2025 Women in AI Expert Series: Agentic AI: Navigating Autonomy, Accountability, and Ethics

WAI CONTENT TEAM
Jun 18
6 min read

Principal author: Karen Jensen

Welcome to the 2025 Expert Series from the Global Ethics and Culture office of Women in AI.

In 2025, we continue our global initiatives in Education, Entrepreneurship, Innovation, and Research to make AI accessible and inclusive for everyone, with a special focus on women and girls.

Like the 2024 Speaker Series, this year's Expert Series aims to boost opportunities for women and girls in AI. We'll feature global women experts sharing practical AI skills that could help you launch a new AI career or reskill for AI roles.

In our first session, titled "Prompt Like a Pro: AI Skills for Students and Young Professionals," our expert offered practical guidance on the essential skill of prompt engineering.

Our expert for today is Dhivya Nagasubramanian. Dhivya is the Vice President for AI Transformation and Innovation at a Fortune 500 company and is currently serving as a volunteer at the Global Ethics and Culture office as our AI Finance Lead.

Dhivya’s informative presentation identifies challenges and pathways for “Agentic AI: Navigating Autonomy, Accountability and Ethics” and highlights why 2025 is the year of Agentic AI!

Understanding the Evolution of AI: From Traditional to Agentic AI

Our session started with a walkthrough of the significant transformation in the AI landscape. Dhivya explained that Traditional AI relies on rule-based or statistical models primarily for narrow, predictive tasks, like image recognition, without an inherent action loop. Building on this, Generative AI, exemplified by tools like ChatGPT, focuses on creating new content based on learned patterns in a straightforward 'prompt-generate-done' cycle. The next step was AI Agents, which introduced the ability to perceive, decide, and act towards specific goals within an environment, as seen in summarizer agents. However, the most advanced form, Agentic AI, elevates this by incorporating memory, sophisticated reasoning, planning capabilities, tool use, and self-adaptation. Its continuous 'perceive-reason-act-learn-repeat' loop allows for true autonomy and continuous improvement, making it suitable for complex roles like research assistants or orchestration agents.

From Automation to True Autonomy: The Five Levels of Transformation

Dhivya then outlined a compelling progression, illustrating the transformation from manual operations to full autonomy using an automobile use case. This evolution consists of five distinct levels. Level 0 represents Manual Operations, where human effort is solely responsible for tasks. As AI is introduced, we advance through Levels 1-3, which involve increasing automation coupled with human supervision, progressing to AI handling significant portions of a process but maintaining critical human oversight. The true leap occurs at Levels 4 and 5, representing semi-autonomous and fully autonomous systems, respectively. Level 4 involves minimal human intervention, while Level 5 is the ultimate goal: complete, human-free autonomy.

How Agentic AI Works: The Core Process

At its heart, Agentic AI operates through a continuous, iterative cycle, mirroring human cognitive processes. First, it Perceives by gathering and processing data from various sources like sensors and digital interfaces to build a contextual understanding. Next, its Reasoning engine, often powered by a Large Language Model (LLM), orchestrates decision-making, leveraging techniques like Retrieval-Augmented Generation (RAG) for enhanced accuracy by accessing proprietary data. Once a plan is formulated, the AI proceeds to Act, executing tasks by integrating with external tools via APIs, with built-in guardrails to ensure compliance. Finally, the system Learns through a continuous 'data flywheel' feedback loop, refining its models and optimizing decision-making to adapt and improve performance over time.

The Lifecycle of Intelligent Behavior in Agentic AI

In Agentic AI systems, "intelligent behavior" is enabled by several vital components. Goals define the overall purpose or desired outcome. To achieve these, the system breaks them down into specific Tasks, which are then assigned to Agents—autonomous workers designed for particular purposes. These agents are equipped with various Tools, such as APIs and plugins, to perform their assigned tasks. Furthermore, sophisticated Memory (including short-term, long-term, and episodic memory) allows agents to retain information and learn from past experiences. Finally, Coordination mechanisms, whether sequential or hierarchical, facilitate effective planning and communication between multiple agents working on complex workflows.

Reasoning: The Major Differentiator

Dhivya emphasized that sophisticated reasoning is what truly distinguishes Agentic AI from other forms of artificial intelligence. Unlike simpler AI, Agentic AI doesn't just execute commands; it processes information and makes intelligent decisions based on its goals, tasks, and past experiences. This isn't a singular process but involves multiple intricate layers that collectively enable autonomous decision-making and problem-solving. These layers include agent action selection, where the AI intelligently determines the next best step or tool to invoke; a planning layer, which sequences multi-step plans to achieve a defined goal; meta-reasoning, allowing the AI to critique its own past actions and revise its strategy; memory-based reasoning, which leverages long-term knowledge and experience for informed decisions; and multi-agent reasoning, crucial for coordinating and negotiating tasks when multiple agents collaborate. This multi-faceted reasoning capability is precisely what grants Agentic AI its unique ability to autonomously adapt, learn, and operate with greater intelligence.

Agentic AI Frameworks: Coding for the Future

The field of Agentic AI is rapidly evolving, supported by a growing ecosystem of robust frameworks designed to facilitate the creation of these advanced systems. For developers, code-first frameworks like LangChain, CrewAI, and AutoGen provide powerful tools in Python for defining agents, building teams of collaborating agents, and orchestrating complex workflows. These offer fine-grained control over logic and interaction. Simultaneously, the rise of no-code/low-code platforms like LangFlow, Flowise, and Botpress empowers business users and AI builders to design and prototype Agentic AI solutions through intuitive drag-and-drop interfaces or visual workflows, democratizing access to this cutting-edge technology and accelerating its adoption across various industries.

An Agentic AI Workflow Example

To provide a concrete illustration, Dhivya presented a practical scenario: a product marketing team's need for daily insights into competitor activity, encompassing product updates, pricing changes, and social sentiment. The ultimate goal is to generate a comprehensive daily summary from various web and internal CRM sources. In this Agentic AI workflow, a Retriever Agent would first identify relevant competitor information and then scrape external news and social media. This data would then be passed to an Analyzer Agent, which processes and summarizes key changes, comparing them with internal CRM pipeline data. Following this, a Strategy Agent would formulate strategic suggestions based on the analysis. Finally, a Reporter Agent would compile all this information into a cohesive report, suggesting possible actions for the marketing team. This entire process is enhanced by the system's memory of past summaries and user feedback, and its access to various tools like Google News API, Twitter API, CRM API, and advanced LLMs.

Navigating the Future: Ensuring Accountability, Ethics, and Governance for Agentic AI

The promise of Agentic AI is vast, but so are its inherent risks. Dhivya's examples of AI failures, like Air Canada's chatbot giving incorrect bereavement advice or a logistics company losing $1.2 million due to automated rerouting errors, forcefully underscore the non-negotiable need for accountability.

To mitigate these risks and build trustworthy systems, explainability is vital. This means implementing tools like Goal-Action Trace Logging to record decision pathways, utilizing Interactive Explainability Dashboards for transparent insights into agent behavior, and providing auto-generated explanations to clarify AI outputs.

Moreover, deploying Agentic AI demands a deep commitment to ethical principles. This includes ensuring Value Alignment, Fairness, Non-Discrimination, Safety, Privacy, and Global/Cultural Sensitivity. Crucially, human oversight serves as our primary safeguard. Humans are essential for preventing malfunctions, verifying critical decisions, maintaining an accountability chain, and providing a training and feedback loop that allows the AI to learn and improve.

To govern Agentic AI effectively, best practices are paramount. This involves establishing clear responsibility boundaries, designing human-in-the-loop protocols for meaningful control, conducting continuous risk assessments, embedding explainability into system architecture, defining escalation and override mechanisms, ensuring regulatory compliance, and actively fostering stakeholder engagement and transparency. By adopting these measures, we ensure that Agentic AI develops responsibly and benefits all.

What are your thoughts on the future of Agentic AI and its impact on our world? Share your comments here on this post and with us on Social Media @WomeninAI to ensure we #MakeitFAIR!

Event recording

You can view the recording of the event using this link.

This Expert Series is presented by the Women in AI Ethics & Culture Office volunteer team, dedicated to A Global Vision for achieving gender parity in emerging technologies through increasing Opportunity, championing inclusive Policies, and fostering practical Action that delivers meaningful and measurable impact.