A pattern is becoming visible in enterprise AI adoption: organisations are deploying AI broadly and mostly at the periphery. They are using it to write internal communications, summarize meetings, help employees navigate documents, and assist customer service agents. These are real applications with real users, and by many metrics they represent genuine progress. They are also, by most definitions, not where the significant value is.
Drawing on McKinsey research, InfoWorld's analysis puts the pattern plainly: AI adoption is most common in IT, marketing, sales, and knowledge management. The core operational systems that actually determine financial performance, including inventory management, supply chain logistics, order processing, and financial transaction processing, remain mostly outside the AI transformation. McKinsey reports that only 39% of organisations have seen any enterprise-level earnings impact from AI so far, and most are still in experimentation or pilot mode.
The reasons are not difficult to understand. Deploying AI in a meeting summary tool carries almost no downside. If the output is incomplete, a human catches it and moves on. Deploying AI in inventory allocation or financial processing carries immediate financial risk: a bad decision in core systems shows up directly in margins, customer service levels, and regulatory compliance. The caution is justified. What is less justified is treating the peripheral wins as evidence of transformation, or assuming the harder work is just around the corner.
Agentic AI, where software agents can plan, execute, and self-correct on complex tasks, is where the enterprise AI conversation has been heading, and it amplifies the problem. Computer Weekly's reporting from Asia-Pacific documents how companies are building impressive agent demonstrations that never reach production. Organisations "build impressive agents in sandbox environments but fail to integrate them with core systems, data flows, and business workflows," one CTO observed. Avanade research puts 44% of organisations still at the proof-of-concept stage. "Innovation theatre" was the phrase used: fragmented pilots that look good in presentations but lack the orchestration, data readiness, and governance to function in live production.
The EY-Parthenon Growth Survey, published this week, shows what corporate leaders believe even if they are not yet acting on it. Seventy-eight percent believe AI will accelerate their organisation's growth rate. Sixty-three percent are currently using AI primarily for efficiency and productivity. Only 14% say they are using it to stay ahead of competitors, and 7% to diversify revenue streams. The trust gap is significant: only 34% trust AI outputs for optimizing pricing, 28% for new product development, and 27% for evaluating acquisitions.
Oracle's approach, as described to Computer Weekly this week, points in the direction enterprise AI appears to be heading: not chatbots at the edge, but agents woven into ERP systems that execute tasks on behalf of human workers with specific business objectives. Rather than a finance manager running reports, an agent monitors cash inflow metrics and surfaces optimized strategies, routing work to human collectors with generated call scripts when needed. Oracle now has more than 1,000 task-specific agents in its Fusion applications. Whether this represents the broader future or a well-resourced vendor's product roadmap remains to be seen. Notably, Oracle is explicit that these capabilities are reserved for its cloud SaaS customers: organisations still on legacy on-premise systems are left out.
The gap between AI in the meeting room and AI in the warehouse has a closing date. As enterprise software vendors integrate agents into core systems, as governance frameworks mature, and as organisations gain the operational experience to trust AI with higher-stakes decisions, the peripheral phase will give way to something more fundamental. The honest question is not whether that transition will happen, but how prepared organisations will be when it does, and who bears the cost of the adjustment when the efficiencies that seemed peripheral turn out to have been preparation for something much larger.