When Should Organizations Let AI Lead? Study Challenges "AI-First" Thinking

A new computational framework challenges the assumption that artificial agents should always take the lead in human-AI collaboration. Instead, the best approach to splitting decisions between humans and machines depends heavily on the structure of the work itself.

The Research Model#

Researchers developed a model examining how two types of decision-makers adapt over time when making sequential choices. The key difference between them: how they remember and learn from past results. Human agents, drawing on behavioral research, tend to weight recent outcomes more heavily—a "recency bias" that shapes their choices. Algorithmic agents, by contrast, treat a window of past outcomes equally, providing consistent, scale-free updates across time.

The framework builds on classical organizational design concepts (NK/NKC models) and tests various configurations of task scope, task complexity, and coupling between agents. By varying these parameters, the researchers could observe how different memory systems perform across modular versus sequenced task structures.

When Adaptation Compounds—For Better or Worse#

Three key mechanisms emerged from the analysis:

• Threshold dynamics: Initial performance creates absorbing regimes that trap agents in either high- or low-payoff states, with later adaptation reinforcing whatever foundation was laid earlier.

• Memory regime effects: Uniform memory (the algorithmic approach) amplifies inherited trajectories—winning repeatedly if starting strong, but compounding failure if starting weak. Recency weighting corrects locally but becomes volatile at larger scales.

• Stochastic escape routes: Random, memoryless adaptation performs worse on average but serves as a rescue mechanism when agents inherit poor starting conditions.

Beyond "AI First"#

The findings overturn the conventional wisdom that AI should always move first. Maximum joint performance emerges when a high-performing human takes the initial steps, followed by an algorithm that can amplify and lock in those gains. When humans perform poorly at the outset, however, broad stochastic search rescues the sequence—randomness becomes an advantage.

This task-structural contingency logic offers organizations a more nuanced decision framework than universal AI-first deployment. The implication: the optimal division of labor between humans and machines cannot be prescribed in advance but must be tailored to what each task demands and how well each agent is likely to perform it.