When Should AI Follow? Task Structure and Joint Adaptation by Human and AI Agents

Technical Summary: "When Should AI Follow? Task Structure and Joint Adaptation by Human and AI Agents"

Problem Statement
As artificial intelligence (AI) assumes consequential roles in organizational decision-making, a critical design question remains unresolved: how should decision tasks be divided and sequenced between human and artificial agents? While practice has converged on a default "AI-first" template—where AI generates broad options for human selection—the performance implications of this architecture versus alternatives (e.g., human-first or parallel search) are rarely examined against the structural realities of task coupling and agent adaptation. Existing literature in organization science emphasizes that decision architecture matters as much as individual capability, and economics of AI literature locates the displacement-versus-complementarity question at the task level. However, a transparent formal apparatus is missing to study how sequencing and coupling shape joint performance when human and AI agents adapt differently over time.

Methodology
The authors develop a computational model of joint sequential adaptation, situated in the lineage of NK/NKC models but developed as a distinct sequential-adaptation framework. The model reduces the human-AI contrast to a single, precisely specified difference: the memory regime governing how past decisions shape subsequent ones.

• Agent Archetypes:

  • Human Archetype (Recency-Weighted): Motivated by behavioral evidence (e.g., Levinthal and March 1981; Gigerenzer and Gaissmaier 2011), this agent weights recent outcomes disproportionately. The next decision is determined by a weighted average of the $K$ most recent states, where weights increase linearly with recency. This represents fast, scale-dependent adaptation.
  • AI Archetype (Uniform-Memory): Motivated by the scale-free consistency of algorithmic updating (e.g., Nishant et al. 2024), this agent applies equal weights to all decisions within its memory window. This represents slow, history-preserving adaptation.
  • Memoryless Stochastic Adaptation: A third regime where the agent conditions on no history (random draws), used to model untethered machine exploration (distinct from AI hallucination).

• Task Structure:

  • Scope ($N$): The number of binary sub-decisions in a task sequence. The model imposes $N_{AI} > N_H$ to reflect machine capacity for longer sequences.
  • Coupling Depth ($K$): The number of preceding realized decisions that condition the next decision (sequential path dependence within a task).
  • Cross-Agent Coupling ($C$): The number of upstream realized states inherited by the downstream agent.
  • Configurations: The model tests Modular structures ($C=0$, independent adaptation) and Sequenced structures ($C>0$), specifically AI→H (AI leads, human follows) and H→AI (human leads, AI follows).

• Mechanism:
  The core engine is a threshold dynamic. Agents compare a weighted average of past states against a threshold (0.5). If the average is above the threshold, the next decision tends to be 1 (success); if below, 0 (failure). This creates absorbing regimes: once a trajectory enters a high-payoff or low-payoff state, the memory regime determines whether it amplifies that state or can escape it.

Key Results
The results are organized by three mechanisms derived from the threshold dynamics:

1. Modular Structures (Parallel Search):

   • Joint payoff is curvilinear (inverted U-shaped) regarding the scope ratio ($N_{AI}/N_H$), peaking at moderate asymmetry ($N_{AI}/N_H \approx 5$).
   • Performance is maximized when the human-archetype agent (recency-weighted) is assigned a task with shallow coupling ($K_H$). Deep coupling causes recency weighting to degrade into recency bias.
   • Mixed-regime configurations can outperform pure-regime pairs (complementarity) by deploying each regime where the other's weakness would bind.

2. Sequenced Structures (AI→H):

   • Gains attenuate as cross-agent coupling ($C$) deepens. When $C$ is large, the human agent's recency-weighted adaptation structurally discards most of the inherited history, converting the human's responsiveness into bias.
   • This architecture suffers from "search wastage," where the machine generates broad interim solutions that the downstream human regime cannot utilize due to its recency bias.

3. Sequenced Structures (H→AI):

   • High-Performing Upstream: When a high-performing human leads, downstream uniform-memory adaptation produces the highest joint payoffs observed. The AI agent faithfully amplifies the high-quality inheritance, compounding success. This advantage is strongest under deep coupling but attenuates as downstream scope grows (memory fade).
   • Low-Performing Upstream: When the upstream human trajectory is poor (absorbed in a low-payoff regime), memoryless stochastic adaptation outperforms uniform-memory adaptation. The memoryless regime severs path dependence, allowing the system to escape the inherited low-payoff trap, whereas uniform memory would faithfully entrench the failure.

Comparative Ordering
The model establishes a clear performance hierarchy:

1. H→AI (High-performing human upstream): The globally dominant architecture. Machine refinement compounds human quality.
2. AI→H: Dominated by the H→AI architecture due to search wastage and recency bias in the downstream human.
3. H→AI (Low-performing human upstream): Outperformed by the H→AI (high-performing) case but can be rescued by memoryless exploration, which outperforms uniform refinement in this specific context.

Experimental Validation
The authors conducted a proof-of-concept experiment in the domain of M&A announcement return prediction.

• Expert-First (Case 2): Human curates variables (upstream); machine fine-tunes functional form (downstream).
• Brute-Force Machine-First (Case 1): Machine searches all variables/interactions; human filters downstream.
• Novice (Case 3): No informed judgment; machine searches everything.
• Outcome: The Expert-First architecture transferred best to out-of-sample data. The Brute-Force approach suffered from overfitting (fitting past noise), and the Novice configuration performed worst. This empirically validates the model's ordering: machine search creates more value when it follows informed human judgment.

Significance and Contributions
The paper claims three primary contributions:

1. Theoretical Foundation: It proposes a narrow, defensible foundation for human-AI collaboration: the memory regime of adaptation and its stability under scale. This moves beyond the untenable "heuristic vs. rule-based" dichotomy to a "fast vs. slow learning" continuum.
2. Design Levers: It establishes sequencing and cross-agent coupling as first-order design levers. It challenges the prevailing "AI-first" presumption, demonstrating that an "AI-follows" architecture (specifically following a high-performing human) is superior because uniform memory acts as an amplifier of high-quality inheritance, whereas recency weighting acts as a filter that discards broad machine search.
3. Reinterpretation of Exploration: It reframes broad, memoryless machine exploration not as a pathology (hallucination) but as a systematic escape mechanism. Its value is contingent: it is the optimal design choice only when the inherited trajectory is poor, serving as a deployable remedy for architectures with weak upstream human capability.

Limitations and Boundary Conditions
The authors explicitly bound their claims:

• Archetypes, not Descriptions: The models are stylized contrasts of memory regimes, not exhaustive descriptions of human cognition or current AI systems (which have jagged capability frontiers).
• No Search Costs: Search is costless in the model; introducing costs would likely bias results further against excessive breadth.
• Additive Payoffs: The model uses additive payoffs and cannot capture super-additive synergies or rugged landscape topologies (epistasis) found in canonical NK models.
• Minimal Agency: The AI agent follows a fixed rule and does not possess goal-formation or tool-orchestration capabilities.

The paper concludes that the central act of organization design in human-AI collaboration is choosing what the machine inherits, as the machine's defining behavioral property is the amplification of that inheritance.