Gartner has put two numbers on the table that seem to contradict each other. By 2026, 40% of enterprise applications will feature task-specific AI agents, up from less than 5% in 2025. In the same breath, Gartner predicts more than 40% of agentic AI projects will be canceled by 2027. Adoption and cancellation are climbing at the same time, inside the same market. That is not a contradiction. It is a description of what a technology looks like while it is still figuring out what it is for.
Two forecasts, one immature market
The adoption number is an eightfold jump in a single year. Gartner also projects agentic AI could eventually drive close to 30% of enterprise application software revenue by 2035 — over $450 billion, up from roughly 2% of revenue in 2025. That is the growth story vendors put in their decks.
The cancellation number is the part that gets left out of those decks. Gartner cites escalating costs, unclear business value, and inadequate risk controls as the three drivers behind failed projects. Anushree Verma, Senior Director Analyst at Gartner, put it directly: most agentic AI work right now is “early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.” A January 2025 Gartner poll of 3,412 webinar attendees found only 19% had made significant investments in agentic AI, while 31% were still watching from the sidelines.
Why so many agents never make it to production
Part of the problem is definitional. Gartner estimates that of the thousands of vendors marketing “agentic AI,” only about 130 actually meet the bar for an agent — a system that observes, reasons, and acts with meaningful autonomy. The rest is rebranded chatbots, robotic process automation, and assistants wearing a new label. Gartner calls this agent washing, and it inflates the adoption numbers while setting buyers up for disappointment when the “agent” turns out to be a workflow with extra steps.
Even genuine agent projects tend to die the same way. CTO Jeremy Ung of BlackLine told CIO.com that pilots look great in a controlled sandbox, then hit production traffic full of exceptions, variable documents, and unpredictable user behavior. “Scaling is where I see most of them fail,” he said. Cost models break down the same way: agentic systems consume tokens unpredictably across multi-step reasoning and retries, so the simple per-call pricing math teams used for a chatbot stops applying once an agent starts making its own decisions about how many steps a task requires.
MIT’s Project NANDA found a similar pattern at the broader generative AI level: despite $30–40 billion in enterprise spending, 95% of organizations saw no measurable return on their pilots, and just 5% of integrated deployments captured most of the value. The researchers concluded the divide wasn’t about model quality — it was about approach. Back-office automation, where scope is naturally bounded, produced the strongest returns.
The governance gap is the real story
McKinsey’s 2026 State of AI Trust survey, covering roughly 500 organizations, found security and risk concerns are now the top barrier to scaling agentic AI — ahead of regulatory uncertainty and technical limits, cited by nearly two-thirds of respondents. Only about 30% of organizations have reached a mature stage on the governance and control dimensions that autonomous systems require. Average responsible-AI maturity rose to 2.3 out of 4, up from 2.0 the year before, but technical capability is advancing roughly twice as fast as the oversight built to contain it.
That gap matters more for agents than it did for earlier generative AI tools. A chatbot that gives a wrong answer produces text a human can catch before acting on it. An agent that takes a wrong action has already acted. When mitigations get bolted on after the fact — extra approval steps, manual review queues, compliance sign-off — they tend to erase the efficiency gain the project was funded to deliver, and that is when finance teams start asking why the project still exists.
What separates the agents that survive
Across the reporting on both sides of Gartner’s numbers, the deployments that reach production and stay there share a consistent set of traits:
- Narrow, task-specific scope instead of a general-purpose agent asked to “handle customer support” or “manage procurement”
- Success defined as a measurable business outcome — cycle time, error rate, cost per transaction — not a vague productivity gain
- Human checkpoints placed before high-stakes or irreversible actions, designed into the architecture from the start rather than added after an incident
- Every decision logged for audit, so accountability doesn’t collapse the first time an agent makes a bad call
- A rollback path designed before launch, not improvised after something breaks in production
None of this is exotic. It is closer to how mature engineering teams already treat any system with real access to production data and real consequences for getting it wrong — bounded permissions, logging, staged rollout, a way back. The projects Gartner expects to survive to 2027 are the ones that were scoped like that from day one, not the ones racing to ship an agent because 40% of the market will have one.
Building for the 60% that lasts
The adoption curve and the cancellation curve are telling the same story from opposite ends: agentic AI is real enough that ignoring it is a competitive risk, and immature enough that deploying it carelessly is a budget risk. The companies that get past 2027 will be the ones that treated agent scope, governance, and rollback planning as part of the build — not as compliance overhead added after the pilot impressed someone in a demo. If your team is scoping an agentic AI project and wants engineering partners who build risk controls in from the start rather than bolting them on later, talk to Luby.