The relentless pace of artificial intelligence development is exposing a critical flaw in how organizations deploy the technology: a broken learning loop. While AI promises rapid innovation, a significant number of projects stall between the pilot phase and full production, often because traditional methods of testing, failing, and operationalizing technology are ill-suited to the unique characteristics of AI systems.
The core issue, according to Mike Manos, CTO at Dun &. Bradstreet, is a mismatch between how organizations measure success and how AI actually performs. “Quarterly reporting and rigid KPIs assume progress is linear and outcomes happen at a defined point in time. AI fundamentally breaks those assumptions,” he said. This disconnect leads to a situation where linear expectations collide with the iterative, often unpredictable nature of AI, creating a distorted view of reality.
Model performance isn’t static; it drifts with changes in data, user behavior, and evolving policies. A KPI set at the beginning of a project can quickly become irrelevant, rewarding the wrong behaviors or failing to recognize genuine improvements. The problem is compounded by the time lag between when a KPI signals a problem and when a CIO can react, allowing issues to escalate across workflows. “By the time a quarterly metric flags a problem, the root cause has already compounded across workflows,” Manos explained.
This isn’t simply a technical challenge. While technical issues like hallucinations – where AI generates fabricated information – and model drift are significant, they aren’t the primary drivers of failure. Mark Baker, chief AI practitioner at Altimetrik, cautions that non-technical factors, such as a lack of governance, organizational unpreparedness, and low user adoption rates, are often more to blame. “But the AI moment – when the pressure to move quickly is combined with unprecedented user leverage – makes everything more acute,” Baker said.
The traditional approach of “fail fast” isn’t always the answer, either. While rapid experimentation is valuable, it can be risky to apply it indiscriminately to operational environments. A more sustainable approach, according to Manos, is “controlled fail fast” – iterating quickly within a trusted data boundary and monitored environment that can detect and correct unintended consequences immediately.
CIOs are increasingly focused on building observability into their AI systems from the outset. This involves designing pilots with deep telemetry to surface early signals of trouble – performance degradation, adoption friction, security vulnerabilities, or integration risks. Amit Basu, CIO and CISO at International Seaways, recommends three key practices: designing for observability, conducting pre-mortems and scenario testing, and deploying AI-assisted diagnostics.
However, simply identifying failures isn’t enough. It’s crucial to distinguish between “good” failures and “bad” failures. A good failure, Basu explains, happens early, is inexpensive, produces clear learning, and ultimately improves the system. A bad failure, occurs late in the process, repeats known mistakes, and damages trust, safety, or compliance.
Yuri Gubin, chief innovation officer at DataArt, emphasizes the importance of learning from failures. A good failure is one where the organization understands *why* it failed and can use that knowledge to improve future iterations. “Failure itself is not the problem. Lack of learning is,” Gubin said.
The key to closing the learning loop lies in shifting the focus from simply measuring output to measuring learning speed. As Basu points out, “In modern technology programs, learning speed is often more important than delivery speed.” This requires a willingness to embrace experimentation, invest in robust monitoring and diagnostics, and foster a culture of continuous improvement.
organizations need to move beyond rigid, templated KPIs that prioritize adherence to process over sound judgment. David Tyler, CEO of Outlier Technology Limited, notes that these frameworks can become a shield against accountability, preventing teams from overriding process with critical thinking when necessary.
successful AI deployment requires a fundamental shift in mindset. It’s not about trying to outrun AI, but about getting in sync with its iterative nature and building systems that are designed to learn from both successes and failures. As Derek Perry, CTO of Sparq, put it, “In an AI-accelerated loop, judgment has to be applied continuously as signals stream in.” This continuous assessment, coupled with proactive iteration refinement and a focus on learning, is the path to unlocking the true potential of AI.
