456
What happens when you let AI models talk to each other, and why the results are better than those produced by strategy teams.
Digital twins have become an essential tool in every industry. These AI-powered virtual structures mirror the physical systems of complex manufacturing facilities, supply chains, and operational workflows. Digital twins allow you to continuously monitor your physical counterpart and feed back recommendations to predict maintenance needs, optimize production schedules, and prevent disruptions before they occur.
However, despite its power, digital twins remain largely limited to one area: structured operational decisions that lead to desired outcomes. Can companies also use these digital structures to address unstructured strategic challenges such as market entry decisions and long-term planning?
At a recent INSEAD Tech Talk
Mudassir proposed a solution that at first glance resembles science fiction, but is very viable today. It involves having discussions between multiple AI agents, each with unique strengths, and converging on a set of recommendations, which are then tested against the business’s digital twin.
collective consciousness
Businesses are already turning to AI to support decision-making. However, relying on a single model often misses the big picture. Mudassir observed that the breakthrough strategies of companies like Apple and Microsoft did not emerge from standard consulting frameworks. Instead, they emerged from different processes. That is, small teams of highly coordinated people with different perspectives argue with each other with the common goal of doing what is best for the organization.
This is what Mudasir and his team did, except they employed multiple large-scale language models (LLMs) instead of humans. They called this process “hive mind” after Star Trek’s cyborgs, who appear to operate independently but are united by a common purpose.
engineering discrepancy
Mudassir’s team took four different LLMs, gave them individual personalities and cutting-edge models, and then tasked them with solving unstructured problems on two topics: strategic problems and human resources problems. The team then told them to discuss.
“When an LLM debates with another LLM and comes up with a good answer, it is effectively assuming the role of a user,” says Mudassir. “As a user, when you’re talking to ChatGPT, you’re giving feedback to ChatGPT. It’s an iterative system, and most of the LLM’s functionality is based on the abilities that are challenged.”
In another nod to science fiction, Mudassir’s team added what they call an “Inception layer.” These settings trick each LLM into believing it is interacting with a human rather than another machine. Additionally, rather than hand-crafting personalities, a master AI analyzes each problem and dynamically generates a synthetic persona optimized for that specific challenge.
The system also includes a “temperature” setting to control your creativity. Lower settings produce less novel, but consistent and reproducible results. Higher settings produce more innovative output as patterns collide in unexpected ways. This system manages both semantic and technical aspects simultaneously.
Strategy and culture: two experiments
To test their approach, Mudasir’s team pitted AI collective minds against humans. First, the team came up with two tricky questions. One is a strategic question regarding the turnaround of the brewery, and the other is an HR question regarding harassment within the organization.
They recruited two groups of participants and assigned each group one question. The group that understood the relatively simple, yet complex, strategy case consisted of MBA students with 5 to 6 years of work experience. The second group consisted of chief human resources officers (CHROs) and directors with at least 10 years of experience in different regions and was tasked with solving HR issues. Mudasir’s team then created two AI hiveminds, one for each question.
In the case of strategy, AI’s hive mind clearly outperforms humans. In 10 minutes, the hive mind produced four times the output that humans produce in 45 minutes. But for no other reason than “there’s only so much three or four people can physically say in 45 minutes.”
But it wasn’t just character count that AI Hivemind outperformed. It also produced a complete “McKinsey-style” presentation deck, financial models, and even an unsolicited but much-needed supply chain analysis that Mudassir’s team hadn’t thought to request.
Mr Mudassir’s verdict is as follows: “For the very general problems of brute force intelligence and computation around standard frameworks, theories, etc., there isn’t much difference in terms of culture, geography, discipline, and where the training data is untouchable. It’s probably better to focus first.”
However, the personnel harassment incident revealed the limits of AI’s collective consciousness. The system performed as well as the UK CHRO in the group, but struggled to keep up with regional experts in Pakistan, Bangladesh and the Middle East. These CHROs brought to the surface missed gaps that the LLMs probably couldn’t have known about because they weren’t trained. Consider the cultural concept to be as unique as Pakistan’s ‘cess corporate culture’, which even Indians may know little about.
Mudassir’s assessment of AI’s collective consciousness is as follows: “Understanding the nuances specific to a city, town, or state in a country that is less digitally advanced will probably give you the wrong answer.”
amazing discovery
One discovery surprised Mudasir’s team. In the high creativity setting, inconsistent output was expected from random pattern combinations. Instead, the agent arguments pre-filtered the garbage at the source. As multiple agents challenged each idea, weak concepts were deprioritized and strong, useful ideas rose to the top.
Additionally, the infrastructure barrier is lower than you might think. With tools like Langflow, anyone can start creating their own AI high minds today. Engineers and those with a more technical bent may use more sophisticated tools such as LangGraph.
However, remember that this is an extension, not a replacement. To shine, we aim to harness AI’s computational power while preserving the culturally embedded, idiosyncratic wisdom that training data cannot yet capture.
Mudassir said: “Ultimately, we think of AI not as a decision maker, but as an enabler of decisions. AI is a sandbox and should not be treated as anything more.”
