(Ranjan, et al. 2025) Agentic AI in Enterprise

LLMs are about content generation.
AI agents leverage LLMs for making decisions and taking action – by planning and executing a workflow. They collaborate with other agents in a goal directed framework.

As a paradigm, agentic AI is about agency – ‘a leap from tools that assist to partners that act’.

The evolutionary path

The 1990s were about rule-based enterprise automation systems with predfined workflows. These were deterministic, and procedural in nature.

The 2000s marked the rise of ML eclipsing the previous generation of rule-based systems. The abundance of data from Web scale systems enabled predictive analytics through ML algorithms that learn from data. Data became more than a record of the past – it served as keyhole to viewing the future.
Two types of algos are fundamental to ML:

  1. Supervised learning (e.g. , classification, regression), and
  2. Unsupervised learning (e.g. , clustering, dimensionality reduction)

‘These helped drive efficiencies across industries. Retailers leveraged ML for demand forecasting, significantly reducing overstock costs and improving stockout prediction accuracy by the late 2000s.’

Improvements in supply chain efficiency and pricing strategies transformed retail operations. Howerver ML systems are constrained by their reliance on curated (structured) datasets, while the vast majority of data is unstructured and hitherto unexploited.

This constraint saw the rise of deep learning in the 2010s.
Deep learning is based on artificial neural networks that process unstructured data with high accuracy.
While ML relies on curated datasets, DL extracts ‘hierarchical features’ from raw inputs – images, audio, and text – enabling breakthroughs in perception and cognition.

Hurdles

  • Cost of computation: Large models consume immense amounts of energy in training.
    GPT-3 is estimated to have used 1,287 megawatt-hours (MWh) of energy for training. This amounts to the annual energy consumption of over 120 US households.

  • Ethics: Training data can contain biases that can result in models generating biased output and unfair outcomes, discriminating against vulnerable groups like native Americans or blacks in America, women and children etc. For example a traditional gender bias that still remains is gender disparity in pay, and racial disparity in the C-suite. This might cause imbalances to perpetuate in hiring systems or facial recognition systems that treat a turbaned person as a terrorist because most ISIS terrorists sported turbans.

The challenge for AI is in the domain of creating responsible, equitable, and sustainable AI systems.

The 2020s marked the onset of Generative AI.
While ML & DL were traditionally leveraged for analytics, Gen AI is used for creative intelligence. The ‘Transformer’ architecture posited by Vasvani et al, was the driving force behind models like OpenAI’s GPT-4 – multi-modal models that generate human-like text, images, and audio.

Gen AI has ‘revolutionized industries by accelerating content creation, automating design processes, and enhancing decision-making’.

It carries its own dark-side:

  • copyright infringement over training data and statistically derived Gen AI content,
  • exponentially higher energy consumption in training (GPT-4 estimated to have consumed tens of thousands of MWh), and
  • upheavals in the workforce (euphemistically referred to as workforce transitions in the book) caused by its generative implications.

…and challenges:

  • maintaining ethical decisions: where does an autonomous decision cross ethical boundaries for example in pursuign a unproven line of treatment or a trading strategy in search of profits
  • mitigating bias in goal prioritization
  • autonomy increases the attack footprint: agents are vulnerable to malicious attacks and operational failures. Mitigation measures for this include:
    • Enforcing attribute-based access control (ABAC) – a more fine-grained version of role-based (RBAC)
    • upping the ante wrt in-transit and at rest data encryption by using AES-256 coupled with secure key management using tamper-proof hardware modules (HSM) like AWS CloudHSM
    • using volatile memory to auto purge all derived data post AI processing using tools like Apache Ignite
    • storing anonymised data for compliance audits

AI governance frameworks play a role in defining policies and practices to guide responsible AI use, ensure transparency, accountability, and compliance.
API governance, on the other hand, establishes security protocols, rate limits, and data access controls to prevent unauthorized interactions with AI agents.

Gen AI has nudged enterprises to change tack from an automation focus to a possibility driven, creational focus. With Agentic this has pivoted towards autonomy and agency – systems that take proactive action and make real-time decisions.

Key aspects of Agentic AI are:

  • goal-driven intent
  • independent or autonomous operation (within parameters)
  • a focus on continual improvement of outcomes

In an Agentic AI–driven workflow, agents can adjust their actions based on real-time data, past experiences, and contextual awareness. It is not constrained by predefined rules, but can reason, adapt, and make decisions with minimal human intervention.

Agentic AI is widely becoming a board manifesto, resulting in hard-driving operational goals such as to reduce operational decision latency by 30%, driving annual productivity gains. It is being viewed as key enabler for growth.
Expectations are gravitating towards proactiveness – agents that anticipate needs, and resolve issues without intervention.

‘However boards need to be wary that Gen AI needs to be used as a tool to enhance human creational ability rather than replace it. I see some boards piovoting to the latter concept, which is a fatally flawed execution idea.’ - NS

Gen AI is a knowledge revolution enabler – machines amplifying human thought.