How AI Is Reshaping Consumer Insights and Marketing Research
AI is quietly rewriting how the consumer insights function actually works, from the personas researchers build to the way trends get validated.
On this podcast episode of the Marketing AI SparkCast, host Aby Varma, Founder of Spark Novus and Marketing AI Pulse, sat down with David Jacobson, Global Senior Director of Applied Consumer Science at The Coca-Cola Company, to talk about how AI is changing consumer insights, synthetic research, creative testing and audience segmentation. Jacobson brings 25 years of marketing strategy experience, including roles at Publicis Sapient and PwC, and now applies AI across nearly every stage of the research process inside a large enterprise insights function. What follows is a look at where AI is earning trust in that role, and where human judgment still has the final word.
What Synthetic Personas Are, and Why Marketers Are Testing Them
A synthetic persona is an AI-built stand-in for a human consumer. Jacobson described several forms it can take, from a digital twin that mimics one specific person to a synthetic survey respondent trained on the same demographic, psychographic and behavioral attributes that would go into a traditional persona slide. The difference is that those attributes now train a large language model instead of sitting in a presentation deck.
The appeal is speed and scale. Traditional research requires recruiting real respondents, which takes time, money and effort, especially for a global brand running campaigns across dozens of markets. Jacobson noted that generating tens of thousands of synthetic respondents, split across specific age ranges and attributes, can happen in a fraction of the time it would take to recruit that same population through conventional methods.
Jacobson has found synthetic personas most useful earlier in the development process, such as narrowing a list of 10 campaign ideas down to the top three or four before committing budget to full production. As ideas move closer to final execution, insights teams tend to shift back toward traditional, human-based testing to validate what the AI surfaced.
Digital twins that simulate a specific individual
Synthetic survey respondents trained on persona attributes
Best suited for early-stage idea filtering rather than final validation
How Insights Leaders Decide When to Trust an AI-Generated Insight
Trust is earned, not assumed. Jacobson said AI models in his experience generally reach 80 percent to 90 percent accuracy when checked against extensive historical research on consumer behavior, switching patterns and purchase occasions. That track record gives insights teams a baseline for judging whether new AI output makes sense.
The bigger risk, according to Jacobson, is not inaccuracy but the illusion of depth. AI can produce a polished, hundred-page report that looks thorough and professional without containing a genuinely useful insight. He compared this to a familiar problem in traditional research: a study can validate or fail to validate a hypothesis without ever surfacing the human tension or need the team was actually looking for.
The guardrail Jacobson points to is a simple one: keep a human in the loop. His organization runs every AI project through an internal governance review, not to slow teams down, but to help people learn from similar projects already completed elsewhere and avoid duplicating work. The best practice is to treat AI output the way you would treat a junior researcher's first draft, useful, fast and worth a second look before it shapes a campaign decision.
Why Data Quality Determines Whether Synthetic Personas Work
Ask Jacobson what separates a reliable AI insight from a misleading one, and the answer comes back to data quality. Insights teams like his train models primarily on internal historical data, supplemented by trusted third-party sources such as Nielsen or Circana, rather than allowing a model to draw broadly from the open internet.
That distinction matters more for a global brand than it might for a single-market company. Data quality and availability vary significantly by country, which directly affects how much confidence the team can place in synthetic research for a given market. Jacobson said the team still wants visibility into local trends, political and economic context and cultural shifts that would not show up in internal data alone, so the approach blends trusted proprietary data with select external signals rather than defaulting to public web data.
The practical takeaway for any marketing organization exploring AI-driven research is that the model is only as good as what it was trained on. Enterprise versions of AI tools that connect to a company's own vetted data sets behave very differently than consumer-grade tools pulling from the general internet, and that gap becomes especially visible once a brand is trying to operate at global scale.
Separating a Real Trend From AI Noise
One of the more practical challenges Jacobson raised is knowing whether an AI-surfaced signal represents a genuine trend or a statistical blip. His rule: one data point is not a trend. If the same signal appears multiple times across multiple markets or time periods, it starts to earn credibility.
Importantly, Jacobson said the team does not use AI to validate AI. Instead, an emerging signal gets checked against real humans, whether that means talking to local market experts, running a smaller focus group or comparing notes with external trend trackers such as Kantar Global Monitor or Morning Consult, both of which now incorporate their own AI models into their reporting.
He also flagged a subtler risk. AI models are built to extrapolate. Given a few early data points, a model may confidently project that a small, localized behavior will scale into a broad cultural trend, when in reality it stays contained to a niche group. For a culturally embedded, consumer-facing brand, mistaking a local blip for a global shift carries real cost, which is why human validation stays part of the process rather than a formality.
How AI Is Reshaping Audience Segmentation
Audience segmentation has long been a strength of data analytics and automation, and Jacobson said AI has made a historically labor-intensive process dramatically faster. Fusing multiple data sets, first-party, retail and third-party, used to require specialized data science talent and significant manual effort in spreadsheets. AI models can now identify shared attributes across data sets and surface high-value audience segments far more efficiently.
What has changed most, in Jacobson's view, is what becomes economically viable. Marketers have talked about personalization and audience optimization for more than a decade, but doing it at scale across many segments and creative variations was historically too costly. AI has lowered that cost enough that even a small, high-performing segment, one that would not have justified custom creative in the past, can now get a tailored offer or message built specifically for it.
That shift moves segmentation from an exercise in finding the largest addressable group to one of matching precise audience signals with the right creative, channel and moment, whether that is Meta, TikTok or another platform. The result is personalization at a scale that simply was not cost-effective before AI made it practical.
What’s Changing in Creative Testing
Before AI, creative testing relied on rudimentary stimuli, animatics, storyboards and early-stage concepts that made it hard for respondents to fully grasp the intended message. Jacobson said AI now allows teams to generate more polished creative stimuli earlier, which removes some of that ambiguity from testing.
AI also plays a direct role in evaluating creative. Some research partners have trained AI models on decades of prior human ad testing, and Jacobson said he has seen confidence levels of 90 percent or higher from some of those partners. That performance gives insights teams room to test more creative variations and adaptations across markets without the cost of testing each one with live human panels.
Even so, Jacobson urged caution as creative gets closer to final execution. An AI model's read on how an ad will land on a specific platform, in a specific cultural context, is not the same as understanding how a human audience will actually experience it in that moment. AI testing has also helped compress the overall timeline from insight to market, a process that could once take a year or more, giving marketing teams more confidence that a campaign still reflects a trend that was valid when the research began.
AI Agents Are Automating the Repetitive Work Behind Insights
Agents are the newest layer Jacobson's team is exploring, and the use case is straightforward: any process that repeats on a predictable schedule is a candidate for automation. Weekly and quarterly campaign performance reports, which used to require hours of manual data consolidation, are a clear example of work an agent can now handle so the team can focus on interpreting results instead of assembling them.
The more advanced application involves agents working with each other. Jacobson described a setup where one agent specializes in a first-party data set while another works with a partner's retail media data, with the two agents communicating to identify where the data sets align and what insights emerge from combining them.
That architecture comes with a clear risk. If one agent introduces a bad data point, a hallucination or a false positive, that error can propagate to every agent downstream, compounding the mistake rather than containing it. Data security and privacy are equally non-negotiable in this setup. Jacobson confirmed that data-sharing boundaries, including tools like data clean rooms, get established before any agent-to-agent workflow goes live, ensuring information never moves outside an approved environment.
Ready to Build Your Own AI-Driven Insights Approach
This approach shows what is possible when AI is paired with the right human oversight, data quality standards and validation process. If your team is trying to figure out where synthetic research, audience segmentation or AI agents fit into your own marketing operation, contact Spark Novus to talk through your specific goals and challenges.
FAQs
-
A synthetic persona is an AI model trained on the same demographic, psychographic and behavioral attributes that would define a traditional consumer persona. It can take the form of a digital twin representing one person or a large pool of synthetic survey respondents used to test ideas at scale.
-
Accuracy varies by use case, but enterprise insights leaders report seeing 80 percent to 90 percent accuracy when AI-generated insights are checked against historical human research. Accuracy tends to be highest earlier in the research process and requires more scrutiny closer to final decisions.
-
No. Marketing leaders who use AI at scale still rely on human-based research such as focus groups, surveys and ethnography, particularly for validating late-stage creative and confirming that AI-surfaced trends are genuine rather than statistical noise.
-
Segmentation results are only as reliable as the data behind them. Enterprise AI models trained on vetted internal and third-party data produce more trustworthy segments than models drawing from general public data, especially when a brand operates across multiple countries with varying data availability.
-
AI agents are most commonly used to automate repeatable processes, such as recurring campaign performance reports, and to help fuse multiple data sets together, including first-party, retail and third-party data, that previously required manual data science work.
-
The main risk is error propagation. If one agent introduces a bad data point or a hallucination, that mistake can pass to other agents and compound across a workflow. Data security and privacy boundaries also need to be defined before agents are allowed to share data with each other.
-
AI is compressing the time between research and market launch. Work that once took a year or more from initial insight to a live campaign can now move much faster, giving marketing teams more confidence that a campaign still reflects a trend that was accurate when the research began.
