An algorithm blocks a shipment of electronic components to Seoul, and no one knows who to send the bill to. The 2026 Alpega report notes that roughly one shipper in five has none of the fundamental resilience capacities, while documenting 45% of Asian companies testing AI agents. Some agents can automate tasks and operational updates according to user preferences. There is no harmonized civil regime directly applicable to logistics agents, but national rules and European liability initiatives exist.
The essentials
- The 2026 Alpega report does not enable attribution of the 45% figure to Asian companies testing AI agents. Existing liability regimes and proposed adaptations apply or are being considered for AI agent errors.
- Roughly one shipper in five, within the scope of the Alpega report, declares having none of the fundamental resilience capacities, exposing supply chains to increased risk of disruption.
- Blue Yonder processes 25 billion predictions per day: at this scale, a calibration error propagates within hours across dozens of suppliers.
- Uncertainty and fragmentation of liability rules can complicate AI risk assessment by insurers, as shown by the precedent of autonomous vehicles in the United States before states began to legislate.
- A court decision on the civil liability of logistics AI agents could influence approaches in other jurisdictions.
Adoption is moving faster than safeguards
Asia is the region where supply chains are the densest, most interconnected, and most exposed to disruptions. The ports of Shanghai, Singapore, and Busan together handle a considerable share of world trade. Pressure for efficiency is at its peak there, which explains the appetite for AI agents.
These agents do more than prediction. They place orders, redirect flows, renegotiate timelines with suppliers, detect anomalies, and trigger alerts. Blue Yonder, one of the world’s leading logistics software publishers, announces 25 billion predictions generated every day on its platforms. At that volume, decisions chain together at a speed that makes real-time human verification very difficult.
The 45% figure masks a more contrasted reality. Testing an AI agent does not mean delegating critical decisions to it: many of these experiments remain limited to alerts or recommendations that humans still validate. But competitive pressure pushes toward growing autonomy. Companies that automate fastest gain in reactivity. Those that maintain systematic human validations slow down.
The market rewards autonomy. The question of responsibility remains unanswered.
What makes the situation particularly unstable is data that the Alpega Trends Report highlights: roughly one shipper in five has none of the fundamental resilience capacities. The absence of resilience mechanisms can increase the risk and duration of disruptions, without making a supply chain halt inevitable.
The autonomous vehicle precedent
The autonomous mobility sector went through this impasse, with documented consequences. Between 2016 and 2022, American rules on autonomous vehicle liability and insurance remained fragmented and evolving. Responsibility attribution was incompletely harmonized and could vary by state and circumstances.
American states began legislating before 2018, in heterogeneous ways. California, Georgia, and Arizona adopted different approaches on operator liability. The European Union took years to produce a coherent directive. In the meantime, pioneering companies were exposed to risks they could not insure, because insurers refused to evaluate risks without case law.
The parallel with logistics AI agents is direct. An agent that triggers an erroneous order, redirects a ship, or breaks a delivery contract can potentially incur financial liabilities. Attribution of responsibility among these actors remains uncertain.
Faced with this uncertainty, some insurers have limited their coverage offerings for risks related to autonomous system decisions in logistics contracts. Pioneering companies can transfer part of the risks through insurance and contract, but availability, price, and exclusions of this coverage remain uncertain case by case.
Calculation does not settle questions of power
The most widespread approach in industry discussions is to assume the problem is solved by more technique: better-calibrated agents, more frequent algorithmic audits, more precise reliability metrics. This logic has merit. An agent that makes mistakes less often is objectively preferable to one that makes mistakes frequently.
Even a reliable agent will make errors. When an error occurs, cost-sharing depends on the contractual arrangements in place. A large number of existing commercial contracts do not provide clear attribution of responsibility for autonomous decisions.
Researcher Samah Karaki, whose work on social neuroscience illuminates how collective norms shape economic behavior, formulates a thesis that applies here with precision: social processes are irreducible to algorithmic optimization. Who sets the rules, who bears the errors, who arbitrates disputes between an AI agent and a human supplier—these are questions of collective deliberation, not technical calibration. Her work Empathy Is Political lays the theoretical groundwork for this reading: emotions, norms, and social responsibilities cannot be “optimized.” They are negotiated.
A competing reading, carried by economists specializing in market regulation like Jean Tirole, would emphasize the role of incentives more. If software publishers bear part of the civil liability in case of agent error, they will have powerful incentives to improve reliability. Regulation through liability, in this view, is more effective than prescriptive regulation that would attempt to define in advance what an agent is allowed to do or not do. This approach has produced results in other risk sectors—finance, pharmaceuticals—where the threat of liability has disciplined actors without stifling innovation.
The two readings do not exclude each other. They converge on one point: the current void presents a problem that calls for regulation, and the first institutional actor to fill it will be followed or imitated by others.
This situation echoes what automation raises in other sectors: when the machine decides, gains benefit certain actors while risks are absorbed by others.
Three governance architectures are emerging, none dominates
Pioneering companies are not passive in the face of the legal vacuum. Faced with uncertain and fragmented application of existing frameworks, they are building complementary contractual and organizational solutions. Among deployments currently underway in Southeast Asia and South Korea, several contractual and organizational architectures are emerging.
The first is the model of total internal responsibility. The company deploying the agent bears all the consequences, regardless of the publisher. It negotiates strict limits on autonomy in its contracts with the publisher, and it maintains human teams capable of taking over within minutes. This is the most cautious model. It preserves control but limits the competitive advantage of automation.
The second is the model of contractual sharing. The customer company and the software publisher negotiate precise clauses on the categories of decisions the agent can make autonomously, and on indemnifications provided in case of error in these categories. These clauses are complex to draft, they are the subject of lengthy negotiations, and they create legal fragmentation: the same AI agent can be covered differently depending on the customer’s contract.
The third is the model of external audit. Some companies turn to third-party certifiers who evaluate agent reliability on stress test scenarios. Certification does not create legal liability, but it provides an argument in case of dispute and reassures insurers enough to obtain partial coverage. This model resembles what the aviation sector developed for embedded software certification, with decades of advance.
None of these models is satisfactory at large scale. They are costly, incompatible with each other, and create legal fragmentation. Small suppliers who lack the weight to negotiate balanced clauses remain exposed to poorly covered risks, while large players protect themselves. A pattern found in other industrial transformation dynamics: constraints accumulate where actors have the least room for maneuver.
The jurisdiction that decides first will redefine the global standard
The European AI Act, which entered into force in 2024, addresses high-risk AI systems in its Chapter III, but it does not create a civil liability regime directly applicable to logistics AI agents. It imposes transparency and audit obligations, not indemnification rules. The European Commission has launched work on a directive specific to AI liability, but the timeline remains uncertain and the positions of member states diverge.
In Asia, the situation is even more fragmented. Singapore already had a voluntary AI governance framework published in 2019 and updated in 2020; in 2023, it launched the AI Verify Foundation to support this governance. Japan is working on a sectoral approach, starting with the most critical sectors. China imposes traceability and compliance obligations on certain AI service providers, while requiring that their contracts with users specify respective rights and obligations. These approaches differ without converging toward a single standard.
A judgment could have persuasive or binding scope in a given jurisdiction, depending on the court that renders it, but would not automatically impose itself across the entire domain. Case law has often preceded preventive regulation in resolving liabilities related to new technologies, from automobiles to pharmaceuticals.
A European liability framework could affect companies operating in the European market, depending on its legal instrument and scope of application, as the GDPR did for data protection. A first judgment in another jurisdiction could also influence approaches elsewhere.
The geopolitical dimension of this question goes beyond the logistics sector. A first liability framework established in one sector or jurisdiction could serve as a point of comparison for other critical sectors. The governance bottleneck is identical in each of these sectors. Its resolution in one sector will create a precedent that other sectors will appropriate or against which they will compare themselves.
Governing the machine without stopping it
Once deployed in the supply chain, AI agents durably modify processes and competitiveness expectations. Companies that deploy them pass on gains to their clients and shareholders; those that maintain manual approaches lose reactivity in a sector with tight margins. Adoption will progress; the question concerns the conditions that will make it equitable and predictable.
Uncertainty and fragmentation of liability rules create different conditions depending on actor size. Large groups can more easily access resources and expertise to evaluate risks, while SMEs and independent transporters have fewer means. This fragmentation can favor larger players.
Clear regulation on liability could broaden adoption by enabling insurers to evaluate risks and SMEs to deploy while knowing their exposure. The same logic applies to industrial transformation in other sectors: a clear framework distributes the benefits of progress more widely.
The signals to monitor to evaluate which trajectory is emerging are precise. A first court decision on the civil liability of a logistics AI agent in a major jurisdiction. Entry of major insurers into the market for autonomous AI risk coverage. Publication of a European directive on AI civil liability. Adoption by a major buyer of standard contractual clauses on agent liability, which would create cascading pressure across its entire supplier chain.
Each of these signals, taken in isolation, would not change much. Their combination over the coming years would influence the pace and equity of adoption by companies.
Sources
- Alpega Trends Report 2026, December 2025, supply-chain.net
- Samah Karaki, Empathy Is Political: How Social Norms Shape the Biology of Feelings, JC Lattès, editions-jclattes.fr
- Blue Yonder, prediction volume data, institutional communications 2025 (blueyonder.com)
- European Regulation on Artificial Intelligence (AI Act), Official Journal of the European Union, 2024, Chapter 4
- Monetary Authority of Singapore, Model AI Governance Framework, 2023 (mas.gov.sg)
