PwC reports that 88% of US business leaders plan to increase their AI budgets over the next twelve months, driven specifically by enthusiasm for agentic AI systems — AI that doesn’t just answer questions but takes actions, runs processes, and operates with a degree of autonomy. Deloitte, surveying the same universe of organisations, finds that only 11% of them are actively using agentic AI in production. And Gartner forecasts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% today.
Read separately, each of these numbers tells a coherent story. Read together, they describe something genuinely puzzling: an industry committing an enormous amount of money to something it has not yet demonstrated it can deploy at scale. Nine in ten leaders are increasing budgets; only one in ten has working systems. The market is accelerating into a gap.
This is unusual behaviour. Enterprise technology investment normally follows a well-worn pattern: pilot, prove, scale. You fund what works. The willingness to fund ahead of evidence — at this scale and speed — suggests that something different is happening with agentic AI, and it is worth understanding what.
Part of the answer is competitive anxiety. In enterprise technology, the penalty for being late to a platform shift is not just a missed efficiency gain — it is a structural disadvantage that can take years to unwind. The CIOs and transformation leads who missed the cloud transition in the early 2010s spent a decade catching up. The leaders who missed mobile spent years redesigning workflows that their competitors had already rebuilt from scratch. Agentic AI is being treated, rightly or wrongly, as a transition of comparable magnitude. You do not wait for proof when the cost of waiting might be irreversibility.
Part of the answer is that the platforms are already inside the door. Salesforce, Microsoft, SAP, ServiceNow, and Oracle are not asking enterprise customers to evaluate a new vendor — they are flipping a switch on software those customers already pay for. This week’s Salesforce Headless 360 announcement is a good illustration: the entire Salesforce platform is now accessible as a set of MCP tools and APIs that AI agents can operate directly, with no human in the loop. The Agentforce capability is bundled. The licence is already signed. The barrier to adoption is not procurement — it is internal readiness, which is a different and more tractable problem.
Part of the answer, too, is that the productivity case is genuinely compelling on paper, even if it has not yet translated cleanly into production. In software development specifically, where the measurement problem is more tractable than in most knowledge work, the evidence is striking. One recent study found that engineers working alongside coding agents achieved a 39% increase in weekly code merges. Another found a 73% productivity uplift for human-AI pairs over human-human pairs on equivalent tasks. These are not marginal gains. They are the kind of numbers that, if they hold at scale, would justify significant front-loaded investment.
The problem is the “if.” Deloitte’s finding that 40% of agentic AI projects are expected to fail by 2027 is not primarily about the AI — it is about the infrastructure underneath it. Legacy systems that were never designed for real-time API access, modular architectures, or secure identity management at the component level cannot easily support agents that need to take autonomous actions across multiple enterprise systems. The governance gap is equally real: only one in five organisations currently has a mature model for agentic AI oversight. You can buy the tool. You cannot easily buy the preconditions for using it well.
This is the gap that the spending numbers obscure. The 88% increasing budgets and the 11% in production are not contradictions — they are a time lag. The investment is going in now. The infrastructure modernisation, the governance frameworks, the identity and permissions architecture, the change management — those will take longer. The question for enterprise leaders is not whether to invest. It is whether they are investing in the right things: not just the agent layer, but the conditions under which agents can actually operate safely and at scale.
There is an analogy worth drawing here. When enterprises adopted ERP systems in the 1990s, the technology was not the hard part. The hard part was the process standardisation, the data cleaning, the organisational redesign that ERP implementation exposed and required. The companies that treated ERP as a technology project — buy the system, flip the switch — mostly failed. The companies that treated it as an operating model transformation succeeded. The investment pattern for agentic AI is tracking the same shape.
The gap between 88% and 11% will close. The Gartner forecast of 40% agent penetration by end of 2026 may prove optimistic — it almost certainly counts some chatbot wrappers alongside genuine autonomous systems — but the directional trend is not in doubt. What matters now is whether organisations are spending their AI budgets on the right side of the gap: not just on the agents, but on the conditions that will let agents actually work.
The ones that get this right will have a structural advantage. The ones that don’t will have spent a great deal of money on software that sits in the same implementation graveyard as the ERP systems they configured but never quite transformed.
