Brands risk losing customers when autonomous purchasing agents act on behalf of humans unless loyalty systems are redesigned to account for agent trust, delegated authority and execution risk. The paper proposes a formal DVM-HALL model and an auditable Net Human‑Agent Score to quantify human–agent alignment and recommends controlled experiments, multi-agent simulations and DeFi testbeds to validate the framework.

Main Finding

The paper introduces the Dynamic Verifiable Multi-Agent Human Agentic Loyalty Loop (DVM‑HALL), a formal model reconceptualizing brand choice in settings where autonomous agentic AIs act as purchasers on behalf of humans. It couples human emotional equity and constraints with an agentic utility that explicitly includes verifiable execution risks (gas, slippage, MEV, oracle and smart‑contract risk, etc.), and it proposes the Net Human‑Agent Score (NHAS) — an auditable, risk‑weighted metric for measuring human–agent alignment across interactions. The model is dynamic: trust and delegation update recursively after each interaction, and choice is modeled by a softmax over composite utilities.

Key Points

• DVM‑HALL reframes loyalty as a tripartite, dynamic loop among human, personal agent, and brand (plus infrastructure agents). Loyalty now bifurcates into human emotional attachment and algorithmic (agent) preference.
• Composite utility: Uhbt = (1 − δht)·UH_hbt + δht·Tha,t·UA_hbt − Ωhbt.
  • UH (human utility) captures emotional brand equity, satisfaction, perceived value, norms.
  • UA (agent utility) captures positive machine‑experience features (APIs, latency, liquidity) and negative execution risks (gas, slippage, MEV, oracle risk, smart‑contract risk, bridge risk). Loyalty rewards enter as Rloy.
  • δht ∈ [0,1] is delegation; Tha,t ∈ [0,1] is calibrated trust; Ωhbt encodes hard constraints.
• Choice probability across brands is softmax of Uhbt with temperature τ (controls exploration).
• Trust and delegation evolve: trust updates as a logistic function of alignment, execution quality, explanation and recourse, losses and constraint violations; delegation updates depend on prior delegation, updated trust, NHAS, task risk, and desire for control.
• NHAS: an auditable interaction‑level score aggregating preference alignment (P), economic efficiency (E), execution reliability (R), explanation quality (X), and hard‑constraint satisfaction (H). Interaction score Si = 10·clip[H·(0.35P + 0.25E + 0.20R + 0.20X), 0,1]. A risk‑weighted Wi penalizes misalignment and catastrophic events (loss, privacy breach, compliance failure). NHAS = 100·(mean Wi).
• NHAS requires four verifiable data sources: human feedback/overrides, execution logs/decision traces, a benchmark engine (counterfactual alternatives), and verifiable receipts/transaction proofs.
• Empirical validation plan: (1) controlled human–agent shopping experiments (manipulate emotional equity, API quality, transparency, delegation), (2) multi‑agent market simulations, (3) DeFi/tokenized loyalty testbeds measuring blockchain execution variables.
• Governance and ethics: calls for oversight frameworks (TRACE, HAIG), attention to trust‑miscalibration (over‑reliance or under‑reliance), risk of dark patterns and persuasive autonomy, and unequal adoption due to security, privacy, and digital‑literacy gaps.

Data & Methods

• The work is primarily theoretical and synthetic: integrates literature on human–machine teaming, decision theory, algorithmic trust, DeFi mechanics, and loyalty research to derive an estimable choice model.
• Formal model elements:
  • Human utility UH_hbt linear in emotional equity, remembered satisfaction, perceived value, normative fit plus noise.
  • Agent utility UA_hbt = βᵀ x_MX_bt − ρᵀ r_EX_bt + β_R Rloy_hbt + noise. Table of concrete DeFi execution variables provided (Verifiability, Gas, Slippage, Liquidity, MEV, Oracle risk, Smart‑contract risk, Bridge risk, Revocation ease).
  • Dynamic recursive update equations:
    • Trust ˜Tha,t+1 = ζ0 + ζ1Tha,t + ζ2Aht + ζ3Qht + ζ4Xht + ζ5Rht − ζ6Lht − ζ7VIOht; Tha,t+1 = sigmoid(˜Tha,t+1).
    • Delegation δh,t+1 = sigmoid(κ0 + κ1δht + κ2Tha,t+1 + κ3 NHASht/100 − κ4 Risktask,t + κ5 Controlht).
• NHAS computation:
  • Per‑interaction Si weighted sum of P, E, R, X if hard constraints Hi satisfied; scale to 0–10.
  • Wi applies ±1 rewards/penalties for high/low Si and subtracts weighted catastrophic flags (financial loss, privacy, compliance); clip to [−1,1].
  • Aggregate NHAS = 100·mean(Wi) over N interactions.
• Proposed empirical data sources and validation:
  • Human subjects (laboratory) for behavioral and trust/delegation dynamics.
  • Simulated multi‑agent markets to test predictive power of machine vs. human variables under different δ and Tha,t distributions.
  • Blockchain/DeFi testbeds collecting verifiable execution logs (transaction hashes, gas/slippage records, oracle feeds, failure/exploit flags).
• Limitations acknowledged: need for large‑scale governance tools, cross‑cultural validation, security/privacy constraints affecting deployment, and measurement challenges for benchmark counterfactuals.

Implications for AI Economics

• Redefines demand: agents acting on behalf of humans change revealed preference signals. Firms will need to optimize for both human emotional equity and for agent‑readable, verifiable execution qualities (APIs, low slippage, verifiability).
• Two‑sided competition: brands compete for (a) human trust/affect and (b) algorithmic preference. Sellers may invest in machine‑facing infrastructure (APIs, attestations, low‑latency execution, smart‑contract security) to attract agentic buyers — shifting investment from pure marketing to execution/infrastructure.
• Price formation and transaction costs: inclusion of verifiable execution variables (gas, MEV, slippage, oracle risk) in agent utility makes fine‑grained transaction costs and microstructure externalities central to demand modeling. Agents will optimize across these execution frictions, potentially fragmenting liquidity and altering price dispersion.
• Platform and gatekeeper power: platforms that mediate verifiable signals, benchmarks, and trust infrastructure (or provide NHAS‑like auditing) obtain new rents; they can influence agent routing via trust signals, default policies, and marketplace governance.
• Measurement and regulation: NHAS offers a candidate standardized metric to audit agentic behavior and human–agent alignment; regulators and firms could use NHAS (or variants) to monitor harm, require minimum auditability, and calibrate liability across agents/brands/platforms.
• Welfare and distributional effects: adoption gaps (security, privacy, literacy) could amplify inequality — those with reliable agents gain better execution and lower costs. Catastrophic execution risks (oracle manipulation, smart‑contract exploits) become systemic welfare concerns requiring public goods (secure oracles, audit standards).
• Strategic behavior and market design: sellers might game agent preferences (e.g., presented reward structures, engineered verifiability signals) — creating a need for robust benchmark engines and verifiable receipts to prevent manipulation. Tokenized loyalty and programmable incentives introduce new dynamic contracting problems between brands and buyers’ agents.
• Research agenda: empirical estimation of the model’s parameters (role of δ and Tha,t), quantifying elasticity of demand with respect to execution risks, testing how NHAS correlates with long‑run retention and welfare, and designing market mechanisms (pricing, disclosure, auditability) that align agentic optimization with social objectives.

Summary conclusion: DVM‑HALL and NHAS formalize how agentic AIs change the economics of loyalty and consumption — shifting the locus of competition toward verifiable execution quality and trust infrastructure, and creating measurable pathways for policymakers and firms to evaluate and govern autonomous commerce.