OpenClaw and the new era of AI conductors: between automation, legal liability and the limits of application of the AI Act
Abstract: The article proposes an interpretation of the OpenClaw phenomenon, characterising it not as an artificial intelligence system in itself, but as an "AI conductor": an orchestration infrastructure that coordinates existing models and transforms intentions formulated in natural language into legally relevant actions executed via APIs on the user's applications. The analysis traces the evolution of automation from RPA to modern orchestration layers, clarifying the distinctions between an AI model, an agent and execution infrastructure. From a legal perspective, the effects produced through automation are examined, together with the differences from smart contracts and the positioning of AI conductors in relation to Regulamentul (UE) 2024/1689 (AI Act). The conclusion is that these systems do not generate will, but rather execute it technically, shifting the centre of gravity of liability from the technology towards the user who configures and authorises the mechanism. AI conductors thus mark a stage in which conversation becomes action, and automation becomes the legal expression of human will.
Keywords: OpenClaw, AI conductor, orchestration layer, digital automation, AI models, ChatGPT, Claude, OpenAI, AI agents, API integration, open source software, legal liability in the use of AI, autonomous automation, smart contracts, AI Act, Regulamentul (UE) 2024/1689 (AI Act), technological compliance, digital responsibility, AI ecosystems, decision automation, legal tech
A. What is OpenClaw?
OpenClaw is a self-hosted, open source "AI conductor" (it can be downloaded free of charge from an open source) designed to execute the user's instructions through already existing artificial intelligence models, such as GPT or Claude. In essence, the system enables the automatic performance of various actions, from organising a calendar and managing correspondence to making online purchases or interacting directly with the user's applications/programs. Therefore, OpenClaw is not an artificial intelligence in itself. It does not "think" on its own, does not learn from existential experiences and does not dream of a robot revolution. It is, rather, a conductor.
Unlike a chatbot, which functions exclusively as a conversational interface, OpenClaw runs locally, on the user's own infrastructure (e.g. on the user's personal laptop or server). This characteristic shifts technical and operational control from the service provider towards the user. The system does not operate in a centralised cloud and does not belong to a commercial entity that intermediates the interaction, but rather functions as a software program installed and configured individually by the user (an essential aspect from the perspective of liability, which we will address below).
OpenClaw operates by connecting to the user's applications through APIs (Application Programming Interfaces), i.e. technical interfaces that enable communication between different computer programs. In this way, OpenClaw can interact with messaging services such as WhatsApp, Telegram or other digital platforms. Through these connections, dirijorul it receives the user's commands and forwards them to external AI models, using the API key and the subscriptions/accounts held by the user themselves (more precisely, in order to operate commands in the messaging area, OpenClaw must be configured with authorised access to the user's services, such as ChatGPT (via API key) and messaging applications such as WhatsApp, through the technical interfaces made available by these platforms).
In other words, OpenClaw is neither an artificial intelligence model nor an autonomous agent in itself. The system does not develop its own artificial intelligence and does not have independent learning mechanisms, representing, in essence, an open source software program made up of publicly accessible source code. From a technical perspective, its role is to coordinate and integrate already existing AI models, giving operational meaning to commands formulated in natural language and enabling their execution beyond simple chat interfaces, by producing concrete effects in the digital applications and services connected by the user.
B. The evolution of automation and AI integration
To understand the emergence of systems such as OpenClaw, we must go back in time a little. Automation did not appear suddenly, but evolved in stages.
In the initial stage (the era of fixed rules, of "disciplined robots"), automation had a rigid character, based exclusively on pre-established rules. Programs executed only the instructions previously entered by the programmer, without the capacity to interpret context or adapt to new situations. The solutions used relied on techniques such as screen scraping, macros or workflow automation, through which software mechanically reproduced the actions of a human user, such as pressing keys, filling in fields or transferring data between applications. In practice, automation represented a combination of hardcoding (entering instructions into the program, so-called classic programming) and mimicking human behaviour, with the user being required to "teach" the robot each action explicitly. From opening a new tab by pressing the plus button, to the way the cursor was moved across the screen, every movement had to be defined in advance, step by step, so that the process could later be reproduced by the software.
This logic was reinforced with the emergence of RPA (Robotic Process Automation) platforms, popularised globally by companies such as UiPath. RPA standardised the automation of graphic interfaces, enabling companies to automate repetitive tasks such as invoice processing, document validation or data entry into internal systems. However, these solutions operated exclusively on the basis of fixed rules: the robot would execute exactly what it had been programmed to do, without any capacity to interpret context.
The second stage (probabilistic automation) began as advances in machine learning enabled software to move beyond the limits of classic, fixed-rule-based automation. Systems became capable of identifying patterns, automatically classifying documents and extracting relevant information from text, without every scenario needing to be programmed/defined by the user. Automation no longer relied solely on express instructions, but began to draw on probability analysis as well. The real shift came with the emergence of generative language models (GLMs), particularly after 2022, when models such as GPT demonstrated the ability to understand and interpret instructions formulated in natural language. From that point on, the user no longer had to describe every step of a process; it was enough to state the objective or desired outcome, with the system taking on the task of structuring the steps needed to achieve it.
Thus, instead of:
"open email → search attachment → extract data → create offer",
the user can formulate a general instruction:
"analyse the client's history from emails and prepare a commercial offer".
The third and most recent stage is represented by the emergence of AI orchestrators. Although artificial intelligence models have become capable of understanding complex requests and generating sophisticated responses, one important limitation has remained evident: they cannot act directly on IT systems and do not have implicit access to applications, files or digital accounts.
A language model can draft an email, but it cannot send it. It can explain the mechanics of a booking, but it cannot make one. This is the context in which the role of an orchestrator such as OpenClaw arises, introducing an intermediary layer between the user, the AI model and the applications being used. The orchestrator takes the command formulated by the user, passes it to the AI model for interpretation, converts the generated response into a set of concrete technical actions and then executes them via APIs or other connected software tools. In this way, the AI model carries out the analysis and formulates the decision, while the orchestrator ensures its effective implementation within the user's digital environment.
C. Key definitions in the context of OpenClaw
To properly understand how an AI orchestrator such as OpenClaw functions, it is useful to clarify the main components involved in this architecture.
AI Model: An AI Model is a program trained on a data set to recognise certain patterns or make certain decisions without further human intervention, applying various algorithms to input data in order to produce the results for which it was designed. Relevant examples of such models are ChatGPT, Claude or Gemini. Essentially, the model is the component that "thinks" in a technical sense: it interprets semantic requests and produces a response. The AI orchestrator is not itself an artificial intelligence model, but functions as an intermediary technical layer that uses AI models and instructs them within an operational framework.
Wrapper: A wrapper is an application or interface built on top of an existing artificial intelligence model, in order to make it usable in a specific context. For example, an application that uses GPT for the automatic drafting of contracts is a wrapper: the model remains the same, but the interface and functionality are adapted to a specific purpose. In this context, OpenClaw also uses such mechanisms to integrate existing AI models and make them usable in a practical way.
AI Agent: An AI Agent is the software component capable of planning and executing successive steps in order to carry out a task. If the AI model provides the analysis (establishing what should be done), the agent is the one that organises the execution. Thus, OpenClaw is not the agent itself, but the technical framework within which the agent operates.
The orchestration layer: This represents the central element of the orchestrator. The orchestration layer coordinates the relationship between the user, the AI models and the connected external applications. In the case of OpenClaw, this orchestrator (i) receives the user's commands, (ii) manages the context, (iii) calls the AI model, (iv) translates the response into technical actions and (v) executes these actions via APIs.
Terminological clarification: the use of the term "AI orchestrator" in this article represents the author's own conceptual proposal, put forward for explanatory purposes, given that Romanian legal and technical literature does not yet establish a settled name for this type of software infrastructure.
D. Automation through OpenClaw versus the concept of a "smart contract"
The concept of a "smart contract" originates from blockchain technology and denotes computer code that executes automatically once certain pre-established conditions are met. For example, a payment may be released automatically upon confirmation of the delivery of a good.
At first glance, OpenClaw appears to function similarly: it can automatically execute actions such as placing an order or initiating a payment when the user formulates a request. The essential difference, however, is structural.
OpenClaw does not operate on a blockchain, does not create an autonomous contractual mechanism and does not generate a legal effect through its mere technological existence. It is an automation tool that carries out actions within contractual relationships already existing between the user and third parties. If a smart contract is “the contract in code”, OpenClaw is “the user's will in execution”.
To illustrate how it works, let us assume that a user makes the following request: “Buy my favourite pizza.” In this scenario, OpenClaw can, through available APIs, access a home delivery application, place the order according to the user's saved preferences and automatically make the payment.
This operational flow bears similarities to a smart contract, but there are essential differences from a legal perspective. Unlike a smart contract verifiable on a blockchain, the automation carried out by OpenClaw is not recorded in a distributed, decentralised ledger, does not benefit from the immutability of the blockchain, and does not have any special legal value conferred by the technological infrastructure.
It is important to emphasise that no “contract” in the legal sense is concluded between the user and OpenClaw (since OpenClaw does not represent a legal person or a service provider, but merely open source software code, with no entity offering the service, setting terms and conditions, or assuming contractual obligations). In reality, the user interacts with the usual services and providers, for example, platforms such as BoltFood, Glovo, or their bank. OpenClaw acts solely as a technical intermediary that facilitates and automates these interactions, based on instructions received from the user.
For the sake of clarity, in the following section we have opted for a question-and-answer (Q&A) structure. This choice of format is not accidental. The field under discussion combines new technical concepts, and much of the confusion arises precisely from the overlap of terminology from the technological sphere, as follows:
Q: „Is OpenClaw a smart contract?”
A: No. A smart contract runs on a blockchain and automatically executes its own conditions. OpenClaw runs locally, as open source software, and carries out the user's instructions within third-party applications.
Q: „Can OpenClaw conclude contracts on the user's behalf?”
A: Not directly, because OpenClaw has no legal personality. However, if the user configures the system to automatically perform actions that amount to accepting a contractual offer (for example, placing an online order or confirming a booking), the legal effects will arise in the user's estate. From a legal standpoint, we consider that the automated action is equivalent to an action carried out through a technical instrument chosen by the user.
Q: „Is there a contract between the user and OpenClaw?”
A: No. Using OpenClaw does not involve concluding a contract, since it does not represent a service provider and there is no legal entity offering the software in exchange for consideration. OpenClaw is open source code made available to the public, and its use is comparable to a situation in which the user were to create their own computer program to automate their own tasks.
E. Legal liability for actions carried out through OpenClaw
As mentioned above, OpenClaw has no legal personality, is not a service provider, and there is no entity that contracts with the user. It is open source code, offered as such, without any warranty as to its functioning or fitness for a particular purpose. In other words, there is no “someone” who is liable for what the system does. There is only the user who installs it, configures it, and grants it access to their own accounts.
From a legal perspective, the relationship between the user and OpenClaw can be conceptualised in terms of agency or representation. The user essentially becomes the principal, that is, the person granting authority, by giving the program access to their own accounts, authentication credentials, and digital tools. OpenClaw acts as a technical executor of the user's will, similar to how a personal assistant would carry out tasks delegated by an employer. Thus, when OpenClaw places an order on an e-commerce platform, makes a payment through a banking application, or sends a message through the user's email account, all these actions are carried out in the name and on behalf of the user, using their credentials and authorisations. From a legal standpoint, it is as though the user had performed those actions themselves, the only difference being the technical means of execution.
The legal consequence is that OpenClaw functions solely as a technical tool through which the user carries out their own decisions and commitments, and liability for the effects produced rests with the user who chose to configure and use it. For clarity, the authors of the code are likewise not liable for its use, since the software is offered as an open source project, without assuming any contractual relationship with users.
Legal responsibilities are governed exclusively by the contracts and terms of the external service providers with which the user interacts through OpenClaw. These may include home delivery platforms (such as BoltFood or Glovo), online shops, banking institutions, communication services, and any other integrated digital applications or services.
The user nonetheless remains bound to comply with the terms and conditions of the third-party platforms with which they interact through automation, including provisions relating to the use of automated tools or software programs acting on the user's behalf. The fact that actions are carried out through an AI agent does not alter the contractual obligations already assumed towards those services. In practice, some digital platforms expressly provide for prohibitions or restrictions on the use of automated scripts, software bots or automation mechanisms, particularly where these may affect the functioning of the service, account security or the correctness of commercial interactions. Using a system such as OpenClaw in breach of such rules may be classified as a contractual non-compliance.
Example: the automated bookings scenario
To better understand the legal implications of using an AI agent, it is useful to examine a practical scenario, inspired by situations already discussed in legal literature on the automation of online orders.
Let us assume that a user configures OpenClaw to automatically manage certain aspects of their personal and professional life. Among the permissions granted is access to an online booking platform for wellness and leisure services, including saved payment methods and user preferences. At one point, the user gives a vague instruction: “schedule me a few relaxation sessions for the next few weeks, I need to unwind”. Based on prior history and preferences saved in the system (therapeutic massages, spa sessions, yoga), the program interprets the request and automatically generates multiple bookings at various wellness centres across the city. To ensure availability and avoid repeated manual confirmation requests, OpenClaw books complete packages at seven different centres, spread over four weeks. The total value of the bookings reaches approximately EUR 3,200, representing 28 individually pre-paid sessions.
After receiving the confirmation emails, the user discovers that the program has booked an excessive volume of services, namely one or two daily visits to different wellness centres, incompatible with their schedule. In attempting to cancel most of the bookings, the user is met with refusal from the service providers, who invoke the cancellation policies applicable to the bookings made.
From a legal perspective, such a position may, in principle, be justified by reference to Article 16(l) of Directive 2011/83/EU on consumer rights, according to which the 14-day right of withdrawal does not apply to the provision of accommodation, transport, catering or leisure activities where the contract provides for a specific date or period of performance. Wellness services scheduled for specific time slots generally fall within this category, since the provider reserves capacity and resources for a specific point in time.
Consequently, the consumer does not benefit from the statutory right of withdrawal once the booking has been made, and cancellation of the services remains governed exclusively by the contractual terms established by the provider.
The fact that the bookings were made through an automated system does not alter the legal situation. From a consumer law perspective, the orders are deemed to have been placed by the account holder, with automation representing merely the technical means by which the user's instructions were carried out.
Therefore, from a legal standpoint, bookings made through OpenClaw are considered valid and binding on the user, as they cumulatively satisfy the following conditions:
First element: the bookings were made through the user's personal account, using their authentication credentials on the booking platform and benefiting from the access rights granted to them.
Second element: payments were authorised through payment methods previously saved by the user in their profile, whether a bank card or a digital account.
Third element: the system acted within the limits of the mandate conferred and the permissions granted by the user.
This example illustrates a fundamental principle of digital law: automation does not eliminate or diminish the user's personal legal responsibility, but merely changes the technical means by which their will is expressed. When a program, whether an AI-based digital assistant, an automated scripting algorithm or another type of automated system, is granted access to personal accounts, service platforms and payment methods, any action carried out within the scope of those permissions produces full and direct legal effects for the user.
In the case of OpenClaw, this responsibility falls entirely on the user, precisely because of the absence of a protective contractual framework. There is no provider assuming duties of diligence, offering guarantees regarding the system's functioning, or ensuring mechanisms of transparency and traceability of decisions. The user cannot request technical explanations, nor can they invoke the liability of an economic operator for any errors or unforeseen outcomes of the program.
From this perspective, it is advisable for a diligent user to explicitly define their own operating parameters for the automation, including clear operational and financial limits, and to ensure that the software architecture used allows for the verification and effective control of the actions performed. In practice, this involves setting value ceilings for transactions, introducing manual confirmation mechanisms above certain thresholds, configuring restrictions for sensitive categories of services, and continuously monitoring the system's activity before granting extended access to accounts or financial resources.
This caution is also justified by the fact that the artificial intelligence systems with which the dirijor interacts do not offer extensive legal liability towards users either. For instance, ChatGPT's terms and conditions provide for the limitation of OpenAI's liability for indirect damages, loss of profit, use or data, with total aggregate liability capped at the greater of the amount paid by the user for the service in the last 12 months and USD 100.
Similarly, Claude's terms of service state that the services are provided “as is” and “as available”, without warranties as to accuracy or fitness for a particular purpose, and the provider's total liability is limited to the greater of the amount paid for use of the service in the last 6 months and EUR 100, except where the law prohibits such a limitation (such as fraud or bodily injury).
Consequently, even when the user interacts directly with commercial AI model providers, the possibility of recovering damages is significantly limited contractually. When these models are integrated into an open-source dirijor such as OpenClaw, responsibility for configuration, oversight and control falls entirely on the user, who effectively becomes responsible for the way in which the automation produces legal and financial effects.
The adoption of the AI Act marks the European Union's first attempt to systemically regulate the use of AI systems through a risk-based approach. The Regulation classifies artificial intelligence systems into several risk categories, establishing different obligations depending on the potential impact on fundamental rights and society. Within this framework, the question arises: where does an AI dirijor such as OpenClaw fit in?
The AI Act primarily regulates AI systems as such, namely software capable of generating outputs such as predictions, recommendations or decisions that influence the environment in which they operate. The legislative focus falls on providers who develop or place AI models on the market, particularly on systems used in sensitive domains: biometric identification, critical infrastructure, education, employment, financial services or the administration of justice. OpenClaw, however, does not fall within these definitions. The system does not contain an AI model of its own, does not train algorithms and does not generate artificial intelligence independently. Its role is that of a technical coordination infrastructure, an orchestration layer that connects existing AI models with the user's applications and accounts.
We therefore take the view that OpenClaw does not fall within the scope of the AI Act, as it does not meet the legal definition of an “AI system”. Under Regulation (EU) 2024/1689, an “AI system” means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.
As set out above, OpenClaw does not “deduce” nor does it “generates” such outputs through its own algorithms and does not have an AI model. It functions as a technical orchestration layer: it takes the user's command, calls an external AI model (GPT, Claude, etc.) via API, and executes the resulting actions in the connected applications. Consequently, if there is an “AI system” relevant to the AI Act in this ecosystem, it is the AI model used (GPT, Claude), not the OpenClaw dirijor.
G. Beyond OpenClaw: what these “dirijori” mean for the law and for society
The emergence of AI dirijori, such as OpenClaw, marks a subtle but profound shift in the way the law will need to understand the relationship between will and action. Until recently, the legal debate on artificial intelligence was focused almost exclusively on generated content: if a model produces erroneous, discriminatory or misleading information, who is liable for that outcome? AI dirijori, however, shift the centre of gravity of the analysis. We are no longer discussing merely words and semantics, but facts and execution. We are no longer discussing merely recommendations, but automatically executed actions: payments initiated, orders placed, bookings confirmed, contracts accepted. In this context, technology is no longer merely a tool for intellectual assistance, but becomes a mechanism for the execution of legal will.
This transformation has an essential consequence: responsibility does not disappear, but is reconfigured. The AI dirijor is an amplifier of the user's decisions. Once the user configures the system, grants it access to accounts, sets its operating parameters and allows it to perform actions on their behalf, the technology becomes an extension of their will. And the law, traditionally, treats extensions of the will as being part of it. It is not relevant that execution is automatic; what is relevant is that authorisation is human.
For the legal world, the challenge lies not only in the technical qualification of these systems, but in understanding how they modify standards of diligence. If, until now, liability was analysed through the lens of acts actually performed by a person, in the era of AI conductors it will also become relevant how that person configured their own automation mechanisms. Legal diligence risks implicitly including technical diligence: the setting of limits, control mechanisms, monitoring and prevention. Not because the law explicitly requires this today, but because technological reality makes it inevitable.
On a broader level, AI conductors bring to the fore a fundamental question about the nature of responsibility in the digital society. The instinctive tendency is to attribute mistakes to the “system”, the algorithm, the code. But code does not decide on its own to gain access to bank accounts, does not set financial ceilings on its own and does not grant itself permissions (at least not yet). Behind every automation there is a human configuration choice and a decision to delegate. Technology accelerates execution, but does not create will. And law ultimately judges will and its effects.
From this perspective, AI conductors do not usher in a world in which liability dissolves into technological complexity, but rather one in which the link between decision and consequence becomes more direct and more visible. If intentions can be turned almost instantly into legal actions, then awareness of consequences must be just as swift and clear. For legal professionals, this is not merely a challenge of normative interpretation, but also one of conceptual adaptation: understanding that the future will not belong solely to those who read contracts, but also to those who understand the technological flows that execute them.
Sources used in drafting this article: 1. Paolo Perazzo, (OpenClaw Architecture) accessed 12 February 2026; 2. Scientific American, (OpenClaw is an open-source AI agent that runs your computer) accessed 12 February 2026; 3. William Lawyer, (OpenClaw: The AI Agent Everyone’s Talking About and Why Your IT Team Should Pay Attention) accessed 12 February 2026; 4. Amanda Silberling, (After all the hype, some AI experts don’t think OpenClaw is all that exciting) accessed 12 February 2026; 5. ADVERSA, (OpenClaw security 101: Vulnerabilities & hardening (2026)) accessed 12 February 2026.
The analysis presented in this article represents solely an interpretation of technologies that are at an early stage of development and regulation. The field of AI conductors and autonomous agents is evolving rapidly, and the legal qualifications proposed reflect the current state of the regulatory framework and technical understanding available at the time of writing. Given the relatively novel nature of these technologies, it is to be expected that both national and European Union legislation, as well as the practice of authorities and courts, will progressively clarify matters relating to legal liability, the applicability of existing regulations and the limits on the use of such systems. Consequently, the conclusions set out herein should be regarded as points for legal reflection rather than definitive solutions, with future legislative and case-law developments expected to provide more precise answers to the questions raised by these emerging technologies.
Lawyer Victor Buju, Buju Stanciu & Asociații Lawyer Mihai Alexandru Alexandrescu, Buju Stanciu & Asociații