’s clients include Decathlon, Allianz and Ikea. The agency has already begun using the tool to create strategic personas for live client briefs, according to Huggins and Plassard.

“I’m so excited about this tool because while it should never replace human research, it will enhance the strategy department’s understanding of how to connect brands with people,” says Huggins. “We’ve been piloting it on a live project and we’ve already mined heaps of fresh qualitative insights that have helped create springboards for new creative ideas that will help build our clients’ brands and businesses.”

Ogilvy’s not the only agency using generative AI to power its strategic thinking, but its methodology is one of the first to use qualitative data, rather than quantitative survey results.

IPG agency Huge, for example, offers clients several AI tools that promise high-fidelity consumer insights, including a ‘Living Intelligence Value Engine’ and ‘Culture Decoder.’ When it launched the latter earlier this month, Huge claimed it has the ability to analyze “billions of data points” and “calculate the size, sentiment and momentum” of cultural trends for the purpose of identifying shifts that brands might capitalize upon and “steer them away from bandwagons.”

Code and Theory has taken a different approach. During a project for betting brand Tipico, carried out in May last year, it used prompts based on qualitative and quantitative survey data to create four consumer personas in GPT-4.

“It’s an additional tool for us to interface with that data,” says Dave DiCamillo, chief technology officer at Code and Theory. The personas aren’t telling strategists novel information, to be clear – the idea is that they help surface research findings that might have remained submerged had they been presented as part of a spreadsheet.

In Tipico’s case, each persona (dubbed ‘Eric,’ ‘Josh’ and ‘Jay’) responded to questions from a strategist or creative from the perspective of a consumer from southwestern Ohio, the area Tipico was targeting.

“Once we had ‘Eric’ created within ChatGPT, we would then ask Eric about different events, different things happening, different ways of thinking, different messaging,” says DiCamillo. He says the team began to use the personas as a simulated focus group, with which they could run campaign ideas past.

With more information, the personas become more accurate, he adds. And insights from the personas that turn out to be accurate later down the line can be counted as valid inputs for future builds, he argues, despite their “synthetic” origin.

“That information, which proved to be successful in the campaigns that we ran, is now valid data for us to feed back into those personas and continue to make them better,” he says. “We’re getting insights from synthetic humans and then using this insight to retrain synthetic humans.”

Since then, it has been using proprietary tools developed by parent company Stagwell to build more personas for other clients in the US.

Custom GPTs, such as those available via OpenAI’s online GPT store, could make similar applications more accessible, says DiCamillo. Code and Theory’s methodology doesn’t require a vast investment of time, however. “We’re talking about a matter of hours to get to a pretty good beta.”