Feature
White Paper
The Future of Leadership Development: AI-Driven Manager Enablement
Research on human-AI collaboration shows that when people team with AI agents instead of other humans, they produce more ads, faster, and with higher-quality text.
But the ads they create turn out to be more similar to each other than ones made by all-human teams, a pattern researchers call "diversity collapse." It’s not the AI itself so much as how people work with it: They delegate more, edit less and talk less like collaborators and more like managers handing off tasks.
Does homogenized output become a liability? And could something as simple as varying the AI's persona or voice counteract it?
Four people who work in consulting, higher education, competitive intelligence and product management at one of the world's largest retailers, have different ideas.
The Outsourcing Trap
Homogenized output can become a liability, but it's usually a result of how organizations choose to deploy AI rather than an inevitable side effect of the technology itself, said Claire Brady, president of Glass Half Full Consulting and author of "AI With Intention."
When teams adopt AI at scale, they’re tempted to treat it as an outsourcing engine: Assign a task, accept the first draft and move on. That approach will almost certainly produce more uniform language and ideas, because large language models are built to generate the most probable response based on patterns they've learned, she explained.
On AI personas, Brady is only cautiously optimistic. A "skeptical analyst" or "experienced dean" persona may change wording, but those remain synthetic perspectives standing in for lived experience and institutional judgment rather than replacing them. The real differentiator isn't the persona but the people around it, she said.
Brady recommended that AI should produce roughly the first 70% of a piece of work, while the last 30%, influenced by expertise, relationships and institutional knowledge, stays with people. That creates value.
Variety Comes From Inputs, Not Voices
AI-assisted work grows more uniform whenever people accept the first draft and skip the questioning, editing and discussion that would normally follow, agreed Daniel Burrus, a technology adviser for more than 30 years and High Point University's Artificial Intelligence Expert in Residence. Varying a persona might change tone, but variety comes from using different sources, asking for competing ideas, assigning roles and keeping humans responsible for the final point of view.
Faster output is useful on its own terms, but sameness becomes a business risk once every campaign, product idea and brand voice starts resembling the competition's, Burrus said. The fix requires collaboration, not a handoff followed by acceptance.
The Persona Experiment That Fell Flat
The most direct rebuttal to the "personas help" camp comes from an experiment. Analysts researching the same competitor question, even using different AI models, tended to land on effectively the same answer, noticed James MacAonghus, Aqute Intelligence's vice president of research.
To test whether personas could change that, MacAonghus asked Claude to answer a question about Salesforce's enterprise strategy twice, once in the voice of a methodical, detail-oriented competitive intelligence professional, and once as a fast-moving, instinct-driven product marketing manager. A separate model, ChatGPT, was then asked to compare the two answers blind.
The two responses were largely interchangeable, delivered in slightly different styles but close enough that one could pass as a rewrite of the other.
Personas with different demographics, work history and attitudes, even styled to mimic individual analysts, produced differences that were real but marginal. They fell short of what would be needed to offset the effort of maintaining separate personas and reconciling their outputs.
To get divergent output, a persona has to be pushed so far from the norm that it becomes unrealistic or self-defeating, such as a business writer instructed to love being controversial, an uncommon trait in B2B communications.
Models are influenced by training data reflecting how people in a given sector already tend to write, so industry-level convergence seeps through no matter what persona sits on top of it, MacAonghus believes.
What Worked at Walmart: More Context
The closest thing to a real-world test case among the four contributors comes from Walmart, where staff product manager Richa Taldar leads AI and generative AI products for marketing planning.
While building an AI-powered marketing brief tool, Taldar grounded the system in only three or four historical briefs. The drafts were usable, but the same structures kept showing up. Changing the persona helped with tone, but did not reliably produce a different point of view once the underlying context and process stayed identical, she found.
What worked was when Taldar expanded source material to roughly two years of historical briefs, and the agent was redesigned to ask targeted questions about audience, goals, channels and constraints before drafting anything, with a human reviewer kept in the approval loop and the knowledge base refreshed over time.
AI sameness is frequently a workflow problem that gets blamed on the model, made worse by the fact that polished, fluent output tends to get edited less critically simply because it feels finished, Taldar said.
The Problem Isn't the Model. It's Us
The underlying cause of sameness is behavioral, not architectural, interviewees agreed. People accept first drafts, skip debate and treat AI outputs the way they'd treat a finished product rather than a starting point.
Where they diverge is on how much a persona fixes that. Brady and Burrus treat varied personas as a modest, partial lever, useful for tone but insufficient on their own. Taldar's experience suggests personas without richer context and structured questioning barely move the needle.
Engaged, But Not Convinced: What's Eroding Beneath Your Strongest Engagement Numbers
Watch Now
Retention Is a Communication Problem: New Data from 6,000+ Employees
Watch Now
MacAonghus's data goes further still, showing that even deliberately engineered persona differences became nearly identical.
That leaves organizations racing to scale AI teamwork with more work than a prompt-engineering trick.
Varying an AI's voice might soften the symptoms. But sameness gets solved by changing how much thinking a team is willing to keep doing itself. Will organizations focusing on speed slow down enough to find out?
The stakes are plain: If every campaign starts from the same structure, language and assumptions, personalization becomes cosmetic, and customers may recognize the machine before they recognize the brand.
Editor's Note: What else happens when people accept AI outputs as final products?
- When AI Writes, Humans Disconnect — As AI polishes our messages, something human gets lost.
- AI Makes Us Sound Better and Communicate Worse — AI cleaned up our workplace communications. It's also made people less likely to have difficult conversations, with real organizational consequences.
- AI Hiring Bias Has an Accomplice: You — If AI has an opinion, people trust it. We need new strategies to stop humans from signing off on AI bias.