Enterprises are the building blocks of business and the extent and ease of AI implementation there, end to end, would decide how far reaching impact it makes on the economy at large and how fast. It would not be wrong to say that we are looking for an AI ERP, we are not sure of whether it would mean one for every industry, sector or just that enterprise. The first wave of generative AI has gained considerable success in companies, particularly in coding assistants and increasing efficiency of the existing SaaS products. As in case of any technology, AI is currently in a nascent stage, and it has delivered commendably.
The real strength of the LLMs is now being unfolded in the second generation of AI powered applications. This is the stage of agent-based systems, built on the same foundational models and taking the capabilities to a new level. AI agents have the operational capability for leveraging the full range of LLM capabilities. They can solve complex problems independently. It operates in its defined autonomy. Now we are faced with a billion dollar question or business dilemma; would it be possible to develop a single, comprehensive application that could solve the company’s problems?
The constraints of the enterprise environment dictate a relatively narrow scope for each individual application to ensure consistent performance and control access to data and tools. The imaginary “super AI application” would require full access to company data and tools. Unfortunately, as in the case of an employee, an agent based application’s access must be limited to what it needs to perform and function. Given this scenario how many AI agents would a company need?* A company with ten departments, each with five core functions, each needing five agents would mean 250 in one company. A realistic estimate.
Without getting into the granularity of it, every organisation can manage only a limited number of applications as per the current practices, it’s called the “complexity threshold.” As the company develops more and more agent-based-applications, maximum complexity is reached and no further applications can be developed. Using frameworks like LangChain, AI based applications are mostly monolithic. What is the road ahead? A new architectural paradigm is needed to create and maintain agent based applications. It can be mesh architecture for LLMs and the associated components required to create agents in the enterprise. It would provide abstraction layers that group different components into uniform object types. This could include; base models, data layer, service layer, orchestration layer and application layer. Service can be decoupled from its components to make enterprise AI happen.
MANAGING “COMPLEXITY THRESHOLD” IS THE KEY TO SUCCESS.
Sanjay Sahay
Have a nice evening.