Want to Deliver a Successful Agentic AI Project? Stop Treating It Like Traditional Software
An MIT study last year found that 95% of generative AI pilots fail to even reach the production stage, and with agentic AI, the situation isn’t much better. However, Dael Williamson, EMEA CTO at Databricks, emphasizes that IT leaders should examine what the five percent club is doing differently—they are making notable tweaks to traditional software processes.
The Shift in Focus
Williamson notes that most companies struggling—comprising the 95%—treat agentic AI development like traditional software. In traditional software, around 80% of the time was dedicated to designing and building. In contrast, with agentic AI, this approach needs to flip; too much time spent on design and build leads to insufficient testing, which is critical due to the probabilistic nature of these systems.
“The companies that are in the 5% are building evaluations and thinking about security, narrowing the context of what an agent does,” said Williamson. Evaluations are essential for fine-tuning agents and ensuring high-quality outputs. Databricks’ report reveals that enterprises focused on evaluations get nearly six times more AI projects into production compared to those neglecting this aspect.
Defining Agentic AI Use Cases
Businesses often misinterpret agents as catch-all tools, but the focus should be on curating them for specific tasks. For example, onboarding and customer support are ideal areas for agent integration. Research from Cisco indicates that agents will handle more than two-thirds of customer support interactions by 2028.
The Rise of Multi-Agent Systems
According to Databricks, there has been a 327% increase in the adoption of multi-agent workflows. Businesses are moving away from individual chatbots toward systems that involve multiple specialist agents working collaboratively. Williamson likens this to a home renovation, where different contractors contribute distinct skills.
Databricks’ analysis shows a growing use of three key types of agents: Information Extraction agents, Knowledge Assistant agents, and Supervisor agents. Information extraction agents account for 31% of all agent usage, reflecting companies’ need to leverage structured and unstructured data efficiently.
The Role of Supervisor Agents
Supervisor agents are crucial in agentic workflows as they oversee and orchestrate the activities of other agents. They ensure that all operations adhere to expected outcomes. Williamson highlights that businesses are placing significant emphasis on these agents to improve governance and reduce risks associated with generative AI, such as hallucinations and rogue bot activities.
“You have the agent that does the work, and then the agent that inspects the work, and a supervisor overseeing that everything aligns with the expected outcomes,” Williamson explained. Companies embracing robust AI governance policies reportedly put over 12 times more projects into production.
The Challenges of Governance
Despite the success of tech companies and digital natives in adopting agentic AI, many businesses struggle with governance. Implementing comprehensive governance across various departments within large enterprises poses challenges. Williamson notes the complexity of managing governance for single market enterprises with multiple functions.
In conclusion, for successful agentic AI projects, organizations must pivot away from traditional software development paradigms. By emphasizing testing, specific use cases, and strong governance, businesses can improve their chances of success in this rapidly evolving landscape.
