It’s not that AI models are failing. The problem lies in how we build and run them. A February 2026 Dev.to report reveals nearly 40% of agentic AI projects are canceled or paused. The culprits: over-optimistic designs, poor data quality, and costly inference methods. This has sparked a shift toward Directed Acyclic Graph (DAG) workflows, tighter data controls, and event-driven AI.

For years, the goal was the "Autonomous Agent"—think AutoGPT—where you hand over a task and the AI figures it out. In practice, "figuring it out" means unpredictability. Picture an agent stuck in an endless loop processing an invoice, racking up thousands in token costs before you can stop it. This unpredictability isn’t just frustrating—it’s expensive and legally risky.

The Air Canada chatbot case is a warning shot. The bot invented a fake bereavement policy and promised a refund that didn’t exist. A tribunal ruled Air Canada legally responsible for the bot’s "hallucination." The takeaway? "The AI said so" won’t hold up in court. This forced a move away from open chat interfaces toward Retrieval-Augmented Generation (RAG) systems that limit AI responses to verified documents. Controlled, reliable AI isn’t optional—it’s essential.

The answer gaining ground is Directed Acyclic Graphs (DAGs). Tools like LangGraph let developers map out exact AI workflows. For example, the AI may choose an email’s tone but must pass through a mandatory "Manager Approval" step. This replaces black-box autonomy with predictable, auditable paths. Experts like Charan Koppuravuri argue this structure is key to taming AI risks.

Data quality is another major failure point. In 2025, Precisely found only 12% of organizations have data good enough for AI. Many projects are built on shaky, poorly governed data—what some call "Shadow AI." Without solid data, even the smartest models fail. The fix? Strong data governance to keep information accurate, consistent, and reliable.

Finally, inference costs are pushing a move from polling to event-driven architectures. Traditional systems constantly check for updates, wasting resources. Event-driven AI reacts instantly to triggers using webhooks and Minimum Viable Co-Pilot (MCP) setups. This cuts latency, lowers costs, and scales better.

In short, AI project failures stem from flawed architectures, bad data, and inefficient inference—not the models themselves. The industry is pivoting to DAG-based workflows, strict data governance, and event-driven designs to build AI that’s reliable and accountable. The Air Canada chatbot fiasco is a stark reminder: control isn’t just smart, it’s mandatory.