New Delhi [India], October 7: A sharp gap in AI oversight between businesses reporting returns and those still experimenting has emerged in KPMG’s latest Global AI Pulse survey, drawing attention to the infrastructure and controls needed to run artificial intelligence in everyday operations.
Among organisations reporting established returns on investment, 86 per cent said they had a formal AI management layer spanning multiple functions or the entire enterprise. The corresponding figure for organisations still at the experimentation stage was 31 per cent.
KPMG India highlighted the Q3 2026 findings on October 5. The survey covered 2,131 senior leaders across 20 countries, territories and jurisdictions, with responses collected between July 23 and August 26. The findings show an association between AI maturity and formal oversight; they do not establish that governance alone produces financial returns.
For enterprises deploying large language models, the operational challenge extends beyond selecting a capable model. Systems must work with changing information, fluctuating demand and security threats while allowing organisations to understand how outputs are produced and when intervention is needed.
“AI is often discussed as if the model exists in isolation,” said Gopichand Talluri, an enterprise AI and data systems researcher. “In reality, an enterprise AI system depends on the quality of its data, the reliability of its infrastructure, how it is monitored, how it responds to change, and whether organizations can understand and govern its behavior over time.”
Data quality is a central concern. Business information is often spread across databases, cloud services, older applications and streaming platforms. Delayed updates, inconsistent records or weak access controls can undermine an AI application even when the underlying model performs well in testing.
The distinction becomes particularly consequential in fraud detection. A system evaluated against historical transactions may encounter different patterns once deployed. People attempting to evade detection can change their tactics or manipulate inputs, creating conditions that a controlled test may not capture.
Assessing such systems therefore involves more than measuring initial accuracy. It also requires examining the integrity of incoming data, performance under heavy workloads, resistance to manipulation and the ability to identify unexpected behaviour.
“In production, reliability is not a one-time test,” Talluri said. “It has to be continuously observed. Organizations need to understand when the environment around a model changes and when those changes begin affecting the quality or trustworthiness of its behavior.”
Ongoing monitoring can help teams identify changes in data or model behaviour that affect performance. Records of system activity can support investigations when outputs become unreliable, while defined escalation procedures can establish who reviews a problem and decides whether a system should continue operating.
These capabilities are relevant wherever automated recommendations influence consequential business decisions. An organisation needs visibility into the information supporting an output, the controls applied to it and the circumstances in which a person must take over.
The supporting data infrastructure presents another operational challenge. Fixed processing schedules and manual resource allocation may struggle with unpredictable workloads. Using operational metadata and workload history to guide processing offers a way to make these systems more responsive, although any gains in efficiency or reliability need to be demonstrated in practice.
Together, these issues put the readiness of the entire deployment under examination: whether information remains dependable, whether failures can be detected promptly and whether responsibility for corrective action is clear. For businesses expanding AI into routine workflows, those questions need answers before a successful demonstration becomes an operational dependency.
