As generative AI continues to evolve, we’re beginning to see the next generation come to life: Agentic AI. Traditional AI is designed to answer a single prompt. By contrast, Agentic AI can perform multi-step tasks and work with different systems to achieve a more complex goal.
Customer service is a good example of an Agentic AI use case. An AI agent might handle inquiries, respond to support tickets, take follow-up actions, and even escalate complex issues to human agents. This ability to automate entire workflows and make decisions across systems is what sets Agentic AI apart. Deployed correctly, it could be a game-changer for many industries.
The promise of Agentic AI is immense. Gartner forecasts that by 2028, a colossal 15% of all day-to-day decisions will be made autonomously by AI agents.
AI agents can drive efficiency, cut costs, and free up IT teams for strategic work. However, deploying them also presents its share of challenges. Before deploying Agentic AI, businesses must address issues that could compromise the reliability and security of these systems.
1. Enhancing model reasoning and insight
As the name suggests, Agentic AI systems use multiple interacting agents to make decisions. One agent might function as a “planner” to set a course of action, while others act as “critical thinkers” that assess and adjust these actions in real-time. This creates a feedback loop where each agent continuously improves its decision-making ability.
But for these systems to be effective, the underlying models need to be trained on realistic, high-quality data — data that reflects the complexities of the real world. This requires continuous iterations, sometimes involving thousands of scenarios, before the model can reliably make critical decisions.
2. Ensuring reliability and predictability
With traditional software, we provide explicit instructions — step-by-step code that tells the system exactly what to do. Agentic AI, however, relies on a more autonomous approach, where the AI decides the steps needed to reach a desired outcome. While this autonomy offers efficiency and scalability, it also introduces unpredictability, as an agent might take a less predictable path to the solution.
This isn’t a brand new phenomenon. We saw a similar situation with the early versions of LLM-based generative AI like ChatGPT. Back then, outcomes were occasionally random or inconsistent. In the past couple of years, however, quality control initiatives like human feedback loops have made these systems more reliable.
The same level of investment will be necessary to reduce the unpredictability of Agentic AI. The technology can’t be useful unless it can be trusted to take reliable action.
3. Protecting data privacy and security
Privacy and security considerations are paramount for the organisations considering Agentic AI.
Since AI agents often interact with multiple systems and databases, they’re likely to have access to sensitive data. Similarly to Generative AI where every piece of data provided to the model gets embedded within the system, Agentic AI could inadvertently expose a business to vulnerabilities, such as data leaks or malicious injections.
To address these concerns, companies can start by isolating data and implementing robust segmentation protocols. Additionally, anonymising sensitive information, such as removing personally identifiable data (like names or addresses), before sending it to the model is key. For example, a financial institution using agentic AI to process customer requests should ensure that transaction details are anonymised to prevent exposure of sensitive data.
At a top level, right now, Agentic AI can be categorised into three types based on its security implications:
- Consumer Agentic AI: These models interact directly with end-users, so security measures are crucial to prevent unauthorised data access
- Employee Agentic AI: Developed for internal company use, these systems carry less risk but can still expose sensitive information to unauthorised employees. For instance, companies might create their own GPT-like system for internal tasks, but it needs safeguards to protect confidential data
- Customer-facing Agentic AI: These systems serve external clients and must be designed to protect both customer data and proprietary business information
4. Ensuring data quality and relevance
For agentic AI to perform at its potential, it needs to be able to draw on accurate, relevant, timely data. Many AI models struggle to deliver that pipeline because they don’t have access to real-time, high-quality data — whether that’s an issue with the data itself, or the pipeline that supplies it.
A Data Streaming Platform (DSP) can address these challenges, allowing businesses to collect, process, and transmit data in real-time from multiple sources. For instance, developers can use Apache Kafka and Kafka Connect to integrate data from various sources, while Apache Flink facilitates communication between different models.
Agentic AI systems can only succeed, avoid errors, and generate accurate responses if they are built on trustworthy, up-to-date data.
5. Balancing ROI with talent investment
Deploying Agentic AI requires considerable upfront investment, not just in hardware and infrastructure, but also in acquiring specialised talent. Companies may need to invest in memory management systems, new GPUs, and new data infrastructures, while in-house teams must be trained to build inference models and manage AI systems.
Although the initial return on investment (ROI) is reliant upon a careful, methodical implementation, the long-term benefits can be significant. In fact, tools like Copilot are already being used to autonomously write and test code, showcasing that businesses can start integrating these systems today.
Despite its challenges, Agentic AI is poised to revolutionise business. With the power to outpace Generative AI, it’ll drive decisions at scale across industries — from healthcare to autonomous vehicles.
Though the path to adoption may be tough, the impact will be massive, reshaping how businesses operate. The key? Investing in quality data, solid security, and the right infrastructure. Once in place, Agentic AI can unlock huge efficiencies, help decision-making, and fuel growth.
- Data & AI
