Why Building Reliable Agentic AI Is More Challenging Than It Seems
338Companies are racing to get their hands on agentic AI, but experts say it’s not as simple as it looks. Curtis Northcutt, CEO of the AI startup Cleanlab, explains that creating these advanced systems involves navigating a lot of tough challenges. He warns that if companies pause or think they’ve got it all figured out, they risk falling behind.
The Rapid Pace of AI Development Makes It Hard to Keep Up
Northcutt points out that AI technology changes so fast that companies need to stay alert. Three years ago, generative AI and large language models (LLMs) shook up the market. Since then, new breakthroughs like retrieval-augmented generation (RAG), multimodal models, and smarter LLMs keep appearing almost weekly. This rapid evolution pushes companies to constantly switch their AI goals and tools to keep pace.
Cleanlab recently surveyed nearly 1,837 tech leaders and engineers. Only about 5% had AI agents in active use. Interestingly, most organizations are swapping out their AI stacks at least every three months. This shows how quickly the landscape is shifting, making it tough for companies to settle on a stable, reliable system.
Building a Solid Roadmap for Agentic AI Projects
Northcutt emphasizes that a good plan starts with clear goals. Companies should define what they want their AI to do and create detailed documents outlining inputs, outputs, and responsibilities. Assigning a dedicated product manager can help keep the project on track. Once a basic prototype is working, the focus shifts to adding safeguards and monitoring tools. These are essential to catch errors and understand how the system performs.
After establishing safety measures, organizations can introduce orchestration—adjusting how the AI responds, routing tasks, or adding fallback options if something goes wrong. Human oversight remains crucial; experts need to guide the AI and improve it through ongoing feedback. This cycle of training and knowledge curation happens even after the system is live, which makes the process complex.
The Road to Mainstream Adoption Is Still Long
Despite all the progress, Northcutt says most agentic AI projects are still in early stages or remain aspirational. Many companies overestimate how advanced these systems are. In reality, true reliable agentic AI with tool integration might not be ready for widespread use until around 2027.
Confusion about which AI technology to adopt also hampers progress. Many organizations struggle with deciding how to implement orchestration, guardrails, and feedback loops. Some spend too much time perfecting prompts or building initial systems before they even start refining them.
Northcutt recommends that CIOs and decision-makers experiment through partnerships rather than large in-house investments. This approach helps maintain control over data and projects while exploring AI’s potential. However, he warns that dependable, day-to-day agentic AI is still several years away. For now, companies should stay cautious and focus on incremental progress, knowing that true autonomous, tool-calling AI is likely to become reliable only in the next few years.
What do you think?
It is nice to know your opinion. Leave a comment.