Agentic AI: A Safety-First Approach

For the past year or so, Agentic AI has become a trend as an evolution of Robotic Process Automation (RPA) in the business process automation domain. It refers to artificial intelligence systems designed with the capability to make autonomous decisions and take actions based on their understanding of the environment and goals.

Despite the flexibility offered by Agentic AI, it must prioritize safety, enforce robust control mechanisms, and ensure responsible decision-making.

Actions taken during the process should be well-defined and adhere to all constraints within the business context of the domain.

The following is a list of considerations for developing a successful and secure Agentic AI strategy, focusing on identified sets of States and Actions.

1. Agentic AI is aware of business context. Without understanding the business context, the AI's actions would be meaningless or even harmful.

2. Possible states in the business domain should be predefined. However, it doesn't mean the AI can't encounter states outside this predefined set, but it will provide a starting point.

3. For practical purposes, there are a finite number of states in the business domain. While the underlying factors influencing the business might be continuous, we need to discretize or categorize them to create a finite set of states that the AI can work with.

4. For safety and control reason there is an Action set with finite number of elements in every business domain. This allows us to manage and predict the AI's behavior.

5. Each state from business domain should be mapped to the subset of elements from Action set. This allows for flexibility and complex strategies, as the AI can choose the most appropriate combination of actions for a given state.

6. Actions defined when Agentic AI was designed and can be added later when needed. It allows for iterative development and adaptation. We start with a core set of actions and then add more as the AI's capabilities grow or as the business needs change. However, adding new actions should be a controlled process, not something the AI does on its own.

7. Nature of actions predefined, and those actions have expected outcome. We need to understand the nature of the actions and have reasonable expectations about their outcomes. This is essential for safety and for evaluating the AI's performance. It doesn't mean the outcomes are certain, as business environments are often uncertain, but we should have a good understanding of the likely consequences of each action.

8. Possible emergent behavior as a result of interactions between Actins is calculated and accepted by organization. Even with predefined actions, complex interactions can lead to emergent behavior. It's important to anticipate these possibilities as much as possible. We should model them and ensure that they are within acceptable boundaries. Thorough testing and simulation are crucial.

9. Agentic AI will not invent a new Action on the fly. This is a critical safety requirement. Allowing the AI to invent new actions would be extremely risky and could lead to unpredictable and potentially dangerous outcomes.

Overall, by emphasizing predefined states, actions, and controlled action selection, we will be able to mitigate the risks associated with emergent behavior and ensure that the AI operates within acceptable boundaries. This approach will allow us to build trustworthy and effective Agentic AI for business.