Civil Law and UAE: Multi-Agent System Liability Conflicts

1. Introduction

Multi-agent system liability arises when several autonomous or semi-autonomous digital agents participate in one transaction or decision—for example, an AI purchasing agent, fraud-detection agent, pricing agent, payment agent, and compliance agent—while no single human or entity directly performs the harmful act.

The difficult legal question is:

Who is legally responsible when the harmful outcome results from the combined conduct of several AI agents, their developers, operators, data providers, platforms, and human supervisors?

UAE civil law does not presently create a separate legal personality for an AI agent merely because it acts autonomously. The more workable approach is therefore to attribute the relevant conduct to the natural or juridical persons who designed, deployed, controlled, supervised, contracted with, or benefited from the system, depending on the facts.

The new UAE Civil Transactions Law, Federal Decree-Law No. 25 of 2025, has been in force since 1 June 2026. It modernises the general civil-liability framework, while special legislation can apply where relevant.

2. Meaning of a Multi-Agent System

A multi-agent system (MAS) is a technological environment in which multiple software agents independently or semi-independently perform tasks and interact with one another.

For example:

Customer → AI negotiation agent → pricing agent → credit agent → payment agent → logistics agent

Each agent may:

• receive information;
• make recommendations;
• make decisions;
• communicate with another agent;
• execute transactions;
• modify its behaviour according to data;
• trigger another agent;
• create an outcome that was not specifically ordered by a human.

This creates a chain of causation and responsibility rather than a simple human-to-machine relationship.

3. Why Multi-Agent Liability Is Difficult

Traditional civil liability generally assumes that a human or legal entity can be identified as the actor.

With multi-agent systems, there may be:

1. Developer
2. System owner
3. AI model provider
4. Agent designer
5. Platform operator
6. Data provider
7. Human supervisor
8. Business that deployed the agents
9. Third-party service provider
10. Several autonomous agents

A harmful result may therefore have multiple causes.

Example

Suppose an autonomous trading system contains:

• Agent A — market analysis;
• Agent B — risk assessment;
• Agent C — trading execution;
• Agent D — compliance monitoring.

Agent A provides incorrect information.

Agent B incorrectly assesses the risk.

Agent C executes the trade.

Agent D fails to stop it.

The company suffers AED 10 million in losses.

The legal issue is not simply:

“Which AI agent made the mistake?”

The more important questions are:

• Who controlled the system?
• Who had a duty to supervise?
• Who selected the data?
• Who designed the safety limits?
• Who had authority to stop the transaction?
• Which person's conduct legally caused the loss?
• Were there several legally responsible persons?

4. Current UAE Statutory Framework

The current Civil Transactions Law provides a particularly important framework for analysing this problem.

Article 245 — Natural and juridical persons

The harmful-act provisions apply to liability arising from harmful acts committed by a natural or juridical person, subject to special legislation.

This is significant for AI because the legislation does not simply make an autonomous software agent the legal defendant.

The likely starting point is therefore to identify the legally responsible person or company behind the system.

5. Article 246 — General Principle of Liability

Article 246 provides that every harmful act causing damage obliges its perpetrator to compensate the injured party, even where the perpetrator lacks discernment.

For multi-agent systems, this creates an important principle:

Autonomy of software does not automatically eliminate civil responsibility.

The court can investigate the human/legal architecture surrounding the system.

6. Article 247 — Direct Act and Causation

Article 247 distinguishes between:

• direct causation, and
• causation through another factor.

Where direct and causal actors combine, the law contains an attribution rule concerning the direct actor.

This is highly relevant to multi-agent systems.

Example

If:

• Agent A produces an erroneous instruction,
• Agent B directly executes the transaction,

the court may have to determine whether:

• Agent B's execution was the legally direct cause;
• Agent A's output was merely a causal factor;
• the system operator's design or supervision was another cause.

Thus, technical causation and legal causation are not necessarily identical.

7. Article 249 — External Cause and Third-Party Acts

Article 249 recognises an external cause beyond the defendant's control, including:

• force majeure;
• sudden accident;
• act of a third party;
• act of the injured person,

subject to the statutory or contractual position.

In an AI dispute, a defendant may therefore argue:

“The loss was caused by an independent third-party model, data feed, cyberattack or external system.”

But this does not automatically succeed.

The court would need to determine whether the external event genuinely broke the causal chain or whether inadequate system design, monitoring, or security contributed to the harm.

8. Article 251 — Attribution to the Actual Actor

Article 251 states, in general terms, that an act is attributed to its perpetrator rather than the person who ordered it, subject to the statutory exceptions.

This raises an important question for autonomous systems:

Who is the "perpetrator"?

An AI agent is not ordinarily treated as an independent legal person.

Consequently, attribution may move through the technological architecture:

AI output → system operator → contractual duty/control → legal person

The court would examine the actual human and corporate relationships rather than simply treating the algorithm as the final legal actor.

9. Article 253 — Multiple Responsible Persons

This is probably the most directly relevant provision to multi-agent liability.

Article 253 provides that where multiple persons are responsible for the harm, each is responsible according to their share, while the court may order equal or joint and several liability. It also allows reduction of compensation where the injured person contributed to the harm.

Therefore, a hypothetical UAE multi-agent dispute could involve:

The court could potentially apportion responsibility according to the evidence.

10. Article 255 — Scope of Compensation

Article 255 provides for compensation according to the loss suffered and lost profit, provided the loss is a natural consequence of the harmful act.

This becomes complicated where a chain of AI decisions produces an unexpected result.

For example:

Bad data → incorrect prediction → wrong pricing → automated sale → market reaction → business loss

The claimant would need to establish the legally relevant causal connection between the system's conduct and the claimed loss.

11. Article 266 — Principal and Subordinate Liability

Article 266 provides that a principal can be liable for harm caused by a subordinate while performing the subordinate's duties or because of them, where the principal has actual authority of supervision and direction.

This is important for human-AI organisational structures.

Although an AI agent is not ordinarily an employee, the principle can become relevant where a human employee operates or supervises the system.

For example:

Company → employee → AI agents

If the employee negligently deploys or supervises the agents while performing work duties, the company may face a separate basis of responsibility.

12. Article 267 — Right of Recourse

Article 267 gives a person responsible for another's act a right of recourse against the person who actually caused the harm to the extent that person is responsible.

This could become important in a multi-agent contractual chain.

Example

A customer successfully obtains compensation from the system operator.

The operator may then seek contractual or legal recourse against:

• the AI vendor;
• software developer;
• data provider;
• cybersecurity provider;
• maintenance contractor.

Thus, the victim's claim and internal allocation of responsibility can be separate questions.

13. Article 271 — Machinery and Things Requiring Special Care

Article 271 imposes liability on a person controlling things requiring special care to prevent harm or mechanical machinery, subject to the statutory exception for harm that could not be prevented.

This provision is potentially significant for:

• autonomous robots;
• connected vehicles;
• industrial AI systems;
• autonomous machinery;
• AI-controlled equipment.

Whether a particular digital AI system falls within such a provision would depend on the factual and legal characterisation of the system.

14. Contractual Liability in Multi-Agent Systems

Multi-agent disputes may also be contractual.

Consider:

Company A hires AI Provider B.

B supplies an autonomous procurement system.

The system makes unauthorised purchases.

The claimant may allege:

• breach of contract;
• failure to meet specifications;
• failure to comply with safety requirements;
• inadequate monitoring;
• breach of confidentiality;
• failure to follow instructions;
• failure to maintain the system.

A contractual limitation clause may then become relevant.

However, contractual allocation cannot necessarily eliminate mandatory civil liability rules.

The current Civil Transactions Law expressly provides in Article 257 that a condition exempting or reducing liability arising from a harmful act is void, while permitting agreed aggravation subject to law.

15. Agency Problem

One of the most important questions is whether an AI agent should be treated as analogous to an agent.

The answer should be approached cautiously.

AI does not automatically become a legal agent merely because it communicates or negotiates.

Instead, the legal analysis may ask:

1. Who authorised the AI?
2. What authority was given?
3. What restrictions existed?
4. Did the system act within those restrictions?
5. Did the principal create an appearance of authority?
6. Did the third party reasonably rely upon that appearance?
7. Did the principal know about the AI's actions?

This question is particularly important because the recent DIFC Court of Appeal decision in Khaled Salem Musabeh Humad Al Mheiri v John Cameron [2025] DIFC CA 008 specifically examined the difficult issue of liability for representations made by an agent under UAE law. The Court considered actual and apparent authority and remitted the matter for reconsideration rather than definitively resolving all aspects of the issue.

16. Multi-Agent Conflict Types

A. Developer–Operator Conflict

The developer says:

“The customer configured the system incorrectly.”

The operator says:

“The software was inherently defective.”

The court must determine whether the loss resulted from:

• design;
• configuration;
• misuse;
• inadequate warnings;
• inadequate supervision.

B. Agent–Agent Conflict

Agent A claims:

“I only generated the recommendation.”

Agent B claims:

“I merely followed Agent A's recommendation.”

This produces a distributed causation problem.

The court may need technical expert evidence to determine which decisions were:

• necessary;
• discretionary;
• automated;
• deterministic;
• overridden;
• foreseeable.

C. Human–AI Conflict

A human supervisor may argue:

“The system made the decision autonomously.”

The claimant may argue:

“The human was responsible for monitoring the system.”

The important question becomes the scope of human supervision.

D. Platform–Third-Party Agent Conflict

A platform may use several external AI providers.

For example:

Platform → third-party LLM → payment agent → identity agent

If the system causes harm, responsibility may have to be divided among several independent businesses.

17. Important UAE Case Laws

There is currently no reported UAE judgment that establishes a complete, dedicated doctrine specifically for liability among multiple autonomous AI agents. Therefore, the following authorities are best used as analogical authorities dealing with agency, attribution, AI-generated material, digital systems, causation, multiple defendants, and technological disputes.

Case 1 — Khaled Salem Musabeh Humad Al Mheiri v John Cameron [2025] DIFC CA 008

This is particularly important for multi-agent liability.

The dispute concerned representations made by an alleged agent and whether the principal could be responsible for those representations.

The DIFC Court of Appeal considered:

• agency;
• actual authority;
• apparent/ostensible authority;
• deceit;
• UAE Civil Code provisions;
• the responsibility of a principal for an agent's conduct.

The Court ultimately held that the reasoning concerning the principal's responsibility was insufficient and remitted the matter for retrial.

Relevance

This provides a useful framework for asking:

If an AI system acts within authority created by its operator, when should the operator bear responsibility?

It does not, however, establish that an AI system itself is a legal agent.

Case 2 — Dubai Court of Cassation No. 141 of 2006

This case is discussed in the later Al Mheiri v Cameron judgment.

It involved a parasailing accident in which the owner of the boat and the person operating it were both involved.

The Court considered when a principal can be responsible for an agent's conduct, including the requirement of actual authority to supervise and direct the agent and a connection between the harmful act and the agent's duties.

Relevance

It demonstrates the traditional UAE approach to:

principal + subordinate/agent + actual authority + harmful act

The same conceptual structure can be tested against an AI deployment architecture.

Case 3 — Klesta Eshja & Hair Creators Salon LLC v Salah Masri & Others, CFI 066/2024

This 2026 DIFC case involved pleadings prepared substantially with the assistance of AI.

The Court found that the AI-assisted defences contained false references and misleading material and ordered the relevant defences struck out, with costs consequences.

Relevance

The case demonstrates an important principle:

Using AI does not transfer legal responsibility from the human party to the AI.

The parties remained responsible for what they placed before the court.

For multi-agent systems, this supports a broader principle of human/legal accountability despite technological autonomy.

Case 4 — Stelian Gheorghe v BSA Ahmad Bin Hezeem & Associates LLP & Jimmy Haoula [2025] DIFC CFI 045

The defendants argued that parts of the claim and evidence might have been generated using AI.

The Court observed that errors of law had no place in witness evidence filed by lawyers and ultimately stayed the proceedings in favour of arbitration under the relevant agreement.

Relevance

The case illustrates:

• AI-generated material;
• responsibility for legal submissions;
• evidentiary reliability;
• human responsibility for court filings;
• interaction between technology and procedural obligations.

It reinforces the principle that AI assistance does not itself create a separate legal responsibility regime.

Case 5 — Techteryx Ltd v Aria Commodities DMCC & Others [2025] DIFC DEC 001

This was heard by the DIFC Digital Economy Court and involved a digital-economy dispute concerning several commercial entities and digital/financial transactions.

The Court dealt with complex digital-economy issues and multiple defendants within the technological and financial environment.

Relevance

The case illustrates why complex digital disputes increasingly require courts to examine:

• technological architecture;
• multiple actors;
• financial transactions;
• digital evidence;
• responsibility across interconnected entities.

It is not a judgment establishing AI multi-agent liability, but it is useful for understanding the UAE's emerging judicial infrastructure for digital disputes.

Case 6 — Anastasiia Denisova v Aleksei Galtcev & Realiste Holding Ltd [2024] DIFC CFI 041

The case concerned shares in an AI-technology platform facilitating real-estate investment. The dispute involved the rights of a former employee/shareholder and the company's ownership structure.

Relevance

This case demonstrates that when an AI enterprise becomes the subject of litigation, ordinary civil and corporate doctrines remain central:

• legal personality;
• shareholder rights;
• contractual rights;
• ownership;
• corporate responsibility.

The technological nature of the business does not by itself create an autonomous legal personality for the AI system.

Case 7 — International Electromechanical Services Co. LLC v Al Fattan Engineering LLC & Al Fattan Properties LLC [2012] DIFC CFI 004

This case is valuable for the multiple-defendant and inconsistent-liability aspect of the problem.

The Court considered the possibility that arbitration against one defendant and litigation against another could result in:

• inconsistent findings;
• duplication;
• multiple proceedings;
• potential double recovery.

The Court emphasised the need to balance those risks against the parties' arbitration agreement.

Relevance

A multi-agent dispute can similarly produce several proceedings against:

• developer;
• platform;
• operator;
• vendor;
• data provider.

The case illustrates why procedural coordination is important when responsibility is distributed.

Case 8 — Krystal Financial Consultants LLC v Nextgen Robopark Investment LLC [2025] DIFC CA 007

The DIFC Court of Appeal dealt with a dispute involving Nextgen Robopark and the jurisdictional/procedural issues surrounding the claim.

Relevance

Although not a decision establishing autonomous-agent liability, it illustrates the broader point that technology-related businesses remain subject to ordinary rules concerning:

• corporate parties;
• jurisdiction;
• contractual relationships;
• procedural responsibility.

18. DIFC Digital Economy Court

The UAE's most explicit institutional response to technologically complex disputes is the DIFC Digital Economy Court.

DIFC Rule 58.7 expressly identifies claims involving:

• artificial intelligence;
• digital assets;
• blockchain;
• substantial or complex databases;
• digital data;
• e-commerce;
• online intermediaries;
• digital payment platforms;
• other digital-economy technologies.

The rules also permit extensive use of information technology and provide mechanisms concerning digital assets and AI-driven smart forms.

This is highly relevant to future multi-agent litigation because the factual dispute may concern the architecture of an entire technological ecosystem rather than one isolated transaction.

19. AI Transparency and Evidence

The DIFC's Practical Guidance Note No. 2 of 2023 is particularly relevant.

It expects transparency concerning AI-generated material and stresses:

• verification;
• reliability;
• disclosure of AI use;
• awareness of bias;
• protection of confidentiality;
• checking AI-generated authorities and content.

It specifically warns parties not to rely on AI-generated material without verification.

For multi-agent liability litigation, this suggests that parties should preserve:

• prompts;
• system logs;
• agent-to-agent communications;
• model versions;
• training-data information where legally available;
• API records;
• human overrides;
• decision thresholds;
• audit trails;
• timestamps.

20. The "Black Box" Problem

A major issue arises where the operator cannot explain exactly why the agents produced a particular result.

Suppose:

Agent A → Agent B → Agent C → Agent D

produces an unlawful payment.

Each provider says:

“Our agent merely relied upon the output of another agent.”

This creates a liability gap.

UAE courts are likely to focus on conventional legal questions:

1. Was there a duty?
2. Was there a harmful act?
3. Was there a breach?
4. Was the conduct a cause of the damage?
5. Was the damage foreseeable/natural?
6. Was there contributory conduct?
7. Who exercised actual control?
8. Was there contractual allocation of risk?

The inability to explain the algorithm does not necessarily eliminate the legal claim.

21. Human-in-the-Loop Liability

A particularly important distinction is between:

Fully autonomous system

No human reviews decisions before execution.

Human-on-the-loop

A human monitors the system but normally does not approve each decision.

Human-in-the-loop

A human must approve significant decisions.

The greater the actual human control and supervisory responsibility, the stronger the argument for examining the human or organisation's conduct.

However, liability remains fact-specific. Merely having a human listed as supervisor does not automatically establish liability.

22. Developer Liability

A developer may potentially face liability where the evidence establishes:

• defective architecture;
• inadequate safety controls;
• foreseeable failure;
• misleading representations about system capabilities;
• inadequate warnings;
• defective updates;
• inadequate security;
• failure to comply with contractual specifications.

But the mere existence of an error does not automatically establish developer liability.

The claimant still needs to establish the relevant legal basis and causal connection.

23. Operator Liability

The operator may face a different risk.

For example, the AI model might be technically sound, but the operator:

• configured an unsafe threshold;
• disabled safeguards;
• ignored warnings;
• failed to monitor outputs;
• connected the agent to an unauthorised payment system.

Here the central question becomes deployment and supervision, rather than software design.

24. Data Provider Liability

A multi-agent system is only as reliable as some of its inputs.

A data provider may therefore become relevant if:

false data → AI decision → financial loss

The legal issue would be whether:

• the data was contractually warranted;
• the provider owed a duty;
• the information was materially inaccurate;
• the inaccuracy caused the damage;
• the provider could reasonably have detected the error.

25. Joint and Several Liability

Article 253 is especially important where several actors contribute to one injury.

For example:

If the evidence establishes responsibility against several parties, the court may consider proportional or joint-and-several responsibility under the statutory framework.

This prevents a technological system from automatically becoming a responsibility vacuum.

26. Contribution Between Responsible Parties

Suppose the injured party receives compensation from the system operator.

The operator may subsequently seek contribution from:

• developer;
• data provider;
• cloud provider;
• cybersecurity contractor;
• maintenance company.

Article 267's recourse mechanism is therefore conceptually important for distributing liability internally.

27. Contractual Allocation of Multi-Agent Risk

Commercial contracts should ideally identify:

1. Agent authority

What may the AI agent do?

2. Spending limits

What is the maximum transaction?

3. Human approval

Which decisions require human approval?

4. Audit rights

Who can inspect the agent's records?

5. Data responsibility

Who warrants the accuracy of inputs?

6. Model responsibility

Who is responsible for model defects?

7. Cybersecurity

Who bears the risk of unauthorised access?

8. Incident reporting

How quickly must a failure be reported?

9. Indemnity

Which party bears specified third-party claims?

10. Insurance

Which risks must be insured?

28. Causation in Multi-Agent Systems

A useful analytical model is:

Input → Agent Decision → Agent Interaction → Human/System Control → Physical or Financial Act → Damage

The claimant should ideally identify each stage.

Example

A financial AI system incorrectly transfers AED 5 million.

The evidence shows:

1. Data Agent received incorrect information.
2. Risk Agent failed to flag it.
3. Decision Agent authorised payment.
4. Payment Agent executed transfer.
5. Human supervisor received an alert but did not intervene.

There could potentially be several legally relevant causes.

The court would then determine the responsibility of each participant rather than simply asking which software component made the final instruction.

29. Defences

Potential defendants may rely upon:

External cause

The loss was caused by an independent third party.

Force majeure

The event was genuinely beyond control.

User fault

The claimant improperly configured or used the system.

Intervening act

Another independent actor broke the causal chain.

Lack of causation

The alleged defect did not cause the loss.

Contractual allocation

The parties allocated particular risks by contract, subject to mandatory law.

Compliance with instructions

The system acted within authorised parameters.

However, contractual allocation does not automatically defeat mandatory civil liability.

30. Evidence Required in Multi-Agent Litigation

A technologically sophisticated case may require:

• source code where legally discoverable;
• system architecture diagrams;
• API logs;
• model versions;
• agent communication records;
• transaction logs;
• access records;
• audit trails;
• system prompts;
• configuration records;
• human override records;
• cybersecurity reports;
• expert reports;
• contractual specifications;
• training and testing documentation.

The DIFC's AI guidance particularly emphasises verification and reliability of AI-generated material.

31. Main Legal Problem: Attribution

The central legal problem can be expressed as:

Who should bear responsibility when no individual actor alone caused the harm?

A useful UAE-law framework is:

Agent action → identify human/legal control → identify duty → establish causation → identify multiple contributors → apply Article 253 → determine compensation → allow recourse between responsible parties.

This avoids treating the AI system itself as an automatic substitute for legal responsibility.

32. Difference Between AI Agency and Legal Agency

These concepts must be separated.

An AI system can therefore be technically autonomous without being legally autonomous.

33. Future Development of UAE Civil Law

The UAE's legal architecture is already moving toward technologically specialised adjudication.

The DIFC Digital Economy Court expressly accommodates AI and other digital-economy disputes, while its procedural framework allows technology-enabled proceedings and digital mechanisms.

The emerging direction is therefore not necessarily to give every AI agent independent legal personality.

A more practical model is:

Human/Corporate Legal Responsibility + Technological Attribution + Risk Allocation + Expert Evidence + Digital Courts

34. Key Legal Principles

Principle 1

AI autonomy does not automatically remove human or corporate responsibility.

Principle 2

The developer, operator, owner and supervisor may have different legal positions.

Principle 3

Several actors may contribute to the same damage.

Principle 4

Article 253 is central where multiple persons are responsible.

Principle 5

Causation must be separated from technical system behaviour.

Principle 6

The person controlling or supervising a system may have a different position from its developer.

Principle 7

Contractual allocation of risk does not automatically exclude mandatory civil liability.

Principle 8

AI-generated evidence remains subject to human verification and procedural responsibility.

Principle 9

DIFC AI/digital-economy decisions are important UAE-based authorities but are not automatically binding precedents for mainland UAE courts.

Principle 10

The absence of a specific multi-agent liability statute does not mean that harmful AI conduct is legally unregulated.

35. Exam-Oriented Case List

36. Conclusion

Multi-agent system liability in UAE civil law is best analysed through existing principles of attribution, causation, agency, principal liability, multiple-tortfeasor responsibility and compensation rather than by treating each AI agent as an independent legal person.

The most important current statutory provisions are Articles 245–255, 266–267 and 271 of Federal Decree-Law No. 25 of 2025. In particular, Article 253 provides a direct framework for situations involving multiple responsible persons, while Articles 266–267 address responsibility for another's conduct and recourse.

The most important conceptual formula is:

AI autonomy ≠ legal autonomy.

A UAE court faced with a multi-agent dispute would likely need to identify who designed, controlled, deployed, supervised, authorised, or benefited from the system; what each actor's duty was; how each actor contributed to the damage; and whether responsibility should be apportioned or imposed jointly and severally.

Because reported UAE decisions specifically resolving multi-agent AI liability remain limited, the existing agency, tort, digital-economy and AI-procedure cases should be treated as analogical authorities rather than as a settled multi-agent doctrine.