Agentic AI: A Solution for Your Product?

Customer service. Workflow automation. Personalization. Research and analysis. Regardless of purpose, agentic AI is a viable solution for various technological issues. Agentic AI technology is advancing rapidly, which is why we spoke with Klika's Tech Lead, Haris Muharemović. We discussed agentic AI and its purpose, limitations, and potential future applications in June's edition of the Techtonic newsletter.

Let's start with the basics: What is an agentic AI, and what is its primary purpose?

Agentic AI is a type of AI designed to act on your behalf. It can understand, plan, make decisions, and take actions based on those decisions. It does not just respond to the question or prompt.

If you constantly perform small-to-medium, repetitive tasks like moving data from one software to another, sending updates, or predicting failures, Agentic AI can help you automate these tasks.

What use cases are best suited for agentic AI today?

At the moment, LLMs (large language models) work best in environments with clear workflows. To effectively use Agentic AI, we need a well-defined use case and a clear workflow.

One example is E-Commerce. Imagine a system that automatically reads an order email, enters the data into the system, notifies the warehouse, updates the CRM, and sends a confirmation email to the customer. This is a clear, repetitive workflow.

Another example is a logistics company that needs to monitor inventory levels in real time, notify the warehouse when stock is low, predict when it will run out, and create supplier orders with the correct quantity.

However, at this stage, it needs a human touch.

Like every other AI component, agentic AI must be controlled by rules and continuously monitored. What technical and interaction constraints must be in place before the launch of agentic AI components?

Agentic AI can perform exceptionally well in many cases, but must be correctly configured and monitored.

The first and most important constraint is clear boundaries. Agents need strict rules and limits on:

• What they're allowed to do

• Where they operate

• When they must escalate to a human

Another critical constraint is the sandboxing, which means that Agents must operate in a controlled environment, with limited access to the system and the data they're working with.

Agentic AI systems must be continuously monitored to minimize risk and impacts on the broader system. That's why we, at Klika, define the rules, the workflow, and the system, and continuously monitor both the agents and the system.

As we discussed this topic before, before the interview, you pointed out that agentic AI is still "unstable" but usable in „safe cases." Could you clarify those instabilities and how we define „safe cases?"

"Unstable" is a term used in system analysis, and this context often differs from everyday usage. It is more like a system that is not predictable. These systems make decisions based on probabilities and assumptions and have limited context. Sometimes, they make sense technically, but they do not always align with human expectations.

For example, an agent can perform a task perfectly nine times but fail the tenth time. Why? The data may be outdated, or the agent has limited context.

So, what are "safe cases"? These are use cases where:

• The impact of a mistake is minimal

• The system is reversible

• There's a person in the process to verify and approve high-impact actions

• The logic and data are well-defined

From a technical architecture standpoint, what's fundamentally different when designing a product powered by agentic AI compared to a more classic AI model-based system?

A classic single AI model-based system usually takes a single input, processes it, and produces output without much autonomy or decision-making. For example, think of a chatbot that can answer questions but not make decisions.

Agentic AI, on the other hand, is more like a team of agents that can work together to achieve a shared goal. It can take multiple inputs, process them, and produce outputs with more autonomy and decision-making. For example, an agentic AI system can read an email, understand the context, decide, and act.

These differences shift the architecture from a simple input-output pipeline to a dynamic and decision-driven system.

Agentic AI is a tool – how can it empower humans in their business and productivity efforts?

Agentic AI is indeed a powerful tool. When used correctly, it can raise productivity and efficiency. There are many ways, of which we can mention a few:

• Delegating complex or repetitive tasks to an agent

• Building proactive systems that monitor operations and predict failures

• Automating processes that require complex logic and decision-making

By reducing operational burdens, agentic AI can help individuals and teams focus on the broader picture, creativity, and problem-solving.

Where do you see the development of agentic AI in the next 18 months?

I'm seeing a lot of development in areas such as multi-agent systems and industry-specific agents. Multi-agent systems involve multiple agents working together to achieve a shared goal. Industry-specific agents, on the other hand, are trained to operate within a specific industry.

Let's also not forget ethics and security. I believe we can focus on these areas as well.