Deloitte: 85% of agentic AI buyers sit in pilots, and the blocker is process, not models

The tightest pairing in the Deloitte data is between process readiness and production. Sixteen percent of leaders say their current processes are prepared for agentic adoption ^[8]; fifteen percent report scaled, orchestrated multi-agent deployments across customer service, IT and engineering ^[4]. One point apart ^[3]. That is not causation, but no question about models or tooling produces a match that close.

The self-reported blockers point the same way. Leaders named the absence of a unified, accessible data foundation (72%), an inability to trust and govern agents (70%), and the cost and complexity of integration (67%) ^[7]. None of those clear on signature. Each is a funded project with an owner, a schedule and a dependency on systems somebody else runs. Readiness scores across the supporting layers are correspondingly thin: vision and strategy 36%, technology infrastructure 34%, data foundation 32%, and risk, security and governance 26% ^[9].

Then there is the calendar. Seventy-four percent of leaders expect nearly half of business processes to be redesigned or rebuilt around agents by 2030 ^[10], while only 31% expect to do that redesign by 2028 ^[11]. That is a 43-point gap ^[2], and it sits precisely where budget authority runs out. Most planning cycles reach 2028. The ambition lives past them. In the interim, the default is to layer agents onto processes that already exist, which Deloitte says can work, while arguing that redesign should be a muscle companies develop rather than a step they skip ^[13].

The workforce numbers have the same shape. One in five organizations say their workforce is ready for agentic AI ^[6], yet 71% report work underway on baseline agent literacy and 65% on upskilling and reskilling ^[16]. Activity at 71% and readiness at 20% is a gap of roughly 51 points ^[4], and half of leaders say outright that their organizations are not investing enough in the workforce transformation that adoption requires ^[14]. Meanwhile 43% anticipate major job disruption as routine and structured tasks become autonomous ^[15]. The people expected to absorb the disruption are being trained by programmes their own leadership grades as underfunded.

Two-thirds of leaders say they are reevaluating their business models because of agentic AI, and only half claim a clear view of the operating model those agents would run inside ^[5]. That is the honest version of the 85% figure ^[1]: appetite is not the constraint, and neither is model capability. What is missing is a described target state that a process owner could build against.

For anyone approving next year's agent spend, the useful read is that token and training costs will land on the same budget as the licences ^[17], and that the 15% at scale did not get there by buying more agents. They got there having already done the unglamorous work the other 85% ^[1] have scheduled for after the current planning horizon. An agent budget without a redesign budget and a workforce budget reliably purchases what it has purchased so far, which is pilots.