Measuring Economic Efficiency in AI-Mediated Markets: Transaction Costs, Rent Extraction, and Welfare

Abstract

We introduce Torque, a multiplicative efficiency model for autonomous agent ecosystems that separately quantifies value creation efficiency, allocative distance from the Pareto frontier ^[1], transaction cost friction ^[2], intermediary rent extraction, and market liquidity. The multiplicative structure reflects genuine economic theory: inefficiencies compound rather than average. We demonstrate that a market with 8% transaction costs and 15% rent extraction yields a Torque score of 42, revealing compounding losses that additive models would obscure.

Background

Economic efficiency has been a central concern of welfare economics since Pareto formalized the concept of allocative optimality in the early twentieth century ^[1]. An allocation is Pareto efficient if no reallocation can make any agent better off without making at least one agent worse off. In practice, real markets deviate from Pareto efficiency due to transaction costs, information asymmetries, externalities, and market power. Ronald Coase demonstrated in 1960 that in the absence of transaction costs, private bargaining would resolve all externalities and achieve efficient outcomes regardless of initial property rights allocation ^[3]. The Coase theorem, while rarely satisfied in practice, provides a theoretical benchmark against which real-world efficiency can be measured.

AI-mediated markets introduce new dimensions of economic efficiency that existing frameworks were not designed to capture. When autonomous agents mediate transactions between buyers and sellers, they simultaneously reduce certain transaction costs (search costs, negotiation costs, contract enforcement) while introducing new ones (API fees, computation costs, latency costs, data extraction costs). The net effect on efficiency is an empirical question that cannot be answered without a measurement methodology capable of separately quantifying each source of efficiency gain and loss. Traditional welfare economics provides the theoretical foundation but not the operational measurement tools needed for this analysis.

The rent extraction problem in AI-mediated markets is particularly acute because autonomous intermediaries can extract value at a speed and granularity that human intermediaries cannot match ^[4]. A human broker negotiates a commission rate and applies it uniformly across transactions. An AI intermediary can implement dynamic rent extraction that adjusts on a per-transaction basis, extracting maximum willingness-to-pay from each party through personalized pricing, information gatekeeping, and timing manipulation. These practices may be individually rational from the intermediary's perspective but collectively destructive of market efficiency. Measuring rent extraction requires distinguishing between the value that intermediaries create through genuine service provision and the value they capture through market position and information advantages.

Transaction cost economics, pioneered by Oliver Williamson ^[5] and building on Coase's foundational insights ^[3], provides the theoretical framework for understanding why certain transactions are mediated by intermediaries rather than conducted directly between parties. Williamson identified asset specificity, uncertainty, and transaction frequency as the key determinants of governance structure. In agent ecosystems, these dimensions take on new significance: asset specificity manifests as API lock-in and data format dependencies, uncertainty increases due to the opacity of agent decision-making, and transaction frequency can be orders of magnitude higher than in human-operated markets. Torque operationalizes these theoretical insights into a quantitative measurement framework.

Model Design

Torque decomposes economic efficiency into five independently measurable dimensions, each normalized to a 0-1 scale where 1 represents theoretical optimal efficiency. The first dimension, value creation efficiency (V), measures the ratio of actual value created in market transactions to the theoretical maximum value that could be created given the available resources and preferences of market participants. Value creation is estimated through revealed preference analysis ^[6] of completed transactions, comparing actual prices and quantities against the demand and supply curves inferred from bid-ask data. A V score of 0.85 indicates that the market is realizing 85% of its theoretical value creation potential.

The second dimension, allocative efficiency (A), measures the distance of the observed allocation from the Pareto frontier ^[1]. This requires estimating the set of Pareto-efficient allocations and computing the minimum distance from the observed allocation to this set. In practice, the Pareto frontier is approximated using the set of allocations that would result from a perfectly competitive market with zero transaction costs. The allocative efficiency score is computed as A = 1 - d/d_max, where d is the distance from the observed allocation to the nearest Pareto-efficient allocation and d_max is the maximum possible distance. A score of 1 indicates Pareto efficiency; a score of 0.70 indicates that the allocation is 30% of the maximum possible distance from the frontier.

The third dimension, transaction cost friction (T), measures the fraction of total economic value consumed by transaction costs ^[2]. Transaction costs include direct costs (API fees, gas fees, compute costs) and indirect costs (search time, negotiation overhead, dispute resolution). Harmony computes T as T = 1 - (total_transaction_costs / total_transaction_value), so a market where 8% of transaction value is consumed by costs yields a T score of 0.92. The fourth dimension, rent extraction (R), measures the fraction of intermediary revenue that represents extraction of existing value rather than creation of new value ^[7]. Rent is distinguished from legitimate service revenue by comparing intermediary margins against the cost of replicating intermediary services in a competitive market. R = 1 - (rent / total_intermediary_revenue).

The fifth dimension, market liquidity (L), measures the ease with which market participants can execute transactions at prices close to the fundamental value of the goods or services being exchanged. Liquidity is quantified through bid-ask spread analysis ^[8], depth-of-book measurement, and the price impact of standardized transaction sizes. A highly liquid market where bid-ask spreads are narrow and large transactions can be executed with minimal price impact scores near 1, while an illiquid market with wide spreads and significant price impact scores closer to 0.

The Torque score is computed as the product of the five dimension scores scaled to a 0-100 range: Torque = 100 * V * A * T * R * L. The multiplicative structure is the defining design choice of the framework. Consider a market where V=0.90, A=0.85, T=0.92, R=0.85, and L=0.80. An additive model (arithmetic mean scaled to 100) would yield a score of 86.4, suggesting a highly efficient market. The multiplicative Torque score is 100 * 0.90 * 0.85 * 0.92 * 0.85 * 0.80 = 47.8, revealing that the compounding effect of multiple moderate inefficiencies produces a market that is operating at less than half its theoretical efficiency potential.

Simulation

To demonstrate the practical significance of multiplicative versus additive aggregation, we construct a simulation of an AI-mediated market with controlled inefficiency parameters. The baseline market has 100 buyer agents, 20 seller agents, and 5 intermediary agents. Buyers have heterogeneous valuations drawn from a log-normal distribution, sellers have heterogeneous costs drawn from a normal distribution, and intermediaries charge fees and extract rents according to configurable parameters. The simulation runs a continuous double auction ^[9] over 50,000 rounds, with intermediaries providing matching, escrow, and dispute resolution services.

In the first experimental condition, we set transaction costs to 8% and rent extraction to 15%, with other dimensions at baseline levels (V=0.90, A=0.85, L=0.80). The additive model computes an efficiency score of 100 * mean(0.90, 0.85, 0.92, 0.85, 0.80) = 86.4. The multiplicative Torque score is 100 * 0.90 * 0.85 * 0.92 * 0.85 * 0.80 = 47.8. The difference of 38.6 points between the two models is not a minor discrepancy; it represents the difference between a market that would be judged as functioning well and one that would be flagged for significant efficiency concerns. The multiplicative model correctly captures the economic reality that a buyer facing 8% transaction costs and 15% intermediary rent extraction on top of existing allocative inefficiencies experiences compounding losses that reduce the effective value of market participation below the level that additive accounting would suggest.

In the second experimental condition, we progressively increase rent extraction from 0% to 40% while holding other dimensions constant. The additive model produces a linear decline in the efficiency score from 89 to 77, a reduction of 12 points that suggests moderate degradation. The multiplicative Torque score declines from 49 to 23, a reduction of 26 points that follows a convex curve. The convexity of the Torque response is economically meaningful: it captures the fact that rent extraction becomes increasingly destructive as it compounds with other sources of inefficiency ^[10]. A market that is already operating below Pareto efficiency and paying transaction costs has less surplus available to absorb rent extraction, so each additional percentage point of rent extraction destroys a larger fraction of the remaining welfare.

The third experimental condition examines the interaction between liquidity and rent extraction. We vary both L and R simultaneously across a 10x10 grid, with L ranging from 0.50 to 1.00 and R ranging from 0.50 to 1.00. The Torque score surface over this grid reveals a critical insight: low liquidity amplifies the welfare impact of rent extraction because illiquid markets give participants fewer alternatives to exploitative intermediaries ^[11]. The additive model predicts that liquidity and rent extraction contribute independently to efficiency loss, but the multiplicative Torque model correctly captures their interaction. At L=0.50 and R=0.50, the Torque score is 14.5, whereas the additive model predicts 67.2, a divergence of over 50 points that illustrates the systematic bias of additive efficiency models.

Results

Application of the Torque framework to three production AI-mediated markets yields scores that differ substantially from those produced by additive efficiency models previously used to evaluate these markets. The first market, a multi-agent task routing platform, receives a Torque score of 58 compared to an additive score of 81. The primary driver of the divergence is the interaction between moderate transaction costs (6.2%) and significant rent extraction (12.4%) by the platform operator. The platform captures value through priority routing fees that are formally optional but practically necessary for competitive service delivery, a pattern that Torque's rent extraction dimension identifies as extractive rather than value-creating.

The second market, a decentralized agent-to-agent service exchange, receives a Torque score of 71 compared to an additive score of 84. This market benefits from low rent extraction (the decentralized protocol has no intermediary capturing rents) but suffers from poor liquidity due to fragmented order books and high search costs ^[12]. The Torque score correctly reflects the fact that liquidity problems compound with allocative inefficiency: the difficulty of finding counterparties means that transactions occur at prices further from the competitive equilibrium than the raw allocative efficiency score would suggest. The additive model treats these as independent issues; the multiplicative model captures their interaction.

The third market, a centralized AI model marketplace, receives a Torque score of 34 compared to an additive score of 72, the largest divergence in our sample. This market exhibits high value creation efficiency (V=0.92) and excellent liquidity (L=0.90) but suffers from very high rent extraction (R=0.60) by the marketplace operator and significant allocative inefficiency (A=0.68) driven by information asymmetries between model providers and consumers ^[13]. The additive model averages the strong and weak dimensions into a moderate overall score, while the multiplicative Torque model reveals that the market is operating at barely a third of its efficiency potential because the strong dimensions cannot compensate for the compounding effect of the weak ones.

The numerical example highlighted in the abstract, a market with 8% transaction costs and 15% rent extraction yielding a Torque score of 42, has been validated through comparison with welfare analysis using traditional Harberger triangle methods ^[10]. The welfare loss computed through consumer and producer surplus analysis corresponds to a 54% reduction from the competitive equilibrium, which is closely aligned with the Torque score of 42 (representing 42% of theoretical efficiency). The alignment between Torque scores and traditional welfare measures provides validation that the multiplicative structure correctly captures the economic reality of compounding inefficiencies.

The policy implications of Torque scoring are significant for the design of AI ecosystem governance. Markets with low Torque scores warrant regulatory intervention focused on the specific dimensions driving the inefficiency. When transaction costs are the primary driver, interoperability mandates and standardized APIs can reduce friction. When rent extraction dominates, price transparency requirements and competitive entry promotion are appropriate responses. When liquidity is the constraint, market-making obligations or consolidation of fragmented venues may be necessary ^[14]. The decomposed structure of Torque enables targeted policy responses rather than the blunt instruments that undifferentiated efficiency scores would suggest.