What Is Agentic AI?
Artificial intelligence is no longer a system that merely produces answers in response to instructions. As society moves deeper into the digital era, the application of AI in professional work has become increasingly central. Contemporary working life demands speed and efficiency. Without AI, much of today’s output would still require a disproportionate amount of human energy and time, not only for substantive intellectual tasks, but more often for non-substantive processes such as administrative coordination, data organisation, and repetitive workflows.
By 2024, this reality became especially apparent. AI applications were no longer confined to general-purpose tools. Early systems, such as search-oriented conversational platforms including ChatGPT and Gemini, or research-focused tools such as NotebookLM, represented only the initial phase of AI adoption. As industry expectations grew, developers began designing systems with greater specificity and autonomy. They were designed to respond to goals or objectives, with the autonomous ability to execute sequences of actions independently, repeatedly analyse their output to see if it answers the objectives and adjust their strategies based on intermediate outcomes. One example of agentic AI is Newgen’s Agentic Credit Decisioning Engine. According to its own website, they guarantee the service to autonomously orchestrate decision workflows, adapt in real time, and ensure explainability and audit readiness. Its purpose is to assist lenders in delivering instant and compliant credit decisions across a range of financial products, including retail loans and complex corporate loans, while remaining aligned with applicable local regulatory guidelines.
This shift marks what is now described as agentic AI, for it is fundamentally different from the AI-generative systems. Rather than operating as a passive instrument awaiting human direction at every stage, agentic AI possesses the capacity to plan, act, and adapt with minimal ongoing human intervention. Its decision-making process is not based on a single prompt response, but works dynamically as the system assesses its own limitations within the objectives assigned, determines what information is missing, and decides how best to acquire it.
In brief, agentic AI is capable of connecting with external resources through its own internal logic and statistical assessment whenever it determines that such engagement is necessary to achieve the goal initially set. With its high-level analytical capacity and its ability to generate subtasks to gather information from external databases in order to prevent a ‘hallucination’ on its part, agentic AI responds to complex objectives differently from conventional generative AI systems. Where a particular task requires access to industry-specific or authoritative data before a meaningful output can be produced, agentic AI does not resort to constructing non-existent information merely to provide an answer. Instead, it actively engages with available online sources, retrieves the relevant data, and synthesises that information to arrive at an output that is grounded in verifiable inputs.
Agentic AI, Data, and Synthetic Outputs
Academic discourse has begun to grapple with the nature of data produced by agentic AI. According to the academic writing of Andrew Dang in his article titled Derivative Data: Rethinking Market Definitions in the Age of Generative AI, data generated through agentic AI processes is described as “synthetic data”, which refers to artificially generated information that mimics real-world data, allowing systems to operate and improve without constant reliance on raw human-generated inputs.
The application of agentic AI involves transforming text into formats suitable for training foundation models. Foundation model providers have adopted this approach to streamline data curation, reduce costs, and accelerate system development. Through the use of customised interfaces, this design choice also lowers the barrier for users, as they not required to understand or work with complex code in order to retrieve information. Instead, users need only articulate a specific goal, allowing the agentic AI to interpret that objective and independently perform the necessary processes required to deliver the intended output.
In this sense, agentic AI functions as an intermediary or agent acting on behalf of the user. It responds to informational needs by navigating across vast networks of data, whether publicly available or otherwise, in order to achieve the desired objective. The system continuously evaluates its own outputs to identify potential inaccuracies, vulnerabilities, or opportunities for optimisation. Where an initial output is assessed as insufficiently aligned with the stated goal, the system autonomously reconstructs its approach, generates alternative lines of inquiry, and conducts further searches. This cycle of analysis and retrieval may repeat multiple times until the system determines that the output sufficiently satisfies the objective originally set.
Unfortunate Downside of the Agentic AI
Despite the progressive and efficient capabilities demonstrated by agentic AI, the system is not without its shortcomings. These shortcomings, however, do not primarily relate to performance or technical competence. Rather, they arise from deeper concerns surrounding morality and ethics. In particular, the function of agentic AI as an intermediary or agent operating on behalf of the user is, by its very nature, largely unsupervised. This autonomy makes the system inherently susceptible to encroaching upon copyright-protected sources in the course of producing accurate and reliable outcomes.
To illustrate, consider a scenario in which the objective is to develop a high-tech application to streamline employee administration within a company, such as digital check-ins, leave applications, or internal approvals. For the agentic AI to function effectively, it would first need to understand the company’s internal rules and operational structure. This may require access to employee guides, policy manuals, or handbooks in order to identify common employment practices and organisational norms. Only after acquiring such contextual understanding can the system proceed to subsequent stages, such as assessing the company’s location, regulatory environment, or workforce profile, before determining the suitability and configuration of the proposed application.
A similar approach can be observed in commercial deployments. Newgen, in securing an effective commercial deployment of its agentic AI system while ensuring compliance with country-specific regulatory requirements, has already incorporates access to frameworks such as Bank Negara Malaysia’s responsible lending guidelines. While such access enhances accuracy and regulatory alignment, it also highlights how agentic AI inevitably relies on pre-existing structured information, some of which may be protected or restricted in nature.
What makes this issue particularly difficult to address is that, unlike conventional generative AI, any encroachment by agentic AI may not manifest in the final output in an obvious or direct manner. The potential infringement may instead occur during the internal, interlayered process of repeated data analysis and information retrieval. This process may involve numerous cycles of searching, analysing, and synthesising data before a final output is produced. By the time the output is generated, it may bear little or no resemblance to the original source that was accessed during the process.
As a result, proving copyright infringement in such circumstances becomes exceptionally challenging. A claimant would face significant difficulty in establishing that their protected material was used at any particular stage, whether for training purposes or merely as part of an autonomous data retrieval process undertaken to fulfil a specific goal. The complexity and opacity of agentic AI’s internal processes raise serious evidentiary obstacles, particularly where the alleged infringement does not result in a recognisably similar output.
Existing litigation already provides a glimpse into this tension. In 2024, Warner Music Group initiated civil proceedings against the AI music platforms Suno and Udio, alleging that copyrighted songs were used as training materials to develop systems capable of generating music that competes directly with human artists. The significance of this case lies in its recognition that advanced AI systems require extensive training in order to deliver seamless and commercially viable services. Such training inevitably depends on source materials. The controversy arises when those materials are obtained and used without permission merely because they are accessible online, notwithstanding their protected status under copyright law.
A similar issue arose in Sarah Anderson v Stability AI Ltd. case, where a group of visual artists claimed that Stability AI, along with other systems such as Midjourney, had used their copyrighted artworks without consent to train AI image generation models. In response, the defendants argued that there was no substantial similarity between the original works and the generated outputs, and that the outputs were sufficiently transformative. The court ultimately set aside most of the plaintiffs’ claims, primarily on the basis that the allegations regarding training and use of copyrighted works were not pleaded with sufficient specificity at that stage to establish direct infringement.
These cases already demonstrate the difficulty of addressing copyright infringement in the context of generative AI. When the same issues are imposed onto the use of agentic AI, the problem becomes even more pronounced. The immediate question that arises is how a claimant could possibly prove that, within an autonomous process of data research and layered analysis, a particular copyrighted work was accessed or relied upon? As agentic AI conducts multiple rounds of independent inquiry, synthesis, and self-correction, establishing a clear and traceable thread of copying obviously has become increasingly difficult.
The Malaysian Perspective and Data Protection
Nevertheless, despite the foreseeable difficulties ahead, there remains a degree of reassurance in the existing legal framework in Malaysia, which may provide meaningful protection for copyrighted materials should these issues arise in the near future.
Under section 36 of the Copyright Act 1987, copyright infringement occurs where a protected work is used by a person who is not the owner of the copyright, and who has not obtained the consent of the copyright owner to use that work. This provision prohibits third parties from exploiting copyrighted works in the course of business, or in any manner that may prejudice the legitimate interests of the copyright owner. In principle, this prohibition would extend to the use of copyrighted materials by agentic AI systems, if it can be established that such works were used as training inputs or analysed in the course of producing commercially valuable outputs without authorisation.
This position becomes particularly relevant when agentic AI systems operate autonomously to retrieve, analyse, and synthesise information from multiple copyrighted sources. While such systems are designed to enhance efficiency and accuracy, their reliance on pre-existing data means that they may inadvertently incorporate protected works into their internal processes. If such use can be positively identified and shown to fall within the scope of section 36, the operation of agentic AI would not be protected from liability merely because the process is automated or technologically complex.
Beyond copyright, agentic AI must also be assessed through the lens of data protection. Although agentic AI may appear, at a surface level, to function similarly to advanced search engines by enabling users to locate information more quickly and efficiently, it does not operate outside the boundaries of Malaysian law. In particular, section 5 of the Personal Data Protection Act 2010 requires that the processing of personal data in commercial transactions comply with prescribed principles governing the collection, use, disclosure, and security of such data. Failure to comply with these obligations carries significant consequences. Under the current legislative framework, non-compliance may result in a fine of up to RM1,000,000.00 and or imprisonment for a term of up to three years. In theory, these requirements should preclude agentic AI systems from freely accessing confidential or personal information embedded within online commercial platforms, such as membership databases or internal organisational systems, as such information would have been very valuable for analytical or optimisation purposes.
As of today, discussions surrounding agentic AI and its associated risks are no longer confined to academic circles. Global industry players have begun developing governance frameworks aimed at mitigating these concerns. Various organisations have proposed structured approaches to managing agentic AI risks, with emphasis placed on transparency, accountability, and data governance. For instance, professional services firms such as Mayer Brown have highlighted the importance of ensuring explainability in AI systems, implementing strong data privacy safeguards, and adopting responsible data practices, including the use of anonymised and ethically sourced datasets, strict access controls, and compliance with applicable data protection regulations.
With all the frameworks introduced, they reflect a growing global and national recognition that while agentic AI offers significant operational and economic advantages, its deployment must remain aligned with existing legal and regulatory structures. In the Malaysian context, this alignment will be critical in determining whether agentic AI can be integrated into commercial and professional environments without undermining copyright protection and personal data privacy.
For further information, please contact:
Ahmad Hafiz Zubir, Azmi & Associates
hafiz.zubir@azmilaw.com
