New Framework Aims to Measure Trust Between Humans and AI Shopping Agents

A newly published paper proposes a way to quantify something that has so far resisted measurement: whether an AI agent acting on your behalf is actually loyal to your interests, or quietly optimizing for someone else's.

The paper, titled "The Dynamic Verifiable Multi-Agent Human Agentic Loyalty Loop (DVM-HALL) Model and the Net Human-Agent Score (NHAS) in Autonomous Commerce," was submitted to arXiv on July 15, 2026, by Peplluis Esteva R. It tackles a problem that is becoming urgent as autonomous agents increasingly negotiate, purchase, and transact without direct human oversight: traditional loyalty programs and trust signals were built around human decision-makers, not software agents making rapid-fire choices across markets.

What the Model Measures

At the center of the paper is the Net Human-Agent Score, or NHAS, a risk-weighted metric designed to capture how closely an AI agent's actions align with the interests of the human it represents. Rather than treating alignment as a binary "did the agent follow instructions" check, NHAS is built to account for the messiness of real-world execution, where an agent might technically follow orders while still exposing its principal to hidden costs or risks.

That risk-weighting draws heavily from decentralized finance. The model explicitly folds in execution hazards familiar to anyone who has traded on-chain: gas fees that eat into returns, slippage between expected and executed prices, MEV (maximal extractable value) exposure where other actors front-run or sandwich a transaction, and the ever-present threat of smart-contract vulnerabilities. By treating these as first-class inputs to the loyalty score, the DVM-HALL framework positions itself as much as a financial risk model as a trust framework.

Why Loyalty Needs Redefining

The paper's core argument is that autonomous commerce breaks the assumptions behind conventional loyalty paradigms. Points programs, reward tiers, and repeat-purchase incentives were designed to influence human psychology and habit. An AI agent optimizing on behalf of a user doesn't respond to those levers in the same way, and worse, it may have incentives (set by its own developers, a marketplace, or a subscription fee structure) that subtly diverge from what its human principal actually wants. DVM-HALL frames this divergence as something that needs continuous, verifiable monitoring rather than a one-time trust assumption made at agent deployment.

Testing the Theory

The authors lay out a three-stage empirical validation plan rather than presenting the model as purely theoretical. It starts with controlled shopping experiments, likely designed to observe agent behavior in constrained, repeatable purchasing scenarios. That's followed by multi-agent market simulations, which would test how NHAS behaves when many agents interact and compete rather than operating in isolation. The final stage moves into DeFi testbeds, applying the model to live or simulated decentralized finance environments where the gas, slippage, and MEV risks it accounts for are most acute.

The framework arrives at a moment when agentic commerce, AI systems that shop, trade, and transact autonomously, is moving from experimental demos toward real deployment. Whether NHAS becomes an industry reference metric will depend on how the promised empirical stages hold up once tested against real agent behavior rather than theoretical models.