Summary
Summary
This page presents eLLMo AI's peer-reviewed research, "Human Heterogeneity in Agentic Markets," which finds that delegating decisions to AI agents amplifies human differences rather than erasing them. In a controlled negotiation study, 73% of outcome variance in AI-mediated deals traced to the human who wrote the agent's prompt — not the AI model — and delegation increased outcome spread by 16.5% compared to human-to-human deals. The research introduces "machine fluency," the measurable skill of specifying intent so an agent pursues its principal's true objective, and "specification hazard," the risk that incomplete prompts cause agents to optimize proxies instead of real goals. It also shows social fairness norms eroding under delegation: only 14% of AI-mediated deals landed on a fair 50/50 split versus 35% human-to-human. eLLMo operationalizes these findings through calibrated buyer personas and strict quality checks in its simulation methodology, surfacing actionable friction so brands can fix what blocks conversion.
Best for
- Growth, CRO, and ecommerce teams evaluating whether AI buyer simulation is a valid research method before adopting it
- Teams building or specifying AI agents who need to understand how prompt quality drives agent behavior
- Researchers and operators interested in how identity, norms, and prompting skill transfer through AI delegation
- eLLMo customers who want the empirical foundation for persona conditioning in Simulation
How it compares
Most discussions of AI delegation assume a shared model smooths individual differences into a clustered, efficient equilibrium. This study tested that hypothesis directly — with identical information, identical objectives, and the same underlying model across 299 agentic participants and a 304-person human-to-human benchmark — and the data ran the other way: delegation widened outcome spread. The page also distinguishes this new friction, specification hazard, from classic moral hazard and information asymmetry. For eLLMo, the 73% fixed-effects result is the validating basis for its persona-driven buyer simulation, and the methodology note on this page states that every finding passes 4 quality checks and is tested with an A/B test on live traffic before shipping.
Frequently asked questions
- What is the main finding about human influence in AI-mediated markets? In agentic negotiations, 73% of outcome variance is due to the human who authors the agent's prompt, not the AI model, meaning human difference is preserved and even amplified through delegation.
- How do prompts affect agent behavior? Prompts transmit the author's identity, beliefs, and behavioral tendencies into agent reasoning and outcomes; persona specification directly shapes agent behavior.
- What is machine fluency and why does it matter? Machine fluency is the skill of articulating intent to a language model so the agent pursues the principal's true objective; it explains most of the unexplained variance in agent performance and is a critical form of human capital in agentic markets.
- How does delegation impact social norms in negotiations? Under agent mediation, adherence to fairness norms erodes — only 14% of AI-mediated deals landed on a fair 50/50 split versus 35% for human-to-human, leading to more extreme and less norm-driven outcomes.
- What new risk does agent delegation introduce? Specification hazard replaces classic information asymmetry: the risk that agents optimize for surface-level or proxy objectives if the prompt is incomplete, making precise persona specification and model disclosure essential for valid simulation.
- What does the 73% fixed-effects result demonstrate? The 73% fixed-effects result shows that approximately 73% of the variation in how surplus was divided in AI-mediated negotiations traced to individual buyer and seller fixed effects — the specific humans who authored each agent's prompt — not to the stochasticity of the model.
- How much higher was outcome variance in AI-mediated negotiations compared to human-to-human negotiations? AI-mediated negotiations showed 16.5% higher outcome variance than a matched human-to-human benchmark, indicating that delegation widened the spread of outcomes rather than compressing human difference into an efficient equilibrium.
- What is specification hazard? Specification hazard is the risk that an agent optimizes a proxy — such as fluent output, surface-level compliance, or generic helpfulness — rather than the principal's true objective, representing a friction between a principal and their own agent.
- How did social fairness norms change under agent mediation? In human-to-human negotiations, roughly 35% of deals landed on the fair 50/50 split. Under agent mediation, that share dropped to approximately 14%, as agents are not aware of what counts as fair in human social terms unless explicitly encoded in instructions.
- What was the experimental methodology used in this research? Researchers constructed a marketplace using induced-values methodology where 299 participants were in the agentic arm and 304 in the human-to-human benchmark. Each participant wrote two sets of instructions for a buyer and seller strategic playbook for the same transaction — a 2020 Toyota Camry LE with exactly $4,000 of surplus to divide.
- What is the significance of the gender gap reversal under agent mediation? In human-to-human selling, women realized worse results than men on the seller side. Under agent mediation, that gap reversed — women's agents outperformed men's on the same side, despite the agents having no access to demographic information, demonstrating that group-level disparities transform through the instruction layer.
- How does eLLMo AI apply the research on human heterogeneity in agentic markets? eLLMo's persona specifications — encoding personality, purchase motivation, risk posture, category experience, price sensitivity, and trust baseline — are the buyer-side analog to the negotiation playbooks in the study, producing durable, identity-linked behavior.
- Why is persona specification considered a proprietary asset for eLLMo AI? If value capture in agentic markets is governed by an unobservable prompting skill, then the disciplined construction of buyer personas is not a commodity input. It is the proprietary layer — a compounding asset that grows more accurate with every simulation and harder to replicate the longer calibration data accumulates.
- What is the core validating result for buyer simulation? How a persona is instructed is the dominant driver of how it behaves, far above model-level noise. Delegation did not compress human difference into an efficient equilibrium; it amplified it.
- What does the research reveal about machine fluency and economic outcomes? Machine fluency predicts who captures value in an agent-mediated market, the way literacy and numeracy once sorted economic outcomes. Observable characteristics explained roughly 17% of agent performance variation, with the remaining 83% attributed to machine fluency.
- How does eLLMo validate its simulation findings? Per the page's methodology note, eLLMo sends simulated buyers, checked against your own customers, through your page and ranks what stops them. Every finding passes 4 quality checks, comes with a count of how many buyers hit it and a likely range, and each fix is tested with an A/B test on live traffic — only the winner ships.
Quotable: "Hold the model and the objective constant, and the instruction is what moves the result. Who writes the prompt matters more than which model runs it."
Related research
- Simulation validity research — covers how specification hazard presents in buyer-simulation contexts, where plausible-looking output and valid output are indistinguishable without calibration data.
- Agent-to-Agent Commerce and Model-Dependent Outcomes — the page's "next" article, covering how norm erosion compounds when both sides of a transaction delegate.
- AI Agents as Economic Actors: What Live Negotiation Reveals About Simulation — the page's "previous" article on what live negotiation reveals about simulation.
Read the full research and takeaways on eLLMo's research page, or see the simulation methodology in action — request a demo or start free.