How AI agents are different from Automation systems (RPA, BPA etc)
David Jitendranath
AI Strategy Developer, Advisor, CxO -- Architect of the absurdly brilliant
AI Agents are systems that are capable of doing complex open ended tasks in the real world just like a real person.? Sometimes they are called Agentic AI. Automation is a word that is synonymously used with AI Agents. In this blog I will point out how AI agents are very different from the conventional automation systems that we have today. How the approach of using AI Agents as automation systems is limiting the potential massive gain from the emerging innovation of Agentic AI.
The approach of using AI Agents as automation systems is limiting the potential massive gain from the emerging innovation of Agentic AI.
Traditional automation systems today range from simple automation programs that can complete one single task or complex orchestrated workflows that can span many days or even weeks to complete.???
Single task automation typically does one specific step on a trigger event. Since the task is not triggered by human intervention it is categorized as automation.? For example, one can think of notification emails when a server reaches 98% CPU utilization. Very often these simple task automation programs are chained to accomplish more complex tasks. Robotic Process Automation (RPA) Tools such as UiiPath would fall in this category.?
In the latter case of more complex orchestrated workflows, Automations are targeted on complex business processes that are asynchronous in nature and span multiple departments or business units over days or sometimes weeks.? Think of processes such as payroll, multiple departments may be involved to process and run payroll for an organization. Some of the tools that have been traditionally used in this area are BizTalk, Mulesoft, Camunda, Zapier etc.
These conventional automation systems are programmatically given definitive instructions. All possible conditions and outcomes are evaluated, defined and tested to ensure the automation systems can execute on the given workflow.?
Whichever type of automation you undertake, the common theme in these simple or orchestrated automations is that the steps are pre-determined.
AI agents on the other hand are very different from the automation systems that are explained above. The following are some of the key characteristics of an AI agent
1. Task planning x Auto-Chaining
Determining the steps to complete a task is something we automatically do with our human intelligence. Given a task or a goal, we break down that goal into simple steps. This breakdown of a goal into simple steps in cognitive science is called planning. This sort of breaking down of a goal into steps is accessible to all of us humans because of language. Our planning requires language to process, determine the steps and articulate it. In other words, we think with language in our heads.?
We think with language in our heads.?
This groundbreaking insight actually comes from the Psychoanalyst and Philosophical tradition of Structuralism. The famous quote by philosopher and French psychoanalyst Jacques Lacan comes to mind, “The unconscious is structured like a language”.
When AI agents are given a task the steps to complete the tasks are not pre-determined like the conventional automation systems.?
AI agents will actually determine the steps needed to complete the task. With the advent of LLMs, language is now accessible to machines. AI agents utilizing LLMs can mimic thinking and do the same type of planning, i.e. breaking down a goal into simple steps.
Language is now accessible to machines
Some of the planned tasks may also be chained to accomplish a larger goal.
This chaining once again is not done in a predetermined way.? The chaining of tasks sometimes may not even follow an optimal path, i.e. a path that may be evident to us as the most efficient one.? On completion of the task, the final outcome can be evaluated on how close to the goal the agent was able to reach and how optimally it was done. The result of the evaluation can lead to learning or agent training. Recursive learning of agents will be a blog for a different day.
Micro-decision Making
In automation systems of today decisions are always known in advance. The various possible outcomes of each step are assessed and defined. In anticipation of these defined outcomes successive and consequent steps are taken. Orchestration frameworks such as Mulesoft use a more flow chart type approach for automation. With the flow chart approach decision making is done by the programmer who designs the automation system. The programmer anticipates various outcomes of the program in the process of testing and each outcome is handled.? All exceptions are thrown into an exception handler.
In the case of AI agents, the agents through reasoning are enabled and allowed to make Micro-decisions.
领英推荐
To illustrate this distinction, consider the example of a conventional automation system that notifies via email when something happens. When the automation system cannot find an email or the email bounces, it simply fails and may notify another generic email of its failure to find the target email.?
Agents on the other hand may be empowered to “decide” to find the phone number of the target contact to notify them. In other words agents would try to stay true to the original goal of “notify” rather than just focus on a specific pre-determined function, to email.
Tools & Agent interaction
In traditional automation all interactions with external systems are typically done with API Calls. These external systems are once again pre-identified and the API interactions pre-programmed as part of system integration projects.?
In the case of AI agents, agents will be enabled and allowed to discover internal and external agents. The choice of these interactions will be left to the agents. The interactions will not be pre-linked and coupled. Agents may even choose to create new agents on the fly using code generation tools or LLMs.?
AI agents with the aforementioned characteristics in the end will produce very different business outcomes and impacts.
Conventional automation systems have been viewed to produce business outcomes such as
Agents on the other hand will offer companies Business scalability.
Agents with an autonomous nature will find a way to augment the workforce. They are not cheaper replacements to your current workforce. Instead they will provide enhancement to your current workforce. Think of a business that can quickly pivot to produce different product lines without disrupting existing product delivery. Currently this sort of nimbleness for organizations is very hard to achieve.??
Secondly, Agents will offer startups and early stage companies a faster way to grow without hiring. This is more an augmentation than a productivity improvement. One can imagine an organization or a company with human employees and agents. The human employees will be more generalists that will have a wide array of responsibilities. AI agents will have very niche roles and will be overseen by the human generalists.
Are you curious to learn more about the world of Agentic AI?? Join our AI Innovators community online. A community of AI innovators and leaders. When you join the community, you will?
A 90 day free trial is currently available.? This offer will end in 90 days.