INSIGHTS
Navigating the Two Phases of Agentic AI
Thuy Lam
CEO and Head of R&D
Published 5 December, 2024
Agentic AI’s evolution will likely happen in two phases. Phase 1 focuses on overcoming foundational roadblocks, like legacy system integration and data quality, while implementing structures for AI agents to make semi-autonomous decisions. In this phase, organisations build insights and decision audit trails, preparing for the leap to Phase 2, where AI agents can make fully autonomous decisions and execute tasks end-to-end. Let’s explore what these phases look like and how they’ll shape the future of AI-driven decision-making and task execution.
What is Agentic AI?
Agentic AI refers to AI models that can strategise, rationalise, and take actions autonomously with minimal human input. In a recent YouTube video, “Andrew Ng – The Rise of Agentic Workflows in AI,” Andrew shares an example of an AI agent writing a research essay. When an API search wasn’t available, the agent autonomously decided to search Wikipedia, completing the task without direct human instruction.
This example shows a simplistic form of autonomous AI agents. Andrew further breaks down the design patterns of agentic AI into these elements:
- Reflection: Examine a draft, fine tune, validate, enrich and reshape the final answer.
- Tool Use: Orchestrate the use of various models and application sources.
- Planning: Structuring how tools are used and questioned (similar to prompt engineering).
- Multi-Agent Collaboration: Connecting multiple agents to work together on task required.
An ideal phase 2 AI agent can autonomously consume instructions, decide on best course of action and complete the task. This has clear commercial potential. However, there are two roadblocks I see that stand in the way of this goal which we need to address in phase 1:
- Limited ability to take action due to legacy systems with immature APIs
- Lack of clean data that can drive these decisions autonomously
Roadblock 1: Immature APIs
The current API ecosystems are often task-specific, point-to-point and require a degree of machine interpretation to successfully execute. For instance, creating a purchase order in SAP requires calling a web service connected to a BAPI and knowing various codes (Vendor, Item, Finance, Tax) and date formats. Automation tools typically handle this by:
- Accessing various mapping tables to generate language suitable for SAP ingestion before invoking the API.
- Encoding lookups directly into the API orchestration layer.
Ideally, an application API should allow an agent to perform all actions a human would. Such as looking up codes if they’re unknown and then generating the call to create a purchase order. However, a majority of applications today lack this level of maturity and may never reach it, purely because their core business focuses on revenue from license sales.
Roadblock 2: Lack of Data or Clear Decision Processes
For AI models to perform well, they need clean data – especially data that focuses on the “why” rather than just the result. Over the years there have been advances on how to work with limited data sets and infer data artificially. However, we can’t make high-risk decisions based mostly on artificially generated data. Additionally, how often have you gone through an organisational policy and found a level of greyness that needed clarification. Feeding this kind of ambiguity into a large language model (LLM) can lead to “hallucinations” or unreliable outputs, which can pose risks in autonomous processes.
Overcoming Roadblocks
Does this mean we should abandon the concept of agentic AI all together? No! Here’s how we solve them:
1. Robotic Automation
The first is easy, we can leverage robotic automation tools, to mimic the functions of API’s. If we had access to the core data models, we could augment API’s with a level of data querying. For example, asking the data model what codes are needed prior to compiling an API. Wouldn’t it be great if we could simply use the database tables of an application? (That’s saved for the next blog).
2. Orchestrated Approach
The second requires a much more orchestrated approach. We need to provide the following tools and capabilities for agents to function in a semi-autonomous mode:
- Constrain or guardrail decisions, by applying clear logical business rules that are definitive and traceable.
- Augment human in the loop to validate insights from AI models and approve the decision.
- Enrich data into a process, that enables the automation to run smoothly in upstream and downstream systems.
These tools, combined with a clear audit trail of decisions, will mature our AI models. Over time, this will provide us with enough data to enable end-to-end autonomous decision-making and automation.
Preparing for Phase 2: Fully Autonomous Agentic AI Agents
I personally believe that organisations are not only ready for Phase 1 but also need to go through it. This phase helps them bank some dollars, build valuable insights, and establish decision audit trails – all essential before moving to Phase 2.
Think of this as growing up. In phase 1, we’re like children and teenagers – somewhat autonomous but guided by rules and regulations that teach us what is right and wrong. This gives us the ability to reason and strategise. Where we need assistance to complete tasks, we have adults to help us. As we gain knowledge and confidence, we’re able to make autonomous decisions without supervision. Similarly, agentic AI needs time to grow and learn before reaching full autonomy.
To learn how you can start implementing Agentic AI, visit www.actionfabric.ai
Agentic AI’s evolution will likely happen in two phases. Phase 1 focuses on overcoming foundational roadblocks, like legacy system integration and data quality, while implementing structures for AI agents to make semi-autonomous decisions. In this phase, organisations build insights and decision audit trails, preparing for the leap to Phase 2, where AI agents can make fully autonomous decisions and execute tasks end-to-end. Let’s explore what these phases look like and how they’ll shape the future of AI-driven decision-making and task execution.
What is Agentic AI?
Agentic AI refers to AI models that can strategise, rationalise, and take actions autonomously with minimal human input. In a recent YouTube video, “Andrew Ng – The Rise of Agentic Workflows in AI,” Andrew shares an example of an AI agent writing a research essay. When an API search wasn’t available, the agent autonomously decided to search Wikipedia, completing the task without direct human instruction.
This example shows a simplistic form of autonomous AI agents. Andrew further breaks down the design patterns of agentic AI into these elements:
- Reflection: Examine a draft, fine tune, validate, enrich and reshape the final answer.
- Tool Use: Orchestrate the use of various models and application sources.
- Planning: Structuring how tools are used and questioned (similar to prompt engineering).
- Multi-Agent Collaboration: Connecting multiple agents to work together on task required.
An ideal phase 2 AI agent can autonomously consume instructions, decide on best course of action and complete the task. This has clear commercial potential. However, there are two roadblocks I see that stand in the way of this goal which we need to address in phase 1:
- Limited ability to take action due to legacy systems with immature APIs
- Lack of clean data that can drive these decisions autonomously
Roadblock 1: Immature APIs
The current API ecosystems are often task-specific, point-to-point and require a degree of machine interpretation to successfully execute. For instance, creating a purchase order in SAP requires calling a web service connected to a BAPI and knowing various codes (Vendor, Item, Finance, Tax) and date formats. Automation tools typically handle this by:
- Accessing various mapping tables to generate language suitable for SAP ingestion before invoking the API.
- Encoding lookups directly into the API orchestration layer.
Ideally, an application API should allow an agent to perform all actions a human would. Such as looking up codes if they’re unknown and then generating the call to create a purchase order. However, a majority of applications today lack this level of maturity and may never reach it, purely because their core business focuses on revenue from license sales.
Roadblock 2: Lack of Data or Clear Decision Processes
For AI models to perform well, they need clean data – especially data that focuses on the “why” rather than just the result. Over the years there have been advances on how to work with limited data sets and infer data artificially. However, we can’t make high-risk decisions based mostly on artificially generated data. Additionally, how often have you gone through an organisational policy and found a level of greyness that needed clarification. Feeding this kind of ambiguity into a large language model (LLM) can lead to “hallucinations” or unreliable outputs, which can pose risks in autonomous processes.
Overcoming Roadblocks
Does this mean we should abandon the concept of agentic AI all together? No! Here’s how we solve them:
1. Robotic Automation
The first is easy, we can leverage robotic automation tools, to mimic the functions of API’s. If we had access to the core data models, we could augment API’s with a level of data querying. For example, asking the data model what codes are needed prior to compiling an API. Wouldn’t it be great if we could simply use the database tables of an application? (That’s saved for the next blog).
2. Orchestrated Approach
The second requires a much more orchestrated approach. We need to provide the following tools and capabilities for agents to function in a semi-autonomous mode:
- Constrain or guardrail decisions, by applying clear logical business rules that are definitive and traceable.
- Augment human in the loop to validate insights from AI models and approve the decision.
- Enrich data into a process, that enables the automation to run smoothly in upstream and downstream systems.
These tools, combined with a clear audit trail of decisions, will mature our AI models. Over time, this will provide us with enough data to enable end-to-end autonomous decision-making and automation.
Preparing for Phase 2: Fully Autonomous Agentic AI Agents
I personally believe that organisations are not only ready for Phase 1 but also need to go through it. This phase helps them bank some dollars, build valuable insights, and establish decision audit trails – all essential before moving to Phase 2.
Think of this as growing up. In phase 1, we’re like children and teenagers – somewhat autonomous but guided by rules and regulations that teach us what is right and wrong. This gives us the ability to reason and strategise. Where we need assistance to complete tasks, we have adults to help us. As we gain knowledge and confidence, we’re able to make autonomous decisions without supervision. Similarly, agentic AI needs time to grow and learn before reaching full autonomy.
To learn more how you can implement Agentic AI, visit www.actionfabric.ai
