Why agentic AI projects stumble and how to steer them right

Agentic AI is spreading fast, but success is far from guaranteed. Some companies see huge wins, while others suffer visible failures. The difference often comes down to data quality and governance, not the AI model itself.

Wayfair's agents now automate 41,000 support tickets a month, and C.H. Robinson's agents read shipping emails and create over 5,500 orders daily, saving 600 person-hours each day. But Klarna has been called the "poster child for bad AI deployments" by one analyst, with a customer service agent meant to replace 850 employees leading to quality issues and lower customer satisfaction.

A 2026 Gartner survey found that over 60% of organizations plan to deploy AI agents in the next two years, but only 17% have done so. Gartner also projects more than 40% of agentic AI projects will be canceled by the end of 2027. A 2026 Fivetran report says only 15% of organizations are fully ready, even though most have already invested millions.

The main culprit is poor data and weak governance. Agents need accurate, complete, and consistent data, plus clear lineage, shared business logic, and access controls. Without these, they can misread situations, take wrong actions at scale, and cause cascading errors.

Risks fall into three categories: operational correctness (acting on stale or missing data), control-plane (wrong permissions or weak security), and human-system (people overtrusting or failing to supervise).

Good use cases are high-volume, repeatable, text- or code-heavy, with clear success criteria and low-cost human review. Poor ones involve high stakes, sparse data, adversarial users, or irreversible actions. AI tends to augment work rather than replace it - radiologists were predicted to vanish by 2021 but saw record employment by 2025.

Building agents from scratch is costly, so a RAG approach that adds your own data to a foundation model is more practical. Roll out in tiers: read-only agents first, then drafting agents, then bounded write-back agents with strict limits and approvals for risky actions. This keeps projects from joining the 40% that get canceled.

Source: dbt Labs Blog