The Agentic AI opportunity: Closing the gap between hype and reality with Decision Intelligence Centric Agents

Hype has a way of obscuring the bigger picture. In the second part of our Agentic AI series, XpertRule Chairman and Founder Akeel Attar explains why a Decision Centric approach is the most pragmatic path to successfully scaling AI Agents in the enterprise.

In part one of this two-parter, I explained why there’s currently a gap between the hype around Agentic AI and the reality of what the technology is currently capable of. (In short: we can’t trust Agentic AI with mission-critical and regulated decisions, because of the weaknesses inherent in the LLMs powering the agents.)

What I didn’t do was outline any viable solutions — which we’ll get to in a moment. Before I do, I want to hammer home my central point:

The need to move from Gen AI Centric Agents to Decision Intelligence Centric Agents.

Why this is important:

The autonomy of Agentic AI has been the cause of much buzz and excitement but it’s this perceived autonomy can also make it a potential risk to businesses depending on the nature of the agentic application. So when can we trust an AI Agents?

LLM powered AI agents have been successfully used for automating search, research, and analysis of enterprise & web content acting as a co-pilot producing draft reports for the human worker to validate and use. In such applications, the agent uses Gen AI for planning, reasoning and for processing enterprise data and communications (chats, voice, emails etc.). This can be a very effective human-agent collaboration framework provided that the time the human saves on research and analysis is much greater than that the worker spends on validating the output of the AI agent!

However, for complex applications requiring the automation of mission critical and regulated decisions, LLM Hallucinations in planning and reasoning tasks, makes it almost impossible for the agent to achieve autonomy and increases the need for frequent invocation the human in the loop to unacceptable levels even when using complex guardrails. To address these problems, we need a new paradigm for Responsible AI agents for mission critical – A Decision Intelligence Centric Agentic AI.

What is A Decision Intelligence Centric Agentic AI and why does it matter?

There are two main components to an AI Agent; a central ‘brain, that plans, reasons, and acts and a ‘perception layer’ for sensing and interacting with its environment (Enterprise systems and communication). The main source of hallucinations in AI Agents results from using Gen AI as the brain of the agent in complex applications. In a Decision Centric Agentic AI, we use a Decision Intelligence Engine for the brain (Planning & reasoning) while using Gen AI for perception (processing unstructured enterprise data and communications).

This represents a Responsible AI framework with a pragmatic hybrid approach to delivering scalable AI agents we can trust with complex, mission-critical tasks.

In a time when ‘shadow AI’ and unauthorised applications are running amok in businesses, a Responsible AI framework is needed to ensure agentic solutions are not just scalable, but accurate, consistent, auditable, safe and adaptable.

Decision Intelligence uses Composite AI technologies— such as symbolic AI, predictive machine learning, optimization, and Decision / knowledge graphs.

“Composite AI represents the next phase in AI evolution,” wrote Gartner in the 2024 Hype Cycle for Artificial Intelligence. But rather than being simply ’The Next Big Thing’ in the hype cycle, Composite AI instead combines the strengths of all of the AI advances that came before it — Gen AI, agents, symbolic, machine learning, etc. — to build business applications that are more than the sum of their parts.

How a Responsible AI framework delivers on the promise of agentic

The cornerstones of this Responsible Decision Centric Agentic AI framework are:

• Reliable, auditable and explainable reasoning. Use a Decision Intelligence engine for decision flow planning and reasoning.
• Exploit the strengths of LLMs. Use LLMs for their very powerful pattern recognition capabilities to extract structured nuanced attributes from unstructured enterprise data (videos, images), documents (written procedures, contracts etc.) and all communications (emails, chats, voice). Use validation rules as further guardrails in mission critical and regulated applications.
• Keep a human in the loop. Integrate human expertise to validate nuanced attributes extracted by LLMs from unstructured text or decisions made by LLMs from unstructured rules, regulations, etc. Driving human interactions intelligently via decision engine inference will reduce the work load on human workers by ensuring that they are only called when needed and given a full context of the decision / assessment they need to make.
• Adapt and learn. The Decision Intelligence engine can use off-line non-gen AI Machine Learning for ongoing learning and adaptation, while the Gen AI engine will learn and adapt dynamically.
• Decision Intelligence Centric Agents can collaborate with Gen AI centric Agents and human workers.

In this Responsible Agentic framework, we see the Decision Intelligence Centric Agentic AI approach in action, with multiple AI technologies working in conjunction alongside human workers, and Gen AI Agents to power better business outcomes.

The above framework can successfully deliver Responsible AI Agents involving structured and unstructured data in many mission-critical and/or regulated complex enterprise applications, including:

• Audit & due diligence
• Structured assets inspection & reporting
• Fraud detection
• Insurance claims processing
• Resource planning and optimisation
• Performance optimisation in Manufacturing applications
• Risk assessment
• Complex & regulated diagnostic and troubleshooting of equipment & machinary

I’d love to know where you are in your Agentic AI journey. Contact me here with any thoughts or questions.

And follow us on LinkedIn to ensure you don’t miss any insights or updates from XpertRule.