DVM-HALL Paper Models How AI Shopping Agents Choose Brands
A conceptual arXiv paper proposes a choice model for AI shopping agents and a Net Human-Agent Score to rate them. Empirical validation is planned but not yet done.
TL;DR
- A conceptual paper models how AI shopping agents pick brands, weighing human emotional ties against API quality, trust, delegated authority and execution risk [^1].
- It also proposes a Net Human-Agent Score for rating such agents, but reports no empirical test; its three-stage validation is only planned [^1].
Background
Loyalty research has long assumed that a human makes the purchase, and the Net Promoter Score has served as its standard measure [^1]. AI agents that can complete purchases on their own change that assumption. The paper argues that traditional loyalty models miss "algorithmic bounded rationality", where technology decides which options a shopper sees at all [^1]. Its model uses a softmax function, a standard way to turn scores for several options into choice probabilities that add up to one.
What happened
Sai Srikanth Madugula and Daya Shankar of Woxsen University in Hyderabad and Peplluis Esteva de la Rosa of the Universitat de Girona propose the DVM-HALL model, short for Dynamic Verifiable Multi-Agent Human Agentic Loyalty Loop [^1]. The human side of a brand's appeal combines emotional brand equity, remembered satisfaction, perceived value and normative fit, such as sustainability or privacy preferences. The agent side adds machine-experience features like API availability, latency, delivery reliability and explainability, minus execution risks, plus loyalty rewards [^1]. A delegation coefficient sets how much authority the human has handed to the agent, a trust coefficient sets how much the agent's assessment counts, and hard limits such as budgets or blocked merchants act as penalties [^1]. Trust and delegation update after every task: alignment with the user's intent, execution quality, explanations and recourse raise trust, while losses and constraint violations lower it [^1].
The authors stress that emotion doesn't drop out. In their words, agentic commerce "transforms the channel through which emotional equity operates; it does not eliminate emotional equity" [^1]. For tokenized loyalty programs and decentralized finance, the model treats gas costs, slippage, MEV exposure and smart-contract vulnerabilities as predictors of which brand an agent prefers [^1].
The second contribution is the Net Human-Agent Score, or NHAS. Each interaction gets a score from 0 to 10, built from preference alignment (weight 0.35), economic efficiency (0.25), execution reliability (0.20) and explanation quality (0.20), and set to zero if any hard constraint was broken [^1]. Scores of 9 or more count as positive, scores of 6 or less as negative, with extra penalties for financial loss, privacy leaks and compliance failures. NHAS is 100 times the average of these values [^1]. Its inputs are human feedback, execution logs, comparison with the alternatives available at the time and verifiable receipts [^1].
Why it matters
If the model holds, brands that sell through agents would compete partly on technical qualities such as API quality and delivery reliability, not only on emotional appeal [^1]. NHAS also offers a way to separate small slips, such as wasting a few loyalty points, from serious failures, such as exceeding a spending limit because of a manipulated oracle [^1]. The authors suggest it could serve as a reward signal for fine-tuning agents or as a metric for monitoring them after deployment [^1].
For now, it is a framework without evidence. The paper synthesizes earlier literature and proposes a three-stage validation: controlled shopping experiments with participants, multi-agent market simulations and DeFi testbeds [^1]. Its published Python code generates synthetic data [^1]. The authors name limits themselves: governance tools to prevent misplaced trust are not yet deployed at scale, and security and privacy concerns hold back adoption in emerging economies [^1].
Practical example
Say you let a shopping agent reorder household supplies within a monthly budget. In DVM-HALL terms, that budget is a hard constraint [^1]. Under NHAS, an order that breaks it scores zero and counts against the agent, however cheap it was. An order that matches your preferences at a good price, with a clear explanation, can score 9 or more and count in its favor [^1]. When you override a choice, that becomes human feedback, and in the model repeated violations would lower your trust and the authority you hand the agent [^1].