Agentic AI has become one of the hottest buzz words in technology today. As is typical with new capabilities, “agentic AI” has no clear, agreed upon definition. To add to the confusion, the big players (Google, Microsoft, Open AI, Anthropic, etc.) have each created a different definition of Agentic AI.
Cynically, this confusion could be seen as a mechanism used by vendors to set the stage to highlight capabilities that they provide, while downplaying those that are not offered.
Alternatively, confusion may be reduced as multiple “flavors” of AI agent emerge as technology matures, allowing for flexibly segmented development.
Is Agentic AI a breakthrough technology, or just the latest hype? Let’s dig a little deeper and see if we can find some answers.
What is agentic AI?
According to this video from Google, “No one seems to agree on exactly what an agent is”. It makes sense to think of agentic capabilities as a continuum of capability and complexity. Not all agents will be equally complex; some will perform very simple tasks. Others may have greater capabilities and the ability to handle more complexity.
For example, information summarization is typically low complexity and offers limited capabilities. Self-driving cars, on the other hand, are highly complex and also offer a slew of capabilities.
However, there are some characteristics that are generally considered to be capabilities of Agentic AI:
- Agents exist within a specific environment
- They have one or more goals to accomplish
- Agents are able to sense inputs or stimulus in their environment
- They are able to reason about the things that they sense and decide on a course of action based on the stimulus
- Agents can act autonomously to accomplish a specific goal. Actions can include:
- Providing notifications if specific conditions are met
- Reviewing documents and providing an answer to a question
- Handing off the interaction to another agent or a human
- Updating records, canceling orders, or kicking off other processes
In short: Agents Sense, Decide and Act within their Environment to fulfil a given agenda.
Additionally, they may be able to interact with other AI agents, analyze and improve their performance over time, have dialogue capabilities, act proactively, and/or affect the real world by taking actions such as preventing a financial transaction when fraud it detected, or placing orders to restock for low inventory.
According to Gemini:
Agentic AI refers to a class of artificial intelligence systems designed to operate autonomously, make decisions, and perform tasks without constant human intervention. The term “agentic” highlights their “agency” – their capacity to act independently and with purpose towards a goal.
How is Agentic AI different from traditional AI?
Gemini says “agentic AI represents a significant step forward from traditional AI, moving towards systems that are not just intelligent but also capable of independent action, planning, and continuous self-improvement in dynamic environments.”
Traditional applications rely on a set of predefined rules to make decisions or perform simple tasks, and are typically scripted. The rules and scripts control the actions. Researchers in the multi-agent systems field contend that legacy “agents” that follow pre-defined rules are really just programs, and are not truly agentic.
Agentic AI does not follow a script. It is given a goal instead. It determines the best route to reach the goal. Typically, the agentic AI is given parameters such as a set of tools, instructions on how to use them, and a description of how to execute the action.
Agentic capabilities include the ability to evaluate different scenarios and predict the outcomes of various actions. They can even assign a value to each action based on how well it aligns with the agent’s goals. By doing so, the agent acts autonomously to choose the action that is most likely to achieve its goals.
Agentic AI uses LLMs (Large Language Models) to understand inputs they receive and to communicate via dialogue. They can provide responses based on the context of a conversation, including information previously provided. This greatly improves the perceived quality of the interaction.
For example, if a user asks, “Can I add my new Chevy to the policy?” the AI can respond with “Sure. To add your Chevy, I’ll need to know…” This confirms, in a fluid way, that the transaction is progressing properly.
Users can add follow on questions, with the agent “remembering” and incorporating previous inputs. For example, “It’s a 1968 model. Is special insurance required?” Here, the agent would “remember” that the vehicle is a Chevy and investigate to provide an answer.
“Can you also add my Nissan?” The agent recognizes that this is a new request to add an additional vehicle. Agentic AI can handle this smoothly, whereas a scripted interaction would require the user to return to an earlier part of the transaction to start down this new path.
“For the Nissan, I only need Collision and Liability, not Comprehensive. How much will my rate increase?” The agent can respond to this request by looking up the information, obtaining additional information if needed, and calculating the amount to provide an answer.
How can it be used?
Agentic AI can be used in many ways, across functions and industries. From a customer experience perspective, agentic AI can be used:
To extend the capabilities of traditional chatbots to provide true problem resolution:
- Retrieve customer data, history, and knowledge base information
- Deliver context-specific responses, or pass customers to the appropriate resources for further assistance
- Take action to resolve a problem, update records, provide status information, etc.
For determining customer intent, even in situations where the information provided is complex or vague:
- Ask follow up questions to gain more context, based on each individual situation
- Determine if there is a question to be answered or if action is required
- Retrieve information to answer the question or route to another agent/program/person to fulfill the action
- Communicate with requestor regarding status or to obtain additional info needed to progress through a workflow
- Select the right tool to perform the required action
To operate across channels (voice, email, chat, SMS, etc.):
- Access full customer history and context from multiple sources (order history, chat history, previous calls, emails, etc.)
- Provide follow up information via customer’s preferred channel (SMS, Chat, email)
To automate documentation of interactions, including:
- Interaction summarization
- Trend analysis
Proactive outbound contacts:
- Follow up information
- Status updates
- Reminders
What are the benefits?
At a high level, some of the benefits that agentic AI offers for customer experience include:
- Automation of more complex processes than simple scripting programs can handle
- Improved, nuanced dialogue that appears natural to customers/users
- Offloading of mundane tasks from humans, allowing them to handle more complex tasks or those requiring empathy or creative thinking
- Faster processes
- Consistent quality
- Ongoing improvement
- 24x7x365 operation
- Handling spikes in volume that would overload traditional operations
Is there a down side?
As with everything, there are drawbacks with the use of agentic AI. It’s much like using a blowtorch. In the right situation, a blowtorch is a fantastic tool. But if used incorrectly, it can burn down a building.
Here are some of potential risks that must be mitigated when using this powerful technology.
Costs are hard to predict, especially at the outset
Agentic AI is typically priced via a consumption model; the more you use, the more you pay. Typically, consumption is measured via tokens, and each provider has a different definition and different way of calculating token use.
This makes it very difficult to forecast costs of an agentic AI solution. Past volume information is often in the form of a number of transactions by channel and/or average handle time. Tokens are calculated based on completely different information. Initial cost projections require many assumptions (that may not be accurate in the long run).
For example, Genesys token costs are based on many factors including the number of users (which can be named or concurrent), number of minutes, number of sessions, number of translations, and other factors. Fortunately, there are token calculators which help predict costs before deployment.
Garbage In, Garbage Out
Before you automate, it’s imperative that your existing processes are optimized and well documented. Underlying data must be clean, classified, and accessible. Optimizing your processes and data can be challenging. Don’t underestimate the level of effort necessary.
Complexity
Supporting agentic AI requires staff with the skills necessary for managing and governing AI agents. There can be many ongoing support challenges due to their technical complexity. They must be managed and optimized on an ongoing basis to ensure that they are operating as expected and don’t drift or experience downgraded quality. Testing and monitoring can be challenging as agentic AI applications and LLMs do not provide the exact same output in every iteration.
Sabotage
Employees who are fearful of future job loss may sabotage AI rollouts.
The Usual Suspects
AI has other risks that must be mitigated, such as:
- Inherited bias
- Possibility of hallucinations
- Security issues including:
- Risk of exposure of proprietary information
- Susceptibility to external manipulation through malicious prompt injections
Conclusion
Agentic AI is a powerful technology that promises automation of increasingly complex workflows. At the same time, poor implementations or lax security have the potential to create a fiasco. It’s important to ensure that your organization has a strong foundation from which to build AI capabilities. Global Tech offers an online quiz that will provide a high-level score of your organization’s AI readiness. We also assist in identifying and assessing the viability of AI use cases and creating strategic plans for implanting AI.
This article was originally published on the Genesys.com/blog The rise, benefits and concerns around agentic AI
