3 ways to stop your AI agents from failing in 2026

The promise of AI agents is enormous, but the path to real-world production is littered with pitfalls. A recent prediction from Gartner warns that by 2027, 40% of enterprises will scale back or abandon these autonomous tools, primarily because governance failures only become apparent after things go wrong in the field. Yet, at the Snowflake Summit in San Francisco, three digital leaders shared how they are bucking that trend. Their strategies boil down to three core lessons: build repeatable frameworks, leverage expert knowledge, and monetize your data.

First, prioritize structured frameworks over ad-hoc experimentation. Matt Luizzi, VP of analytics at wearable tech company Whoop, explained that his team collects biometric data around the clock. They use Snowflake’s CoCo coding agent to accelerate development, from running A/B tests to analyzing results and proposing new features. The key breakthrough came when they stopped treating each agent interaction as a unique event. “We learned fast that context was everything,” Luizzi said, emphasizing the need to anchor agents in a semantic layer where information is organized and repeatable. His team now formalizes evaluation frameworks, allowing them to scale agent rollouts with confidence, mirroring the disciplined data architecture they’ve used for years.

Second, pair your agents with human experts who know the business cold. Madeleine Want, VP of data at Fanatics, oversees data engineering and machine learning for the company’s betting and gaming division. She noted that early success came not from powerful models alone, but from well-bounded domains where expert analysts could coach the agent. “Where we had expert analysts who understood the business domain top to bottom, they were able to coach the agent,” she said. This partnership reduced the required investment in context layers over time, while the accuracy of agent responses improved through scaled evaluation frameworks. As confidence grows, Fanatics is embedding agent APIs into third-party tools, letting users access data-driven insights directly within their existing workflows.

Third, think about turning your data into a revenue stream, not just a cost center. Sriram Sitaraman, CIO at Synopsys, observed that AI agents can now handle tasks once reserved for junior employees, such as running queries and generating reports. His team created specialized agents for finance and data center ticketing, evaluating them across three dimensions: quality, speed, and cost. For the first time, they found AI improved all three simultaneously. “Start with data , monetize your data using AI,” Sitaraman advised. He stressed that AI scales linearly with data volume, so the more you feed it, the smarter it becomes. However, he also cautioned against confusing automation with autonomy. “Do you want to automate a process or do you want to actually create an agent?” he asked, noting that each path has different cost structures and governance requirements. He urged professionals to carefully define the scope of each agent, as a sales ops agent can easily drift into becoming an analyst agent without clear boundaries.

The common thread across these stories is that successful AI agent deployment is not about the technology alone. It requires deliberate governance, a commitment to repeatability, and a clear-eyed assessment of what you want the agent to achieve. Without those foundations, the hype will remain just that , a promise unfulfilled.