Gartner's agentic AI cancellation forecast lands two years before its adoption forecast

Two Gartner dates sit inside the same argument, and the order they arrive in does most of the work. The forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027 ^[4] runs two years ahead of the forecast that 70% of enterprises will deploy agentic AI as part of IT infrastructure operations by 2029 ^[3][13]. Read in sequence, the 70% is a survivors' number. Moving from under 5% in 2025 to 70% in 2029 is more than a fourteenfold increase across four years ^[12], and the same research house expects a large share of the attempts to be abandoned mid-climb.

The case for moving now rests on a failure mode described as getting worse, and there the record is thinner than the framing. Uptime Institute's 2025 Annual Outage Analysis puts nearly 40% of organizations at a major human-error outage within three years ^[1]; that figure describes a single measured point, not a trend line. The directional claim is narrower: the share caused specifically by staff not following established procedures has climbed year over year ^[2], with no figure published alongside it ^[14]. The record shows the direction climbing but not the size of the climb, so a business case built on acceleration is extrapolating past what has been measured.

Agent cost behaves unlike licence cost because it scales with reasoning rather than seats. Song Pang, CTO of NetBrain, writes that agentic reasoning revises and rechecks plans across several model calls per task, so the bill tracks how much thinking a resolution took ^[9]. That is the logic behind his prescription: a deterministic layer for routine diagnostic steps, reasoning models held back for judgment calls, hard caps on loop iterations and delegated subtasks, per-agent budget limits, and a metric of cost per resolved incident rather than raw tokens ^[10].

The trade-off hides in one phrase of that prescription, "as much as reliability allows" ^[10]. Cheaper models absorb the workload until they stop being trustworthy, and where a buyer draws that line is the thing actually being purchased. The harness has the same shape: continuous context, network intents to reconcile against and bounded guardrails are prerequisites, and Pang's own claim is that skipping any of them leaves an organization needing more human oversight, not less ^[8]. That places the spend on describing intended state, such as an application holding a latency target between two sites ^[11], ahead of the spend on models.

The CTO of a company selling agentic NetOps ^[6] is also the one specifying how much foundation you must buy first, and that foundation is not free. The useful reply is that two items in his Forbes Tech Council piece ^[15] are checkable without the vendor: whether documented network intents exist in your environment today, and whether a contract can be written against cost per resolved incident instead of token volume ^[10].

The idea that teams learn about inference cost only after deployment does not match this account. In Pang's telling, engineers at Cisco Live asked what it would cost almost without fail, immediately after asking how the agent decides what to look at ^[7]. Gartner's cancellation causes still lead with escalating costs ^[4], and its 2026 Hype Cycle names FinOps for agentic AI alongside agentic governance and security as signs of concern about economic sustainability ^[5]. The question is being asked before purchase and answered after it: a contracting gap, not a curiosity gap. This quarter's decision is whether the intent layer and the cost metric go into the pilot's terms; next quarter's consequence is whether anyone can tell a cost overrun from a scope expansion.