Study identifies a delegation gap in user acceptance of intelligent agents on matchmaking platforms between sending and receiving

Digital platforms and matchmaking apps are increasingly adopting a new design pattern based on delegating autonomous intelligent agents, powered by large language models, to speak and conduct conversations on behalf of users. However, the success of this pattern faces a structural obstacle identified in a new paper published by researchers Daria Leshchikova, Valentina Kuskova, Dmitry Zaytsev, and Valery Klimov on arXiv. The study reveals that the viability of these systems does not depend solely on an individual's willingness to deploy their own agent, but hinges on an overlooked condition: their willingness to receive messages managed by an intelligent agent representing the other party.

The paper drew on two large-scale surveys of active users on a major platform. The first survey covered 2,894 users to examine generative profile features, while the second surveyed 2,617 users to evaluate autonomous conversational agents and was conducted in two different languages. The research team developed a latent variable measurement model using graded response models with latent regression. Statistical analysis demonstrated that the willingness to send an intelligent agent and the willingness to receive one are statistically distinct constructs. Although closely correlated at 0.92, they remain distinct, with a Bayesian Information Criterion difference of 52 and measurement invariance holding partially across languages.

The study identified a phenomenon it termed the "asymmetric delegation gap," where a user deploying their own agent requires a much lower acceptance threshold compared to their willingness to interact with an agent deployed by another party.Measurements showed that the acceptance threshold for deploying a personal agent was -0.38, whereas the acceptance threshold for interacting with an interlocutor's agent jumped to +0.32, reaching +1.39 when measuring full interaction. In practice, this means users are on average roughly three times more inclined to deploy their own agents than to engage with agents sent by others.

When applying a random pairing scenario based on stated acceptance levels, results showed that only 4% to 13% of directed dyads actually combine an agent deployment on one side with recipient engagement on the other, alongside a clear gender-directional imbalance. These figures demonstrate that allowing the system to operate unassisted without accounting for recipient acceptance causes a large share of potential interactions to collapse, as the majority seeks the convenience of having an agent speak on their behalf while refusing to spend time reading automated responses from a potential partner's agent.

Empirical models in the paper showed that design interventions clearly determine the efficiency of agentic recommender platforms.Testing mandatory reciprocity halved or more than halved total interaction volume, as roughly two-thirds of potential agent deployments were screened out. Conversely, routing agent communications based on recipient receptivity tripled the interaction rate per contact, an improvement that maintained accuracy and robustness in out-of-sample validation with an area under the curve of 0.88 and a 3.1-fold gain across cross-validation folds.

The paper concludes that developing agentic recommender and matchmaking systems requires rethinking transparency mechanisms and speaker disclosure, adopting explicit opt-in protocols, and engineering matching algorithms that account for mutual acceptance levels. The success of autonomous agents in dyadic interaction environments will not come simply from equipping users with a language tool to act on their behalf, but demands balancing sending efficiency against receiving flexibility to ensure interactions remain worthwhile for both sides.