In June 2025, Gartner put a specific number on a problem most enterprises were still describing in vague terms. More than 40% of agentic AI projects will be canceled before the end of 2027, the firm predicted, and the cause isn’t model capability. Gartner’s own explanation names three drivers: escalating costs, unclear business value, and inadequate risk controls. Anushree Verma, the Gartner analyst behind the forecast, was blunt about the underlying issue: most agentic AI projects right now are early-stage experiments driven by hype, and organizations are misapplying them without understanding the real cost and complexity of running an agent at scale.
More than a year later, that prediction hasn’t aged into irrelevance. It’s aged into confirmation. Enterprises kept deploying agents through 2025 and into 2026, and the pattern Gartner described, projects that clear a demo but never clear a governance review, has shown up repeatedly enough that it’s no longer a forecast. It’s a description of what’s already happening.
What “Agent Washing” Actually Costs
Part of Gartner’s warning is about honesty in the market itself. The firm estimates that only around 130 of the thousands of vendors marketing agentic AI products are building something genuinely agentic, a system that plans, acts, and adjusts across multiple steps with real autonomy. The rest are largely “agent washing”: chatbots, assistants, and robotic process automation tools rebranded with agentic language but without the underlying capability.
That distinction matters for buyers because it shapes where the real risk sits. A chatbot that answers questions poorly is a customer experience problem. A genuine agent that reads a data source, decides on an action, and executes it is a different category of risk entirely, because a bad decision doesn’t just produce a bad answer, it produces a bad action, taken with whatever access the agent was granted. Enterprises that evaluate the two the same way, on capability alone, consistently miss the part of the evaluation that actually determines whether the project survives to 2027: what governs what the agent is allowed to do once it’s live.
Governance Was Never the Afterthought It Got Treated As
The organizations Gartner expects to see canceling projects aren’t the ones that picked a weak model. They’re the ones that built agent access the way most companies build any new integration: quickly, with broad permissions, and with a plan to tighten things up once the pilot proved itself.
That sequencing is backwards for agentic systems specifically. An agent that can read a shared drive and send emails on someone’s behalf isn’t a feature waiting for a governance layer to be added later. It’s already an operational risk the moment it’s connected, whether or not anyone has written a policy for it yet. Security researchers have repeatedly demonstrated what happens in that gap: agents manipulated through hidden instructions in a document or web page into taking actions nobody authorized, using access nobody had gotten around to reviewing. The incidents differ in detail. The root cause is consistent: permissions were granted before anyone built a way to govern them.
This is why the enterprises most likely to be in Gartner’s surviving 60% aren’t necessarily the ones spending the most on AI. They’re the ones that treated agent governance as infrastructure from the start, not as a compliance step bolted on after a pilot succeeded. That distinction increasingly has a name in enterprise architecture conversations: an AI control plane, a centralized layer that governs which agents can access which systems, what actions they’re permitted to take, and what gets logged when they take them, rather than leaving each team to make that call independently for every agent it connects.
Where the Failure Pattern Actually Breaks
The practical version of this shows up around a specific piece of infrastructure many enterprises didn’t have on their radar two years ago: the protocol layer connecting AI agents to the tools and data they act on. Model Context Protocol, or MCP, has become the standard most AI vendors now use for exactly this kind of connection, and it’s grown fast enough that most organizations running more than a handful of agents are running MCP connections whether they’ve formally adopted the term or not.
The problem is scale without a control point. Five AI agents each with their own direct connections to internal systems is manageable through informal oversight. Fifty is not. Once an organization crosses that threshold, without a centralized way to see which agent can reach which system, security and compliance stop being answerable questions, and that’s precisely the “inadequate risk controls” category Gartner cites as a cancellation driver. This is the specific gap that products like the open-source Obot MCP Gateway are built to close: a single point that authenticates every agent connection, enforces which tools each one can actually invoke, and keeps an audit trail that can answer, after the fact, exactly what an agent did and why it was allowed to.
The Shakeout Isn’t a Reason to Slow Down
It’s tempting to read Gartner’s forecast as a caution against agentic AI itself. That’s the wrong lesson. The 40% figure isn’t a verdict on the technology. It’s a filter that will separate deployments built on real operational discipline from ones that were never going to survive contact with production regardless of how capable the underlying model was.
The agents still running in 2028 will be the ones with clear ownership, scoped permissions, and a control layer that can prove what happened when something went wrong. That’s not a technical nicety reserved for the largest enterprises. It’s the same deployment discipline companies already apply to every other system with access to real data and real consequences, applied to a category of software that, for the last two years, mostly hasn’t gotten it.
