Why AI pilots stumble when they meet the realities of business

Scaling AI takes more than a successful model: businesses must align data, workflows and oversight, then track whether deployment improves speed, accuracy and costs.

Moving artificial intelligence from a successful proof of concept into day-to-day business use is often where the real difficulty begins. In an interview with Analytics Insight, Sukruth Pillarisetti, Straive’s senior vice-president for analytics and AI, argued that models that work neatly in a controlled environment can face very different pressures once they are exposed to the messier realities of enterprise operations, where data is scattered across systems, workflows overlap and exceptions are routine. IBM has made a similar point, saying that scaling AI is not simply a matter of building a model, but of managing it across complex production environments with the right infrastructure, governance and operating discipline.

Pillarisetti’s advice is to start with the business problem rather than the technology. In fraud management, for example, the aim should be to help investigators rank suspicious transactions more effectively, cut false positives and resolve cases faster. That requires the model, the data and the operational process to be designed together, rather than treated as separate pieces of a project. NIST’s AI Risk Management Framework reinforces that view by stressing lifecycle-wide risk management, including whether a system is meeting its intended purpose, how risks are handled and how deployed systems are monitored as conditions change.

He also drew a clear distinction between traditional machine learning and generative AI. According to Google Cloud’s guidance, predictive models remain well suited to structured data and defined decision tasks, while generative systems are better aligned with document-heavy, unstructured or conversational work. In practice, the two approaches can complement one another. A workflow may use machine learning to score or classify information and generative AI to read documents, extract meaning or assist analysts, with domain expertise providing the glue between them.

The final test, Pillarisetti suggested, is whether the system produces measurable operational gains. He pointed to processing time, decision accuracy, cost-to-serve and capacity as the most useful indicators of value. If those improvements do not materialise, the issue may lie not with the model itself but with how it has been integrated into the business. That emphasis on post-deployment scrutiny is echoed in NIST’s guidance, which says organisations should keep monitoring, feeding back and adjusting AI systems as risks, expectations and operating conditions evolve.

Source Reference Map

Inspired by headline at: ^[1]

Sources by paragraph:

Source: Fuse Wire Services