Blab AI
Agentic Workforce Grows, but Diversity Gaps Threaten Performance and Market Dynamics
AI agents—systems that grasp context, make decisions, and carry out complex tasks—are steadily weaving into corporate workflows. Developers lean on coding agents to write, test, and review code; call centers deploy agents to triage calls and suggest solutions; supply‑chain planners turn to agent teams to monitor demand, draft plans, and coordinate between customers and suppliers.
The scale of this trend is evident from recent industry observations. A McKinsey & Company interview noted that the company’s workforce has grown to 60,000, with 20,000 of those roles filled by AI agents—an increase from just 3,000 agents a year and a half ago. NVIDIA’s CEO envisions a future in which the firm staffs 50,000 people alongside 100 million AI assistants spanning all business units.
Yet the gains from agentic AI—boosting productivity, sparking innovation, fueling creativity—are curtailed when the workforce is not diverse. The source cites studies that show agent teams selected with diversity in mind outperform individual agents by 25 % on software‑engineering tasks. Another study found that two diverse agents can match or exceed the performance of 16 homogeneous agents, underscoring how varied perspectives reduce correlated errors and accelerate problem‑solving.
The root of the diversity problem lies in the underlying technology stack. Most enterprises rely on a small set of foundation models, retrieval architectures, and data sources. The source reports that when clients ask about diversity, they often mean different personalities or cultural settings for the agents, but the underlying models remain the same. Because the stack is uniform, cosmetic changes to agent personas do not alter the agents’ reasoning or knowledge base.
Research also shows that prompting for personality types produces binary outcomes—agents are either very outgoing or very introverted—while most humans exhibit these traits on a continuum. A study by Atari and colleagues found that major large‑language models respond to psychological profiling tests in a way that mirrors people from Western, Educated, Industrialized, Rich and Democratic (WEIRD) societies, and fail to capture the diversity of other populations with different values.
The lack of diversity in foundation models has organizational, market, and systemic consequences:
1. Correlated errors – If the entire sector uses the same models, fraud‑detection systems in payments or insurance may miss the same false negatives simultaneously, creating a systemic risk. 2. Market convergence – Retailers that rely on identical recommender or pricing systems may converge on the same equilibrium, compressing competitive differentiation. 3. Loss of edge‑case insight – In insurance, convergent models may fail to spot novel fraud patterns. In consumer markets, firms may miss emerging preferences, limiting experimentation and new customer segments.
To mitigate these risks, the source outlines seven imperatives for enterprises:
- Diversify the tech stack – Use different foundation models (e.g., Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini, Meta’s Llama, Mistral’s open models) for reasoning, generation, and evaluation. Diversifying the retrieval layer, orchestration framework, and guard‑rails layer also reduces correlated failures. - Enrich training data – Incorporate multi‑dimensional psychometric datasets (e.g., Big‑Five personality traits) and cultural surveys (e.g., World Values Survey) to train agents that reflect a broader range of human traits. - Fine‑tune with internal data – Leverage enterprise HR systems, employee surveys, and psychometric evaluations to adapt agents to the composition of the workforce. - Train agents by shadowing humans – Use email communications and meeting transcripts to teach agents teamwork styles from diverse geographic and cultural contexts. - Implement model‑portfolio governance – Boards should set limits on the proportion of critical decisions that depend on a single model vendor, treating concentration risk like any other supplier risk. - Use cultural red‑team testing – Expand existing red‑team practices to evaluate agents for bias, societal impact, and cultural sensitivity. - Create agentic talent marketplaces – Develop platforms that allow companies to recruit agents and teams with varied roles, nationalities, skills, personality types, and cultural backgrounds.
Large enterprises are still in the early phase of agentic AI implementation, focusing on pilot use cases and basic AI training for staff. By adopting the above practices now, firms can avoid future problems and strengthen the diversity of emerging agentic teams.
In summary, the agentic workforce is expanding rapidly, but its effectiveness depends on genuine diversity in the underlying models and data. Enterprises that diversify their tech stack, enrich training data, and govern model portfolios are better positioned to reap productivity gains, reduce systemic risk, and maintain competitive differentiation. The next wave of AI adoption will likely hinge on how well organizations address these diversity gaps.
The scale of this trend is evident from recent industry observations. A McKinsey & Company interview noted that the company’s workforce has grown to 60,000, with 20,000 of those roles filled by AI agents—an increase from just 3,000 agents a year and a half ago. NVIDIA’s CEO envisions a future in which the firm staffs 50,000 people alongside 100 million AI assistants spanning all business units.
Yet the gains from agentic AI—boosting productivity, sparking innovation, fueling creativity—are curtailed when the workforce is not diverse. The source cites studies that show agent teams selected with diversity in mind outperform individual agents by 25 % on software‑engineering tasks. Another study found that two diverse agents can match or exceed the performance of 16 homogeneous agents, underscoring how varied perspectives reduce correlated errors and accelerate problem‑solving.
The root of the diversity problem lies in the underlying technology stack. Most enterprises rely on a small set of foundation models, retrieval architectures, and data sources. The source reports that when clients ask about diversity, they often mean different personalities or cultural settings for the agents, but the underlying models remain the same. Because the stack is uniform, cosmetic changes to agent personas do not alter the agents’ reasoning or knowledge base.
Research also shows that prompting for personality types produces binary outcomes—agents are either very outgoing or very introverted—while most humans exhibit these traits on a continuum. A study by Atari and colleagues found that major large‑language models respond to psychological profiling tests in a way that mirrors people from Western, Educated, Industrialized, Rich and Democratic (WEIRD) societies, and fail to capture the diversity of other populations with different values.
The lack of diversity in foundation models has organizational, market, and systemic consequences:
1. Correlated errors – If the entire sector uses the same models, fraud‑detection systems in payments or insurance may miss the same false negatives simultaneously, creating a systemic risk. 2. Market convergence – Retailers that rely on identical recommender or pricing systems may converge on the same equilibrium, compressing competitive differentiation. 3. Loss of edge‑case insight – In insurance, convergent models may fail to spot novel fraud patterns. In consumer markets, firms may miss emerging preferences, limiting experimentation and new customer segments.
To mitigate these risks, the source outlines seven imperatives for enterprises:
- Diversify the tech stack – Use different foundation models (e.g., Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini, Meta’s Llama, Mistral’s open models) for reasoning, generation, and evaluation. Diversifying the retrieval layer, orchestration framework, and guard‑rails layer also reduces correlated failures. - Enrich training data – Incorporate multi‑dimensional psychometric datasets (e.g., Big‑Five personality traits) and cultural surveys (e.g., World Values Survey) to train agents that reflect a broader range of human traits. - Fine‑tune with internal data – Leverage enterprise HR systems, employee surveys, and psychometric evaluations to adapt agents to the composition of the workforce. - Train agents by shadowing humans – Use email communications and meeting transcripts to teach agents teamwork styles from diverse geographic and cultural contexts. - Implement model‑portfolio governance – Boards should set limits on the proportion of critical decisions that depend on a single model vendor, treating concentration risk like any other supplier risk. - Use cultural red‑team testing – Expand existing red‑team practices to evaluate agents for bias, societal impact, and cultural sensitivity. - Create agentic talent marketplaces – Develop platforms that allow companies to recruit agents and teams with varied roles, nationalities, skills, personality types, and cultural backgrounds.
Large enterprises are still in the early phase of agentic AI implementation, focusing on pilot use cases and basic AI training for staff. By adopting the above practices now, firms can avoid future problems and strengthen the diversity of emerging agentic teams.
In summary, the agentic workforce is expanding rapidly, but its effectiveness depends on genuine diversity in the underlying models and data. Enterprises that diversify their tech stack, enrich training data, and govern model portfolios are better positioned to reap productivity gains, reduce systemic risk, and maintain competitive differentiation. The next wave of AI adoption will likely hinge on how well organizations address these diversity gaps.
