Agentic AI: Deep Dive into the Next Leap in Technology

AI agents are looking to be the next big thing in the tech industry as investors rally around the new concept. As one example, Amazon.com has made a groundbreaking leap by employing AI agents to predict sales for its astonishing 400 million products in over 185 countries. Artificial support agents are also running entire departments, handling the workloads of hundreds of people.

This technology enables AI agents to collaborate with humans, setting specific goals and devising comprehensive strategies. But how does it actually work? And do we have the right data to power these increasingly autonomous agents?

Let’s start from the top.

Quick Primer on Agentic AI

Agentic AI distinguishes itself from generative AI by offering a more focused and efficient approach. In short, it’s capable of achieving complex tasks with minimal human oversight. We liken the idea of agentic AI to having another colleague in Slack to whom you can talk and receive task updates.

Agentic AI encapsulates the ability of an AI system to dynamically fulfill intricate goals within complex environments with minimal direct supervision. These systems are engineered to pursue complex objectives and workflows on their own. They are capable of decision-making, planning, and adaptive execution to fulfill multi-step tasks.

By tapping into structured data and knowledge graphs, agentic AI agents can provide precise, contextually relevant recommendations and optimizations, outperforming today’s large language models.

Key Capabilities of Agentic AI

Agentic AI systems are distinguished by their autonomy, adaptability, independent decision-making, and long-horizon planning.

• Autonomy. It autonomously executes intricate tasks, adapts to changes, and resolves issues across various languages and contexts.
• Reasoning. It employs advanced reasoning techniques, such as reinforcement learning and cognitive computing, to make informed decisions and solve complex problems.
• Flexible Planning. AI agents can manage complex infrastructure projects, coordinate large-scale disaster responses, and identify pollution sources and affected populations.
• Communication. AI chatbots with agentic AI capabilities enhance customer service by solving complex problems on behalf of customers.
• Optimization. These systems are capable of improving their own performance by analyzing data and improving workflows.

The potential of agentic AI is vast, from transforming supply chains to accelerating drug discovery. In healthcare, for instance, agentic AI systems independently monitor vital signs, administer medication, and assist in surgical procedures, enhancing patient outcomes.

Agentic AI is being explored in applications such as IT support, help desk ticketing, and human resources workflows including routine payroll, employee onboarding, and benefits management. As this technology advances, it has the potential to revolutionize enterprise operations and drive innovation across industries.

Here is a handy table referencing what it’s capable of:

In the future, virtual corporations comprising thousands of artificial agents could orchestrate incredibly sophisticated business processes.

However, the proliferation of agentic AI also raises ethical and societal concerns, necessitating robust regulation and ethical frameworks for responsible deployment. As startups embrace agentic AI, they must prioritize value and goal alignment to ensure these systems respect human preferences and boundaries.

The Evolution of Agentic AI Systems

The evolution of agentic AI systems has been nothing short of transformative, reshaping the landscape of conversational AI and language models. This evolution spans from early systems like IBM's Deep Blue to advanced platforms such as AutoGPT and LangChain.

These systems integrate three primary components: perception, reasoning, and action.

The Levels of Intelligent Process Automation, defined by Cognilytica in 2019, range from 0 to 3. Currently, most automated processes and integrations operate at Level 0 or Level 1, with only a few organizations achieving Level 2, where AI systems can manage exceptions autonomously.

Agentic AI is still in its nascent stages but is being explored for various applications, including IT support, human resources tasks, customer support, financial fraud monitoring, and healthcare operations. In healthcare, AI agents are delivering personalized patient care and diagnostics at an increasing rate. The finance sector is utilizing AI agents for predictive analysis and automated trading, with data showing their effectiveness.

The shift towards designing for AI agents (AX) instead of human customers (CX) demands new design principles focused on efficiency and transparency. Embracing this evolution while addressing the associated challenges will be crucial in shaping the future of agentic AI.

Examples of Agentic AI in Development

The emergence of agentic AI has triggered a competitive drive among tech titans and research entities to forge ahead with pioneering solutions. These leaders are at the helm of crafting multi-agent frameworks, open-source modular agents, and agentic architectures. These innovations aim to transform business operations and redefine human-AI interactions.

Microsoft's Project AutoGen

Microsoft's Project AutoGen represents a seminal effort in developing a multi-agent framework. This platform simplifies the process of building workflows and applications supported by large language models (LLMs). It incorporates specialized agents, which can be tailored with various LLMs for human-AI collaboration. Project AutoGen's design facilitates the efficient integration of human and artificial intelligence, unlocking the potential for businesses to leverage agentic AI.

Lumos

The Allen Institute for AI has spearheaded Lumos, an open-source modular agent. Lumos stands out for its natural language understanding, planning, and autonomous action execution. This agent architecture enables businesses to efficiently manage complex tasks, reducing human oversight, and leading to notable efficiency enhancements and cost reductions.

Moveworks’ AI Copilot

Moveworks, a forerunner in AI-driven service desk automation, has introduced its Next-Gen Copilot, a transformative agentic AI solution. This copilot is reportedly capable of autonomous goal-setting, planning, and execution. Through its integration with enterprise systems and data sources, the goal is to facilitate workflow optimization, decision-making enhancement, and superior customer experiences.

As businesses adopt various cutting-edge solutions, they are poised to gain a substantial competitive edge, enhance operational efficiency, and unlock new avenues for growth in the agentic AI era.

Challenges of Agentic AI in Enterprises

Agentic workflows bring numerous advantages, including enhanced performance through collaborative efforts, increased adaptability for real-time adjustments, and higher efficiency through the autonomous simplification of complex tasks. However, they don’t come without potential challenges:

Autonomy and Transparency Challenges

The autonomy of agentic AI systems is both a strength and a potential risk. They may propose actions that lead to issues due to not capturing rare edge cases, resulting in operational problems. Moreover, the ‘Black Box’ nature of their decision-making processes makes it challenging to understand and trust their choices. To address these issues, companies must prioritize transparency and implement robust monitoring and control systems.

Workflow Challenges

Looping AI agents into existing operations isn’t a simple feat and most organizations would struggle with proper governance. Oversight is of paramount importance with this, especially because, as we discussed in our last exploration, agentic AI is built on generative models - which means they ‘hallucinate’ and come to wildly inaccurate conclusions without proper guidance.

Responsible Implementation Strategies

To leverage agentic AI responsibly, companies should adopt several key strategies:

• Ensuring the AI provides explanations for its decisions to avoid the ‘black box’ problem.
• Providing humans with oversight and thoroughly testing the AI can help mitigate risks associated with autonomy.
• Implementing controls and steering mechanisms within predetermined boundaries for agents not to take on tasks they aren’t designed to tackle.

Adopting agentic AI in businesses should involve a phased approach, starting with targeted pilot projects to demonstrate benefits and build confidence. Investing in evolving skillsets to align with agentic AI requirements and reengineering processes to facilitate human-AI collaboration is crucial for the responsible implementation of this transformative technology.

By designing systems carefully and establishing safeguards, companies can maximize the productivity benefits of agentic AI while minimizing risks related to its autonomy.

The Future of Work (If We Get This Right)

Agentic AI is set to enhance employee capabilities by automating mundane tasks, allowing people to concentrate on strategic initiatives and high-judgment responsibilities. While AI agents may automate certain jobs, it is also expected to create a plethora of skilled roles, much like with the emergence of web developers during the early days of the internet.

For example, in HR, AI's utilization leads to a substantial reduction in time spent on recruitment and employee engagement tasks. Similarly, in customer support, agentic AI can function as co-pilots, assist in training, and perform sentiment analysis, enabling support staff to deliver more tailored service to customers.

As it continues to evolve, the focus will transition from individual technologies to developing infrastructure that supports truly autonomous AI agents. This evolution will emphasize the agent's ability to make decisions and take actions independently, moving beyond mere information retrieval or response generation.

By integrating external resources and human expertise into their problem-solving processes, this kind of AI can take over routine tasks independently, freeing up employees for uniquely human tasks.

Speaking of uniquely human - we’re still the ones in charge of Ground Truth and the kinds of data we train these models on…

The Role of Data in Agentic AI Systems

Data is pivotal in empowering agentic AI to maneuver through complex business environments. To operate on their own, AI agents must possess rudimentary reasoning and domain-specific knowledge. This includes goal-setting, planning, and adaptability, alongside specialized info for specific industries or functions.

We can create digital representations of business entities, also known as digital twins, to help collect the right data. These twins mirror existing assets, people, or processes virtually for AI to copy. This digital twin acts as a blueprint or knowledge graph, aiding AI agents in navigating the business landscape and executing tasks with intelligence.

Integrating Agentic AI with Enterprise Data Sources

To fully leverage agentic AI, integrating these systems with diverse enterprise data sources is essential. This integration enables agentic AI to access and analyze relevant information in real time, tailoring its actions to the latest insights.

Key considerations for integrating agentic AI with enterprise data include:

• Establishing secure and reliable data pipelines
• Ensuring data quality and consistency across sources
• Implementing robust data governance practices
• Leveraging advanced data integration technologies, such as API-based connectivity and data virtualization

By integrating agentic AI with enterprise data sources, organizations can get the best of both worlds - an AI that understands your process on the level of your best experts.

Combining Synthetic and Real-World Data

While real-world data provides authentic insights, it often comes with limitations such as scarcity, privacy concerns, and inherent biases. Synthetic data, on the other hand, offers a controlled environment with the ability to generate diverse scenarios and edge cases, albeit with potential limitations in mirroring real-world complexities.

The combination of synthetic and real-world data presents a powerful solution for model training. Gartner predicts that almost all future training will happen this way:

By integrating synthetic data into the training process, AI models can benefit from enhanced diversity, increased data volume, and reduced biases. Synthetic data can be used to address specific gaps in real-world datasets, simulate rare events, and create balanced representations for improved model robustness.

In computer vision, synthetic data has been instrumental in training models for object recognition and autonomous driving, supplementing real-world images with simulated scenarios. In natural language processing, synthetic text generation has aided in tasks like machine translation and sentiment analysis.

There’s a strong expectation that agentic AI will use synthetic data alongside real-world data to learn as well. Of course, it’s important to curate the quality of this kind of data. Ensuring consistency and compatibility between synthetic and real-world data is also no easy feat. Generative models are prone to regression over time, with performance dropping after a certain number of iterations.

That’s why data experts at Keymakr recommend carefully considering your organizational needs and sourcing ethical real-world data alongside with a human-supervised synthetic data plan. If you plan on training your own AI agents, starting with a data plan is crucial for success.

Humans and AI Agents - Setting Boundaries

Agentic AI, impressive as it may appear, needs human expertise and validation to function. The output generated by these systems needs to be accurate, relevant, and aligned with ethical and societal standards. That’s where we come in - domain experts, technical specialists, and just average people who know how models ‘think’ and process information.

We’re in charge of what data we feed to our AI in training - and we’re in charge of what it ‘sees’ in the real and virtual world. The continuous involvement of human validators and curators in the process of training ensures that AI remains a reliable and trustworthy tool.

This is especially true for agentic AI capable of far more independent operations than chatbots we’re all used to seeing by now. Our goal is to ensure we use tomorrow’s autonomous agents to augment our skills rather than replace them - and by the time we create AI that can optimize and improve itself, we better be ready to provide it with the right data and guidance to do so.

Startups would do well to focus on this emerging niche, as larger organizations are already starting to employ AI agents in vast numbers. It’s not a question of whether we get them, but of what they look like when we do, and whether we’re ready for it.