Why Enterprise AI Fails: It’s a Politics Problem, Not Tech

Large enterprises have spent billions deploying artificial intelligence tools, yet virtually none have fundamentally rebuilt their core operations around the technology. The barrier is no longer software capability. According to senior executives across consulting, financial services, and insurance, the bottleneck is organizational politics, workflow rigidity, and a persistent habit of bolting new models onto outdated processes.

Marc Warner, global chief technology officer at Accenture, estimates the number of major corporate or government processes truly reinvented with AI at the center remains close to zero. Speaking alongside technology leaders from Mastercard, AXA, and Mistral AI during a recent industry discussion, Warner argued that most organizations confine generative AI to peripheral tasks—marketing copy, meeting transcription, and basic coding assistance where returns are easily quantified.

Altering the fundamental economics of an industry requires targeting core business processes, a move fraught with operational risk. Warner compared the current phase of enterprise AI adoption to early 20th-century factories that initially used electricity solely for lighting before eventually redesigning entire production floors around electric motors. The technology was available decades before the structural imagination caught up.

Greg Ulrich, who leads technology initiatives at Mastercard, reinforced this historical parallel by warning against isolated automation. Speeding up a single step in an 80-step operational process does not yield efficiency; it merely creates a backlog at step two. Mastercard has pursued aggressive restructuring of its product development lifecycle, shrinking team sizes and accelerating turnaround times. But Ulrich emphasized that the AI agent itself is rarely the primary constraint. The actual work lies in remapping how value flows through the company.

This remapping transforms technology deployments into what Warner bluntly calls “politics projects.” While building the underlying systems remains a technical exercise, integrating them into a legacy corporation demands navigating entrenched departmental interests. Ulrich noted that Mastercard relies on direct CEO sponsorship, including weekly progress reviews with top executives, to force alignment. Even so, he acknowledged that no organization has fully solved the internal friction generated by massive structural change. Matthieu Caillat of AXA echoed the sentiment, noting that as model capabilities advance, the challenge shifts entirely toward bringing people along rather than managing code.

AXA illustrates both the ambition and the rapid obsolescence defining current enterprise AI strategies. The insurer initially cataloged roughly 400 potential use cases to build internal momentum before narrowing its focus to high-impact targets. Its operational plan for 2027 through 2029 aims to extract €500 million to €700 million annually in recurring pre-tax value from AI by the end of the decade, driven largely by unlocking insights buried within approximately 4 billion unstructured documents. Yet the pace of model advancement forces constant recalibration. Caillat admitted that AXA’s traditional “measure twice, cut once” engineering philosophy is obsolete because strategic conclusions expire before certainty arrives. Ulrich reported that custom consulting tools Mastercard built just a year ago are already considered quaint, requiring complete rebuilds.

Warner projects this relentless acceleration will continue until roughly 2030, when physical energy constraints in Western markets may begin throttling compute expansion.

The experimental nature of these deployments has also forced a pragmatic approach to failure. When autonomous agents produce errors or hallucinate outputs, assigning blame can paralyze innovation. Warner advised companies to avoid dwelling on fault, urging leadership to reward experimentation even when individual projects collapse. He outlined a general framework where business owners remain accountable for an agent’s economic value, technologists own its functionality, and chief information security officers govern its safety. However, he carved out strict exceptions for critical incidents, such as major data breaches, where immediate accountability is non-negotiable.

Marjorie Janiewicz of Mistral AI, drawing on a decade of cybersecurity experience, agreed that finger-pointing stalls progress. She shifted the focus from blame to control, arguing that the vital question for any enterprise is whether it actually governs the technology it adopts. Mistral’s strategy of selling open-weight models designed to run directly on customer infrastructure reflects a broader industry realization: relying on external API providers introduces latency, privacy risks, and dependency issues that complicate the deep workflow integration executives now recognize as essential.