Agentic AI has experienced double digit growth in deployments over the past year. Gartner predicts that it will be making 15% of the day-to-days by 2028, and that 33% of software applications will be designed with agentic AI as part of their infrastructure by the same year.[1] However, despite demand and interest, the implementation of agentic AI remains challenging. The Deloitte Emerging Technology Trends survey found that while 38% of companies are piloting a solution, only 11% are actively using them and 35% have no formal strategy at all.[2]
The challenge is that companies remain uncertain as to what agentic AI can achieve, how to implement it within existing systems and what they need to do to ensure it does what it says on the AI tin – it reinvents and reimagines the way the business operates.
“Agentic AI is AI that’s capable of both acting and answering,” says Dimitri Denissiouk, Managing Director at the IBA Group South Africa. “These systems can set goals, plan steps, choose how to reach those goals and even use external tools to monitor results and adjust along the way. A digital teammate, if you will.”
It’s a not-so-subtle shift from traditional automation, which follows fixed rules and predefined triggers. Agentic AI introduces reasoning and autonomy and can carry out multi-step tasks with minimal human input. In theory, this means agentic AI can improve business output by taking over repetitive operational work and handle full workflows.
“Done well, agentic AI can help companies work more intelligently across disjointed environments, reduce duplication, prevent manual errors and create a more efficient working rhythm,” says Denissiouk. “However, all these promises come with barriers that CIOs need to address early, particularly around risk, governance and strategic alignment.”
The first challenge is the unpredictability of agentic AI as it can choose how to achieve its goals independently and this may result in it taking unexpected actions. This is powerful, but when it comes to messy, real-world data, it can misinterpret context and take inappropriate actions. This is a risk if it has access to key systems with production data.
“Security and governance are the second major barrier as agentic AI systems require access to corporate systems and sensitive information in order to execute tasks, and that expands the attack surface,” says Denissiouk. “Giving a semi-autonomous system access to sensitive data opens up new vulnerabilities that traditional governance and audit models are not equipped to handle. They were in place long before AI so there’s a risk of data leaks or compliance issues if strong controls aren’t built in.”
The complexity of integration can also become a practical constraint. Agentic AI may still rely on smooth interoperability between systems, but many companies still operate on fragmented technology stacks alongside more technical limitations such as reasoning and memory constraints. This impacts consistency across long workflows, which means agentic models require careful testing before they can be trusted inside core business processes.
“Strategic alignment and a lack of clear ROI are also common barriers,” says Denissiouk. He notes that buy-in from executives is essential, but AI requires upfront investment in governance, infrastructure and training. “It’s often hard to prove business value upfront because while the output can be easier to quantify, the initial investment is harder to estimate because companies don’t always know what it will cost to put the right capacity, security controls and governance in place before implementation.”
Implementing AI requires a pragmatic approach that starts small and gradually evolves upwards while assessing governance, security, data protection and processes. The technology is not another trend, it is a fundamental shift in how AI is used, making this the perfect time to approach its integration with both caution and meticulous planning.
Share