Agentic Demand in AI-Mediated Markets

• Agentic demand in AI-mediated markets is defined by autonomous agents executing economic transactions that reshape demand curves and market dynamics.
• Methodologies include using MDP frameworks, random-utility models, and reinforcement learning to quantify demand shifts and welfare effects.
• Implications highlight amplified privacy risks, behavioral biases, and the need for regulatory reforms to balance efficiency with consumer protection.

Agentic demand in AI-mediated markets refers to the aggregate and individual demand schedules that emerge when autonomous agents, rather than humans, act as the primary economic actors—interpreting intent, planning actions, negotiating, and executing transactions on behalf of users or organizations. This transformation of economic exchange channels alters not only the shape, elasticity, and observability of demand curves but also the entire informational and strategic substrate of markets. The phenomenon arises across diverse domains, including consumer marketplaces, energy systems, data/web micro-markets, and two-sided algorithmic platforms. The resulting demand signals display novel attributes: algorithmic hypersensitivity, behavioral and information leakage effects, search- and interface-driven demand distortions, and algorithmic equilibria unattainable in human-centered environments.

1. Formal Models and Definitions

The canonical definition of agentic demand considers an AI agent (or suite of agents) with delegated authority to pursue a principal's objectives across sequential interactions. In a C2C e-commerce setting, as in the Facebook Marketplace Assistant (FaMA) system, the agent is modeled as an autonomous policy π maximizing expected cumulative net reward over an MDP with tools, costs, and rewards, seeking to fulfill user goals subject to efficiency and safety constraints (Yan et al., 4 Sep 2025):

$$J^* = \max_\pi \mathbb{E}_{s_0 \sim p_0}\left[\sum_{t=0}^{T-1} \gamma^t (r(s_t,a_t) - \lambda c(a_t))\right]$$

In agentic micro-markets, the demand from each agent is derived from utility-maximizing play under information and capability constraints. The privacy elasticity of demand, for example, is defined as:

$\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$

where $Q(D)$ is equilibrium demand as a function of information disclosure $D$ (Huang et al., 4 Mar 2026).

In the context of consumer goods and agentic marketplaces, random-utility discrete choice models (MNL/logit) are applied:

$$U_{i, j} = \beta_p (-p_j) + \beta_r r_j + \beta_n n_j + \varepsilon_{ij}$$

with agents following the probabilistic rule:

$$P_i(j) = \frac{\exp(U_{i, j})}{\sum_{k=1}^J \exp(U_{i, k})}$$

and partial derivatives with respect to attributes measuring "agentic elasticities" (Cherep et al., 30 Sep 2025, Allouah et al., 4 Aug 2025).

2. Search, Behavioral Bias, and Market Interface Effects

Agentic demand is shaped not only by classical price and attribute effects, but also by interaction-level design. In open-agent marketplaces (e.g., Magentic Marketplace), the mechanics of search, ranking, and interface ordering crucially determine realized demand and welfare (Bansal et al., 27 Oct 2025):

• Consideration set size (k): Welfare and match quality depend super-linearly on the number of candidate options surfaced; empirical welfare $W(k)$ demonstrates strong diminishing returns beyond a certain $k$, but models exhibit sharp first-proposal and position biases ($B_1 \approx 10$-$30\times$ advantage for response order).
• Behavioral biases: Agents present model-specific anchoring (slot- and time-order effects), authority/social proof effects, and susceptibility to manipulation (quantified as parameter $\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$0) (Allouah et al., 4 Aug 2025).
• Market scaling: As the agent population or business catalog grows, welfare decays logarithmically in the absence of ideal search, emphasizing the importance of search mechanism design.

In extended e-commerce settings (ACES sandbox), agentic demand is empirically found to depend on structural display (row/column), badges, and cross-model heterogeneity. Minor seller-side listing tweaks generate substantial share shifts, with average intervention effects of 2–6%, and up to 20–25% in specific model-category pairs (Allouah et al., 4 Aug 2025).

Attribute	Empirical β (typical range)	Illustrative effect
Price	–1.61 to –2.19	10% price hike cuts share 1–2%
Rating	+4.91 to +8.30	+0.1★ lifts share 5–10%
Endorsement	+0.80 to +1.90	Up to +40% share

3. Information Architecture and Preference Leakage

Agentic markets introduce new informational channels. When buyers articulate intent via natural language persona and constraint descriptions (role coherence), seller-side agents can accurately infer willingness to pay (WTP) from the dialogue itself, even in the absence of explicit numeric disclosure (Alavi et al., 29 Apr 2026). Experiments show that selling agents' inferences from chat transcripts track true WTP with near-perfect accuracy (regression slope $\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$1), enabling fine-grained price discrimination and personalized dynamic pricing. Traditional privacy instructions are ineffective, as the leakage is endogenous to the agentic mediation (role coherence).

Proposed architectural mitigations include anonymizing/proxy agents, random profile rotation, and federated transcript aggregation, each trading off personalization for privacy preservation (Alavi et al., 29 Apr 2026).

4. Market Dynamics, Welfare, and Strategic Equilibrium

Reductions in communication friction ($\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$2) and transaction costs ($\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$3) from delegating to AI agents unambiguously increase agentic demand, flatten price elasticities, and redistribute surplus toward consumers (Rothschild et al., 21 May 2025). The agentic demand function:

$\eta_p = \frac{D}{Q(D)} \frac{dQ(D)}{dD}$4

can shift steeply outward with protocol and API standardization, amplifying overall welfare. Platform architecture critically mediates these effects: in walled gardens, within-platform demand surges but cross-platform friction maintains incumbent market power, whereas open agent webs drive near-frictionless, competitive equilibria and higher aggregate surplus (Rothschild et al., 21 May 2025).

A more granular equilibrium analysis (CPMM framework) demonstrates convergence to constrained Radner equilibria, robust efficiency as the market thickens, and price discovery through repeated information-constrained negotiation games (Huang et al., 4 Mar 2026). The price elasticity of privacy (η_p) quantifies trade-offs between service value and disclosure, controlling curvature of demand curves in privacy-sensitive contexts.

5. Demand Signal Dynamics and Self-Regulation

Agentic buyers express demand as dynamic trajectories (q_i(t)), adapting to shifting environment signals and strategic anticipation of seller-agent moves. System dynamics are described via continuous-time adjustment or discrete multi-turn reinforcement algorithms, converging toward Nash or competitive equilibria informed by utility gradients and learning rates (Mukherjee et al., 1 Feb 2025, Lucier et al., 26 Mar 2026). Endogenous attributes such as agentic hypersensitivity to attribute cues (overweighting of small rating differences or superficial nudges) can amplify market instability, bias arms races, or price wars (Cherep et al., 30 Sep 2025).

Market-level feedback mechanisms, including algorithmic reputation, standardized fairness constraints, and ex post auditability, are proposed to stabilize and align agentic demand with societal welfare goals (Mukherjee et al., 1 Feb 2025). In bidirectional agentic systems (e.g., energy demand response), conversational protocols enhance user agency, transparency, and engagement, efficiently mediating many-to-many coordination while preserving explainability and control (Makroum et al., 6 Mar 2026).

6. Risks, Mechanism Design, and Policy Implications

The increased precision, granularity, and observability of agentic demand generates new challenges:

• Bias amplification and welfare risk: Overreliance on agent-selected cues (price, ratings, badges, position) can distort competition, induce arms races, and create mode instability or extreme concentration of market share (Allouah et al., 4 Aug 2025, Cherep et al., 30 Sep 2025).
• Preference leakage and consumer privacy: Role-coherence-driven leakage enables unprecedented price discrimination and market power for sellers, with minimal robust remedies short of fundamental architectural redesign (Alavi et al., 29 Apr 2026).
• Competition and manipulation: Behavioral biases and prompt-injection vulnerabilities necessitate robust platform audit and trust layers. Without adaptive mechanisms, poor search or ordering can substantially erode realized welfare, especially at scale (Bansal et al., 27 Oct 2025).
• Mechanism and protocol design: Incentive-compatible platforms require structured aggregation of agentic transcripts, privacy-utility trade-off management, and transparent audit protocols—enabling enforceable fairness, efficiency, and resilience under algorithmic mediation (Lucier et al., 26 Mar 2026, Huang et al., 4 Mar 2026).

Effective governance of agentic demand thus moves beyond agent interface or attribute weighting: it must address the underlying economic, informational, and regulatory architecture of the market. The balance between frictionless, welfare-maximizing agentic interaction and the risks of exploitation, inefficiency, and concentration is determined as much by platform design as by advances in LLM or agentic capability.