In this blog series, we explore various ways how large language models (LLMs) can support different levels of enterprise automation. In part 1, we began with the emerging methods of connecting LLMs to internal data and the notion of ambient intelligence. Now, we’ll progress towards more advanced cases that we see as having a growing impact in the future, explaining the main drivers behind each level of automation.
As we delve deeper into the potential of large language models for enterprise automation, we arrive at the concept of intelligent agents. These are autonomous entities capable of observing their environment and directing their actions towards specific goals. The real question then becomes, can we transform large language models into such goal-oriented intelligent agents?
The transformation of LLMs into intelligent agents primarily hinges on two capabilities: self-reflection and code generation.
Self-reflection refers to the model’s capacity to evaluate its output, using it as context or input for subsequent responses. This iterative process allows the model to assess whether its output aligns with the set goal and adjust its responses accordingly. The LLM essentially becomes a self-regulating system, continuously refining its output based on self-analysis and goal alignment.
The capability to generate code, on the other hand, empowers the model to leverage a range of software tools effectively. By generating code that interfaces with various tools or APIs, the model can overcome potential limitations while interacting within digital environments like web browsers.
With these capabilities, the AI system evolves from a passive assistant or ‘copilot’ to an intelligent agent. The system shifts from merely responding to queries towards proactively seeking information and performing tasks that align with its goals.
This shift has profound implications for enterprise automation. It represents a move towards artificial intelligence (AI) systems that can initiate interactions, undertake tasks and even delegate responsibilities. Such proactive intelligent agents can dramatically improve efficiency, streamline processes and significantly reduce the human workload.
In the enterprise context, intelligent agents can be applied in various ways. They can manage and optimise workflows, auto-generate reports and even conduct predictive analyses to aid strategic planning. They can handle routine administrative tasks, thus freeing up people’s time to do more complex tasks.
Moreover, intelligent agents can function as personal assistants, scheduling meetings, managing communications and even proactively providing information or reminders based on context and user preferences. They can also play pivotal roles in customer service, not only by responding to customer queries but also by anticipating and addressing potential customer issues proactively.
The increased capability and autonomy of intelligent agents inevitably come with associated costs, particularly those related to API calls and computations. A viable solution to manage these costs is implementing a hierarchy of agents with varying skill levels. In such a system, lower-skilled agents can handle more straightforward tasks, thus conserving resources. More complex tasks or issues can be escalated to higher-skilled agents. This arrangement optimises the cost-effectiveness of the system by ensuring that resources are only expended when necessary.
As AI systems become increasingly autonomous, a critical challenge emerges: alignment. We need to ensure that the goals of these intelligent agents align with our own. Misalignment can lead to undesirable outcomes, even when the agent is perfectly optimising according to its goal.
In our discussion on how to use large language models in automating enterprises so far, we’ve mainly focused on knowledge work and digital interactions. However, another domain, which is often underrepresented in these discussions, is the world of physical interaction and movement – the domain of robotics. Although it is considered a field of slower progress compared to AI, the potential influence and advancements of LLMs in robotics cannot be underestimated.
The rise of intelligent agents extends beyond virtual environments into the real world, interfacing with APIs that allow them to impact our physical environment. However, there’s another step where the agent doesn’t merely interface with our environment but traverses and manipulates it as a fully embodied entity.
LLMs were previously confined to the realm of processing and generating text. However, advancements over the years have led to successful experiments in linking these models to robotic systems. Consequently, we are beginning to witness a convergence in the progress of large language models and robotics.
Many enterprises are already exploring the benefits of embodied intelligence with humanoid robots under development. These intelligent systems can perform physical tasks, from organising stock in a warehouse, to carrying out maintenance tasks in industrial environments or assisting employees in their daily operations. The impact of these systems within enterprise environments could be far more extensive than anticipated.
One of the most fascinating aspects of embodied intelligence is its potential to redefine how we perceive intelligent systems. As these AI models begin to move, interact and respond in the physical space, they become more tangible and relatable entities. This tangibility increases our capacity to trust and collaborate with these systems, thus strengthening the human-AI synergy within enterprises.
However, the advent of embodied intelligence also amplifies the importance of safety considerations. As these intelligent systems begin to interact with the physical world, the stakes for potential errors or mishaps significantly increase. For instance, a system that incorrectly interprets an instruction could cause damage to physical property or, worse, harm to human lives. Therefore, ensuring that these systems operate safely within their designated parameters becomes a paramount concern.
Regional Head of Data, AdriaticBoris is a machine learning expert with over 15 years of experience across various industries. As Regional Head of Data, he is dedicated to growing our data capabilities in the Adriatic region. Throughout his career, Boris has co-founded and led a data science consultancy, established an AI department in an IT services company, and advised governments on AI strategy and standards. He is passionate about innovation and has a good track record of spotting key technology trends early on. Boris also enjoys sharing his insights through speaking engagements, advising start-ups, and mentoring. Recently, his main areas of interest include large language models and other generative AI technologies.
13 November 2023
Delving Deeper Into Generative AI: Unlocking Benefits and Opportunities
07 November 2023
Retrieval Augmented Generation: Combining LLMs, Task-chaining and Vector Databases
19 September 2023
The Rise of Vector Databases
20 July 2023
Large Language Models Automating the Enterprise – Part 1
01 June 2023
Challenges for Adopting AI Systems in Software Development