How enterprises fail at agentic AI (and how to succeed)

For many enterprises, the chatbot era of the AI revolution has already peaked. The new frontier is agentic AI.

At their full potential, AI agents can plan, act, and improve workflows, delivering real value for enterprises. But achieving that potential can be challenging, and too many firms burn money on oversized models and shallow pilots that never scale into production.

Image credit: Timmy Loen

I see some common stumbling blocks when enterprise-level companies adopt AI, whether by clinging to closed-source models or rushing into agentic systems unprepared.

The real opportunity in enterprise AI is to build systems that are private, workflow-native, context-rich, and agentic.

Pitfalls

Size isn’t everything, especially when it comes to AI models. One recurring mistake I see enterprises make is choosing the largest, most famous models for their use case.

In reality, these large models often perform poorly on specific workflows. They also tend to drive up costs, and the company ends up with a generic AI co-pilot that fails to capture enterprise complexity and provide tangible value.

Relying on closed APIs is also an issue, since it creates compliance challenges and limits control of sensitive data.

Image credit: Timmy Loen

In addition, I see many enterprises treating AI as a simple Q&A tool, rather than deploying it as something that can take action on a business process of value.

This last point raises a crucial question: How can firms know which workflows AI can safely run on its own? Keeping the shackles on AI can reduce its potential for value creation, yet unleashing it on workflows not ready for autonomy can be disastrous.

Enterprises can use this checklist to gauge the right level of AI autonomy for their use cases.

• Predictability of outcomes: Workflows with clear inputs, rules, and KPIs like demand forecasting or invoice reconciliation are safer. Open-ended processes like strategic planning can have too many possible answers.
• Tolerance for error: Workflows with built-in checks, such as fraud detection alerts, can tolerate mistakes. High-stakes areas like clinical diagnoses need human oversight and stricter processes.
• Data readiness: Even the most advanced AI agent will fail without clean and connected data relevant to the context.

Recipe for success

Step by step

Skipping the trial and error stage