Can AI fight corporate groupthink?

Imagine a workplace where every voice gets heard, brilliant ideas emerge from unexpected corners, and decisions stem from collective wisdom rather than boardroom politics.

Kim Polese believes this dream can become reality, and as chairwoman of CrowdSmart, she's using AI to cut through organizational complexity and expose cognitive blind spots that slow down decision-making. 

"AI can do what is impossible for humans to do, which is listen to many people, hear what they say, and then balance and weigh that input," she says. 

Polese isn't just another tech optimist bubbling over about the potential of AI. She has a long history of spotting tech's next big turn.

Thirty years ago, she led the launch of Java at Sun Microsystems; its "write once, run anywhere" philosophy revolutionized software development. She later founded Marimba, pioneering internet-based application management, and started SpikeSource to automate open-source management. Today, CrowdSmart works with organizations like NATO, insurance giant Alera Group, and the 90-10 Institute, a biotech industry think tank using AI to surface insights buried in organizational hierarchies.

“People-first AI dissolves organizational friction by amplifying human wisdom,” says Ashwin Ballal, CIO at Freshworks. “It surfaces insights hidden in hierarchies, enabling leaders to act faster, smarter, and with greater impact.”

While many organizations get bogged down in complex AI implementations, CrowdSmart proves that AI-assisted general collective intelligence (GCI) can amplify human wisdom at scale. Says Polese: "When every person's insights can be heard, evaluated fairly, and combined with others to create something greater than the sum of its parts, organizations don't just improve incrementally. They transform."

What inspired you to apply AI to a collective intelligence platform that helps organizations cut through complexity and deliver sharper, more accurate decisions?

It began with an attempt to address the limitations of the venture capital investment process. The traditional VC decision-making path has always been very relationship-based and limited to a small group of decision makers. But this arrangement has historically poor hit rates—especially at early stages where most companies never reach Series A.

We wondered whether collective intelligence and AI could improve the accuracy of due diligence at scale. So we built a platform with 2,500 experts—professional investors, angel investors, and domain specialists—who evaluate companies sourced from universities and accelerators such as Techstars, Berkeley, UCLA, and the University of Michigan. Each company gets reviewed by 30-45 people over a few weeks.

It sounds like you discovered that complexity was impeding efficiency in traditional VC processes. How does AI change that dynamic?

It starts by gathering data through open-ended questions in the application: Do you think this company has a high likelihood of a profitable exit? If so, why? If not, why not? Then, drill down on domain-specific issues: Is this technology unique? Is the IP protected? Is the go-to-market strategy good? Then it drills down on product-market fit, team, traction, and other factors. 

The AI utilizes large language models to learn semantic meaning from responses, develop themes, and optimize for diversity of thought. After people share their thoughts, the system prompts them to rank the ideas of others, seven ideas at a time. The seven you get are different from what others see, based on what might resonate with you and what might challenge you. The system then generates a predictive computational model of the collective mind. An investment decision is based on all these inputs.

How did that work out?

Over the course of two years, this approach proved to be over 80% accurate in predicting which companies would go on to raise a follow-on round, a very significant predictor of an ultimately profitable exit. 

What was particularly surprising was that more than 40% of the high-scoring companies ended up being founded by underrepresented founders—women and underrepresented minorities. This wasn't because evaluators couldn't tell they were underrepresented founders (there were live Q&As and Zoom calls), but because instead of focusing on "who do you know, what's your network, where did you go to school," the focus was on product-market fit, team, and traction.

What did you learn from this experience? 

That two key elements drive the best outcomes. First, you need participants who are willing to say what they think. In traditional boardroom or leadership situations, you get "I'll do your deal if you do my deal" dynamics. But in our system, identities are masked, so people really say what they think.

Additionally, you need a diversity of perspectives. For example, we were examining a company called Cocoon Cam, based at UC Berkeley. It was essentially sensors on cribs that measured infant heart rate and breathing to protect against SIDS and monitor baby health. Among the people doing the evaluation, we had experts in AI vision technology, professional investors in the medical device space, NICU nurses, parents, and people who understand how to take a product like this into retail at scale through Walmart.

Read also: How long should your AI transformation take?

How does your system actively combat groupthink and reduce the hidden complexity that slows organizations down?

Groupthink happens for several reasons. Sometimes it's going along with whoever has the loudest voice in the room. Or when it looks like the conversation is going in a specific direction and people fold in and go along. That doesn't happen here because there's no benefit to siding with one view or another—you're objectively responding to and ranking other people's ideas. The incentives and motives for going along with the flow don't exist.

The AI system is an objective analyzer that resists groupthink by optimizing for diversity of thought. When you're presented with lists of other people's input, that list is selected both to challenge you and to resonate with you. The system is constantly prompting you to respond to what other people think and to disagree when you do.  

How do you handle conflicting opinions so that they still translate into actionable insights?

We're actually looking for conflicting opinions. That's where the insights lie. The system always asks not just what you think, but why you feel that way. That's where you discover hidden knowledge. That's also what changes people's minds and evolves their positions.

When someone's mind changes, it means they've learned something new. The system actively optimizes for diversity of thought by presenting you with input that both challenges and resonates with you. When you disagree with someone, you explain why, and that often leads to new insights and changed minds.

What's the optimal group size you engage?

The minimum is about a dozen people, and there's no upper limit because AI scales—the more people you have participating the better, because you get greater diversity of perspective. 

What are the key challenges in your approach?

Getting people to participate is always the key challenge. In a company setting, it may be part of your job to perform this task. You're obligated. But it really helps if people care about the question being asked. We use text-based responses rather than voice to reduce bias; some people are more articulate than others, and accents or speaking ability shouldn't influence the outcome. The most crucial factor is whether people care about the problem, because then they'll come back again and again to re-rank and respond to new ideas as they emerge.

How do teams respond when AI highlights their blind spots? 

It's typically very constructive, as you're hearing from people within your organization, including those whose voices may not be heard. What is done with these insights depends on the leader. Many leaders embrace this new knowledge.

Let's take a look at our customer NATO. They've got 32 member nations, a lot of people deep in the organization, many at lower levels, who have insights that often don't get surfaced in NATO's command-and-control environment. NATO uses GCI for collaborative decision-making, running wargaming experiments with human and AI agents collectively reasoning, and countering opponent battlefield innovations in ongoing combat. They see GCI as a way to gain an advantage in a high-risk, complex, and fast-changing environment. They've got to find and surface insights that exist within the organization, and they want to get better at that. 

It's one thing to acknowledge the bias or the gaps. It's another thing to change and integrate different ways of doing it. How do you help effect change? 

The organization decides what to do with the information we surface. Sometimes they conduct another round of engagement to gain deeper insights into how to address a specific issue. Now that we've learned about all the problems, what should we do? What are the solutions? Remember, it's not the AI that is deciding on any of this. The AI helps with learning and listening. The stakeholder has the answers, or at least is working toward them. AI is just a tool. It surfaces knowledge that’s already in your organization but buried.

What’s your best advice for IT and customer service leaders who want to implement collective intelligence tools to create faster, smarter service?

First, assemble a diverse group of participants. AI can handle unlimited group sizes, whereas traditional facilitation caps out at seven. Use that option.  

Then, get laser-focused. Define specific problems and concrete outcomes. Instead of vague goals, ask targeted questions, like: What's your confidence level we can boost customer recommendation rates 20% in 90 days? Follow that up with: What specific actions drive your confidence or concerns? What are the three most critical moves?

Think of the AI as a super facilitator, understanding each person's ideas in order to get a sense of overall group sentiment. Once you have these insights, you can have the conversations that lead to better organizations and more accurate decisions.