Agentic AI projects are hitting a wall, and the cause is rarely the model itself. Real-world results vary wildly: Wayfair's supplier-support agents now handle 41,000 support tickets a month, and C.H. Robinson's agents read shipping-request emails, connect data across messages and attachments, and automatically create over 5,500 shipping orders per day, saving 600 person-hours daily. But Klarna's customer service agent, meant to replace over 850 employees, led to quality issues and declining satisfaction—one analyst called it the "poster child for bad AI deployments."
The difference comes down to data trust. According to the 2026 Gartner CIO and Technology Executive Survey, over 60% of organizations plan to deploy AI agents in the next two years, yet only 17% have done so. Gartner also projects more than 40% of agentic AI projects will be canceled by 2027. A 2026 Fivetran report found only 15% of organizations are fully ready, even after most have invested millions. The key limiting factor is poor data quality and governance, not model capability.
Agents act on data continuously, unlike humans who can intuit or investigate context. Without governed, consistent context—including lineage, semantic layers, and access controls—agents can misanalyze, act wrongly at scale, and cause cascading errors. Public failures illustrate this: Air Canada's chatbot gave incorrect bereavement pricing info, and Replit's copilot deleted a production database during a code freeze. Risks fall into three categories: operational correctness (stale or missing data), control-plane (wrong permissions or weak auditability), and human-system (overtrust or poor supervision).
Good agentic use cases share traits: high volume, repeatable structure, text-heavy inputs, clear success criteria, low-cost review, reversible actions, and authoritative data. Poor cases have ambiguous accountability, high stakes, sparse data, adversarial users, or irreversible actions. AI tends to augment work, not replace it—Geoffrey Hinton's 2016 prediction that radiologists would be extinct by 2021 never materialized; by 2025, their pay and employment hit record highs.
To build trustworthy agents, start with read-only retrieval, then drafting with human review, then bounded write-back actions with approvals. Prohibit autonomous execution for sensitive or irreversible actions. Fivetran's own team used this approach: a conversational analytics agent on Jira data via BigQuery and a Claude Skill, plus a Zendesk support agent drawing from multiple internal sources. Both are narrow, high-volume, and grounded in authoritative data.
Source: dbt Labs Blog