Gartner Says 40% of Agentic AI Projects Will Be Cancelled, But Enterprises Are Doubling Down Anyway

There is a recurring pattern in enterprise technology that nobody seems willing to learn from: a new capability emerges, executives pour money into it, projects fail at staggering rates, and then everyone acts surprised. We watched it happen with blockchain. We watched it happen with RPA. And now we’re watching it happen with agentic AI, except this time, the failure rate predictions are landing before most organizations have even moved past proof of concept.

Gartner dropped a prediction in mid-2025 that should have stopped boardrooms cold: over 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. But here’s the paradox that should concern every CIO: investment in agentic AI is accelerating at the same time cancellation rates are climbing. Enterprises aren’t just failing to learn the lesson, they’re actively doubling down while the evidence mounts against them.

The prediction nobody wants to believe

The Gartner forecast isn’t speculative hand-wringing. It’s based on observable patterns in how organizations are approaching agentic AI right now.

Anushree Verma, Senior Director Analyst at Gartner, put it bluntly in the press release accompanying the prediction: “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied. This can blind organizations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production.”

The word “misapplied” is doing a lot of work in that sentence. What Verma is describing isn’t a technology problem. It’s a strategy problem. Organizations are grabbing agentic AI and throwing it at problems without asking the fundamental question: does this use case actually benefit from autonomous decision-making, or would a well-designed automation workflow accomplish the same thing?

A January 2025 Gartner poll of 3,412 webinar attendees revealed the investment posture: 19% had made significant investments in agentic AI, 42% had made conservative investments, and 31% were taking a wait-and-see approach. That means over 60% of organizations had already committed capital to a technology where Gartner itself was predicting 40%+ failure rates.

Agent washing makes the problem worse

One of the most insidious dynamics Gartner identified is “agent washing”: vendors rebranding existing products like chatbots, RPA bots, and AI assistants as “agentic AI” without adding any genuine autonomous capabilities. Gartner estimated that only about 130 of the thousands of agentic AI vendors are real.

130 genuine vendors out of thousands. That ratio should terrify procurement teams.

I spend my days architecting AI-powered systems for enterprise environments, and the agent washing problem hits different when you’re the one evaluating vendor claims. The gap between what gets marketed as “agentic” and what actually exhibits autonomous reasoning, planning, and execution is enormous. A chatbot that follows a decision tree isn’t an agent. An RPA workflow with an LLM summarization step isn’t an agent. An API integration with a natural language interface isn’t an agent.

But try explaining that distinction to an executive who just saw a demo where the vendor’s “agent” appeared to handle a complex customer service scenario end-to-end. The demo was scripted. The production environment won’t be.

Forrester’s parallel warning

Gartner wasn’t alone in its pessimism. Forrester had published its own prediction that 75% of firms attempting to build their own agentic architectures would fail. The reasoning was even more specific: the systems are “too convoluted,” requiring diverse models, sophisticated RAG stacks, advanced data architectures, and niche expertise that most organizations simply don’t have.

Sam Dorison, co-founder and CEO of ReflexAI, wasn’t surprised by the Gartner numbers. In an interview with HR Brew, he offered a pointed diagnosis: “Where you get impact in professional settings, particularly enterprise settings, is through clarity and clear goals, and if you’re implementing AI, because you think AI can address everything, you do not have clear goals.”

That tracks with what I’ve observed across the standards bodies I participate in, including CoSAI and IETF AGNTCY. The organizations asking the most sophisticated questions about agent security and governance tend to be the ones with the most focused use cases. The organizations asking the vaguest questions tend to have the broadest, most ambitious agent deployment plans. That correlation is not coincidental.

The automation-that-amplifies trap

The failure pattern that keeps recurring goes something like this: an enterprise has a broken or poorly documented workflow. Rather than fixing the workflow, they layer an AI agent on top of it, assuming the agent’s intelligence will compensate for the process dysfunction. What actually happens is that the agent automates the dysfunction, faster and at greater scale than any human could.

As one analysis from Codal put it, organizations “think they can just parachute those agents in and then things will just happen. In reality, it’s not different than most technologies, and it’s definitely not different than hiring people.”

Based on industry estimates cited in the same analysis, over 60% of organizations still use at least one legacy system. These aren’t environments where you can drop in an autonomous agent and expect clean execution. They’re environments where a human’s institutional knowledge papers over gaps that a machine will stumble into, hard.

The appeal is understandable. Agentic AI promises customization without full rebuilds, flexibility as processes change, and orchestration across complexity. But those promises assume the underlying data is clean, the processes are documented, and the governance frameworks are in place. In most enterprises, at least two of those three conditions aren’t met.

The investment paradox explained

Why are organizations continuing to invest heavily despite warnings? Three forces are converging.

First, competitive fear. Executives are more afraid of missing the agentic AI wave than they are of losing money on failed projects. When your board sees competitors announcing agent deployments, the pressure to match their announcements, regardless of substance, is enormous.

Second, vendor pressure. The enterprise software market has collectively decided that agentic AI is the next revenue growth engine. Every major platform vendor is embedding agent capabilities into their products and pricing accordingly. Not investing in agentic AI increasingly means not upgrading your existing platforms.

Third, the pilot illusion. Small-scale agent proofs of concept are easy to make look impressive. An agent that handles 50 customer inquiries in a demo environment performs beautifully. That same agent handling 50,000 inquiries across 30 different product lines, integrating with five legacy systems, and complying with regulations in 12 jurisdictions is a completely different challenge. The gap between pilot and production is where the 40% cancellation rate lives.

Gartner itself acknowledged the tension. Despite the cancellation prediction, the firm also forecast that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, and that 33% of enterprise software applications will include agentic AI by the same year. The technology works. The problem is organizational, not technical.

What the survivors will do differently

Verma’s guidance was precise: “To get real value from agentic AI, organizations must focus on enterprise productivity, rather than just individual task augmentation. They can start by using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.”

That hierarchy matters more than most executives realize. Not every problem needs an agent. Some problems need automation. Some need assistants. The organizations that correctly match the capability to the use case will be in the 60% that survives. The organizations that deploy agents indiscriminately will be in the 40% that doesn’t.

From my work building AI-powered systems that serve over 170,000 users, I’ve learned that the governance question has to come before the technology question. Specifically: before greenlighting any agentic AI project, you need to answer “who is accountable when this agent acts wrong?” If you can’t answer that question with a specific person’s name, the project isn’t ready for deployment.

Here’s what that looks like in practice:

Define governance before autonomy. If you can’t govern it, don’t deploy it. Start with human-in-the-loop for all agent actions and remove the human only after demonstrated reliability. Map what each agent can do, read, write, execute, before it touches production data. Budget for agent monitoring and audit capabilities, not just agent development. And apply a rigorous buy-versus-build analysis to every agent project, because the Forrester data suggests that building from scratch fails at three times the rate of buying purpose-built solutions.

The technology hype cycle has a body count

Every transformative technology follows the same arc: hype, overinvestment, reality check, recalibration, and eventually sustainable value. We saw it with cloud computing. We saw it with big data. We’re living through it with agentic AI right now.

The difference this time is that the failure mode is more dangerous. A failed cloud migration left you with a higher hosting bill. A failed big data initiative left you with an expensive data lake nobody queried. A failed agentic AI deployment could leave you with autonomous systems making bad decisions at machine speed: decisions that affect customers, employees, and compliance posture in ways that are difficult to unwind.

The Gartner prediction isn’t a warning that agentic AI doesn’t work. It’s a warning that agentic AI deployed without discipline, governance, and focus will fail in ways that are expensive and, in some cases, harmful. The 60% that survive won’t be the ones with the biggest budgets or the most ambitious plans. They’ll be the ones that treated agent governance as a prerequisite, matched capabilities to problems honestly, and had the discipline to kill projects that weren’t delivering measurable value.

Forty percent failure seems high until you remember that Gartner estimated only 130 out of thousands of agentic AI vendors are real. The industry built the failure into the foundation.