From RPA to AI Agents: the evolution of software robots towards autonomous decision making

This article discusses the evolution of software based robots employed in automation-and process orchestration in the age of Generative AI. Starting from traditional Robotic Process Automation,? we? discuss GenAI-based solutions, agentic systems and? the interlinks of these technologies. Leveraging data-driven use cases, we review the advantages and disadvantages of software robots from various perspectives. The paper is intended to provide the reader with a high-level overview and guidance to navigate in the realm of automated software robots.

Robotic Process Automation

Robotic Process Automation (RPA) aims at automating slow, repetitive, time-consuming, manual and error-prone tasks. RPA software robots are commonly deployed to solve data-intensive repetitive problems previously executed in a manual way. A common characteristic of the processes that are great candidates for RPA-based automation is that they have low implementation risk due to a high level of predictability. Typically, rule-based workflows with structured data can be easily addressed by RPA. If a workflow can be covered by a set of well-defined rules, then it is worth considering for RPA based automation.?

It is important to add that RPA is a process orchestration approach: an RPA software robot can be composed of various consecutive and concurrent steps or tasks. Each step is a deterministic software module implementing a well-defined functionality. To review, a deterministic software will always provide the same output given the same input, as opposed to more recent developments in software, such as non-deterministic generative artificial intelligence, as discussed later in this paper.

An RPA software robot can be viewed as a 2-tier architecture composed of the following layers:

• Orchestration layer
• Task layer

The orchestration layer is responsible for carefully arranging individual tasks and executing them in a predefined order.? The task layer is composed of a set of independent steps each of which implements a well-defined deterministic functionality. The orchestrator layer runs an execution plan based on which the individual tasks within the task layer are triggered in a predefined order.

Consider the case of a Business Intelligence (BI) application used in management accounting to? provide decision makers with up-to-date insights. A data-intensive RPA solution can be created to provide the BI application with daily data. The RPA solution can be composed of 3 consecutive? steps: extracting data from a relational database containing financial information, calculating prescribed statistics related to business KPIs, and loading the results to a Business Intelligence (BI) application that is used for reporting purposes. As each of these steps is deterministic and well-defined, they are great candidates for RPA. A deterministic orchestrator module can be implemented to automate the process. The orchestrator can execute each of the steps in a well-defined order and handle the data transfer in transition between them. The advantage of this RPA software solution is the efficient automation of manual and data intensive activities, eliminating manual, error-prone, human workloads and thereby increasing operational efficiency. RPA bots enable the scaling of processes while increasing the execution frequency.?

Generative Artificial Intelligence

Generative AI is a newer field of Artificial Intelligence. This technology goes beyond rule-based automation and analysis of large structured datasets; it has the capability to synthesize entirely new information. Traditional AI excels in analysing big data volumes and building models to capture hidden patterns; however, it cannot synthesize new information and has no reasoning capability.

GenAI, in turn, excels at content creation and reasoning across various modalities. GenAI is non-deterministic, meaning that the same input can generate a variety of outputs. It can analyze big data volumes and synthesize new information. While traditional RPA addresses workflows that can be covered by rules and structured data, GenAI has generalization capability based on which one can address problems that cannot be covered by pre-defined rules. It can boost productivity, content scalability, and the user experience.

In a retail webshop, capturing customers’ attention requires compelling images and detailed product descriptions. Traditionally, these tasks were handled by human workers by photographing products from multiple angles and crafting descriptive summaries. Creative content generation was a time-consuming, slow and high-cost activity. However, with advancements in Generative AI and Large Language Models (LLMs), the process of creative content generation can now be automated and scaled efficiently. These technologies enable the automatic creation of diverse content formats, including text, images, and even videos, reducing the need for manual human effort while enhancing productivity and scalability.

GenAI-augmented RPA

Traditional RPA can be augmented with generative AI capabilities. As aforementioned, RPA is implemented as a 2-tier architecture: one can distinguish the orchestration layer and the layer of individual tasks. The orchestration layer encodes an execution plan which triggers the individual functional tasks in a predefined order. In RPA software robots, both of the tiers are realized as deterministic software components.

GenAI-augmented RPA extends traditional software robots by replacing some deterministic tasks with GenAI-augmented functional components. Making use of the data synthetization and reasoning capabilities of Generative AI, it provides us with the opportunity to automate more complex and dynamic processes, where traditional, deterministic automation would fail.? Situations requiring content generation, knowledge work and decision support are worth considering for GenAI-augmented RPA.

Let us evolve the Business Intelligence example discussed previously. Assume that decision makers need customizable BI reports, in which diagrams are dynamically generated in response to the user’s input questions. This is a GenBI (Generative Business Intelligence) use case. A decision maker can ask a question from the system so that it responds with diagrams and visuals most suitable to the input question. While a traditional RPA software bot is efficient in implementing a deterministic process flow of data extraction and deterministic calculation, always showing the same output diagrams, the custom visual generation necessitates a GenAI-augmented step.

AI Agents

Most recently, the evolution of Artificial Intelligence is defining its next innovation: agentic AI. Agentic systems incorporate autonomous orchestration modules that can make decisions, adapt, and take initiative in dynamic environments.

Further expanding? the capabilities of regular RPA software robots, an agentic system expands the capabilities of the same? 2 tiers:

• Orchestration layer: responsible for generating an execution plan and running the individual tasks accordingly. The execution plan is dynamically generated, and not hardcoded in the orchestrator as in traditional RPA.
• Task layer: composed of a set of deterministic and/or GenAI-based tasks. Each task can be viewed as a well-defined functional component.

An agent realizes a software robot holding autonomous process-orchestration capability. Though GenAI-augmented RPA incorporates generative Artificial Intelligence, it is employed at the task level. In contrast, agentic AI? allows the orchestrator to be realized by a generative AI module, such as a Large Language Model (LLM) with reasoning and data synthesis capability. Provided some external input, the orchestrator creates a plan and executes the individual tasks (deterministic or non-deterministic) accordingly. The execution plan is dynamically generated by the agent in response to the input query. Depending on the output returned by the individual steps, the orchestrator can update its execution plan and act accordingly. Once the orchestrator receives all the responses from the tasks, it can synthesize a response to the user.

The autonomous character of the agentic orchestrator implies that there is no pre-defined execution sequence of the tasks. In RPA and Gen-AI augmented RPA solutions, the orchestration is hardcoded by a predefined execution sequence of the individual tasks. In an agentic architecture there is no predefined execution plan: a set of tasks is defined; however, their execution is not hardcoded, but the agent can dynamically generate an execution plan using its own decision-making capabilities.

Let us consider an agent responsible for generating SQL queries. The user can feed a natural language question to the agent. It extracts database schema information, interprets the question in the context of the extracted database information and asks additional clarifying questions from the user if needed; finally, it generates a SQL query. This system is composed of 3 individual tasks within the task layer:

• Database schema (metadata) extraction:
• Input validation:
• SQL query (program code) generation:

In an agentic system the above 3 individual tasks have no predefined order. The agent must interpret the incoming question (input validation), extract the database metadata, ask the user for clarification (if needed) and generate a SQL query.

Multi-agent systems

An emerging field of agentic systems is the development of multi-agent architectures. Individual agents can be used to solve specific problems with a high level of autonomy. Provided a set of different agents addressing well-defined problems, one can combine them – as if they were composable software modules – to create a collaborative, multi-agent architecture. In multi-agent collaboration, a supervisor agent is defined to orchestrate the work across various agent-based modules within the task layer, named task agents. Each task agent can be viewed as an autonomous system solving a well-defined problem by grouping multiple elementary tasks orchestrated and executed dynamically. The supervisor agent is responsible for orchestrating the resolution of a complex problem that can be decomposed to a set of consecutive tasks taken by individual task agents. The supervisor interprets the input and creates an execution plan, then it calls the individual agents in the order it deems optimal.?

Measuring the performance of AI-augmented RPA

In our last paper, Why GenAI projects fail on the path to production, we proposed a three-tier measurement framework in order to mitigate implementation risks and quality issues due to the non-derministic nature of Generative AI. In order to manage the uncertainty of non-deterministic factors integrated into a solution by the use of GenAI, we proposed the addition of a technical KPI layer to the traditional measurement of business and operational KPIs. The same concept can be replicated to manage the non-deterministic factors brought by the use of agentic AI. Selecting the right technical KPIs to create guardrails on potential outputs by agentic AI components ensures robust decision-making capabilities by the AI agents and mitigates the? uncertainties caused by non-deterministic characteristics, such as the potential for? hallucination. It is critical that decision-makers and technical teams define relevant KPIs in each level to ensure that AI solutions create the expected value upon implementation.?

The framework defines 3 levels of KPIs:

• Operational: measuring the value of solving the problem for the user
• Business: measuring the business value
• Technical: measuring the value of the output of the AI solution

When a software solution is implemented, it is expected that it creates value to the user of that solution. This is the operational KPI layer. In traditional RPA, just as GenAI-augmented and agentic systems, KPIs in this layer often revolve around employee productivity and operational efficiency, or improving the customer experience in the case of external-facing applications. This measurement dimension does not change across the evolution of automated software.

When operational KPIs are realized, business value is expected to be created, either in the form of reduced costs or increased revenue. This is the business KPI layer.?

With both GenAI-augmented RPA and agentic systems, a novel non-deterministic element is introduced to the solution that must be managed carefully in order to ensure high quality and accuracy. We propose the addition of a technical KPI layer to the business and operational layers to measure and quantify the output generated by Gen-AI modules and agents. Ideally KPIs are measured throughout the implementation, deployment and maintenance of agentic systems, this way quality assurance can be enforced throughout the entire solution lifecycle. Technical KPIs provide guardrails to ensure that the variable outputs fall within the required success criteria. For example, in the case of a Generative AI-based software module that performs some form of content generation, image or text accuracy metrics – defined in partnership with appropriate stakeholders – can be employed to execute. In the case of an agentic system that provides an interactive user experience, metrics related to interaction and reasoning can be employed, such as task completion rate, instruction adherence, misinterpretation rate or harmful content detection rate. As we proposed in our previous paper, this is a critical consideration to prevent the failure of non-deterministic AI solutions in production.?

When deploying GenAI modules and agentic systems to augment traditional RPA capabilities, we must consider the technical KPI layer to ensure the expected quality and accuracy of the system.

Comparing different automation solutions

While traditional RPA robots can efficiently automate simple, data intensive and repetitive tasks, they are restricted to realize rule-based processes and workflows. Unfortunately, they often cannot accommodate the complexity of reality. Rule-based automation cannot cover many challenging, real-world scenarios where it is infeasible to prescribe rules as the basis of automation.?

In exchange for added implementation complexity and costs, GenAI can? generate content, predictions, and insights. GenAI excels at content-heavy, user-facing tasks, this way they can significantly improve the customer or user experience and boost revenue or productivity.

Combining GenAI with traditional RPA robots leverages the data synthesis capability of generative AI while executing well-defined workflows. Traditional RPA is successfully adopted in automation of backoffice processes at relatively low cost and implementation risks, it reduces costs but rarely generates revenue. GenAI-augmented RPA can provide us with? highly visible, customer-facing application flows, such as chat bots, voice bots and generative BI applications. GenAI realizes an individual task, while GenAI-augmented RPA stands for a complex process (workflows) incorporating various steps (tasks) so that a step can be a GenAI component.

Agentic systems go one step further, promising autonomy and decision automation. Making use of the complex reasoning capabilities of LLMs, they can realize autonomous decision mechanisms based on which sophisticated problems can be solved. Though their transformative power might influence business models, this emerging technology currently has high upfront costs and a wide range of risks, from integration challenges to reliability, unpredictability and latency issues.

Figure 5 provides a semantic illustration of various software robot applications: RPA, GenAI as standalone module, GenAI-augmented RPA and Agentic AI. Traditional RPA comes with the lowest implementation risk and integration complexity. In exchange for that it has limited capability. GenAI application has evolved capability (e.g. data synthesis and limited reasoning), but it increases its implementation risks and integration complexity. GenAI-augmented RPA provides enhanced capabilities in terms of software bot based process automation. It can address complex workflows requiring data synthesis and minimalistic reasoning capability. Agentic systems and multi-agent solutions provide sophisticated capabilities by leveraging reasoning and autonomous decision making, however, they have significant implementation risks and integration complexity.

Conclusion

RPA offers a reliable and cost-effective solution for automating structured business processes, enhancing efficiency and reducing operational costs. Generative AI provides advantages in content creation, personalization, and improves user experience, but requires higher costs, talent and careful oversight to ensure accuracy and compliance. Agentic systems, while giving the promise of autonomous decision making, come with greater complexity, requiring substantial investment, specialized expertise, and regulatory safeguards to unlock their full potential responsibly. We recommend adapting the discussed measurement framework when considering the introduction of software robot based automation. Starting from operational and business-level KPIs, one can pre-qualify candidate projects for automation. Provided the business-level validation, technical KPIs can help us understand and mitigate the risks of a particular solution during its development and refinement. Whenever possible, first we recommend to consider traditional and GenAI-augmented RPA solutions as they come with moderate costs, risks and talent needs. Given the higher risks of integrating agentic systems, we advise them in situations where a strong business case is identified and supported by KPIs at all levels. Agentic systems promise a completely new level of digital experience, leveraging this technology can drive companies to strategic advantages, but it requires well-defined business cases with proper risk management frameworks. We look forward to continuing to advance the capabilities of agentic AI to minimize its disadvantages and generate increasing value within appropriate use cases.