The reason the second track cannot be run like the first is that the two price failure differently. With chat-based tools the unit of value is a single task, and a weak draft simply gets edited [11]. An agent earns its keep by executing multistep processes across systems, and because its outputs become actions, a misfire books the wrong shipment or approves the wrong invoice [12]. Emily Lewis-Pinnell, who works on AI adoption at Evaila, argues in her Forbes Tech Council column that her two-track framework of broad enablement plus deep focus in one domain still holds, and that what changed is which track leads [13][19]; her point is that fluency can be democratized while autonomy has so far resisted the same treatment [14].
Gartner's forecast that more than 40% of agentic AI projects will be canceled by the end of 2027 cites escalating costs, unclear business value and inadequate risk controls [7]. Lewis-Pinnell reads all three as the signature of organizations handing out agents the way they handed out co-pilots, without redesigning the work underneath [15]. Read that way, the forecast is about rollout method rather than model capability, which is the more useful reading for anyone allocating next year's budget.
Two of the survey figures are worth restating in other units. PwC's 56% is roughly 2,494 chief executives saying the money has not arrived [1]. Deloitte's split works out to four organizations that put AI on top of unchanged work for every one reporting redesign at scale [2], a ratio about implementation choices rather than about tools.
This is self-report all the way down, and PwC's most quoted finding is an association rather than a cause: the 12% of CEOs reporting both revenue and cost gains are two to three times more likely to have embedded AI extensively into specific products, services and decisions [6]. That objection holds. What the surveys support is narrower than causation, namely that the firms reporting returns differ from the rest in a way an operator can check, and Lewis-Pinnell sets out the pattern (one bounded workflow, redesigned end to end, with the human role defined and the before and after measured) without naming a company or function that ran it [16]. The two 12% figures should also stay apart, since one comes from PwC's CEO panel and the other from Deloitte's sample of professionals, different instruments with different populations [3].
The sequencing explains why the gap persists: adoption dashboards measured the broad track adequately and then travelled into the agent era unchanged, which is how a program shows activity while the P&L shows nothing [17]. What a redesign forces next is different in kind, because Deloitte counts workflows and roles together [8], and a process rebuilt around what an agent reliably does raises a role question a co-pilot rollout never did [16]. Deloitte's own conclusion is the one to carry into the planning cycle, that adoption metrics and transformation metrics are not the same [10]. Teams that never measured the manual version will spend next year arguing about attribution instead of banking a number.
