In the summer of 2024, McDonald’s quietly pulled an AI ordering system out of more than 100 drive-thru locations. The system had been built with IBM, backed by real enterprise infrastructure, and tested for nearly three years. It still ended up going viral for the wrong reasons: a customer charged for 260 Chicken McNuggets, bacon added to an ice cream order, a bot arguing with itself while a family waited in line. McDonald’s called it a “thoughtful review.” Everyone else called it what it was, a very expensive, very public AI failure.
That story matters here because it isn’t the exception. It’s a preview. In June 2025, Gartner put a number on what McDonald’s had already lived through: more than 40% of agentic AI projects will be canceled before the end of 2027. Not paused. Not quietly descoped. Canceled, due to a combination of escalating costs, unclear business value, and inadequate risk controls.
If you’re currently running, funding, or about to greenlight an AI agent project inside your organization, this is the number that should be sitting on your desk right now, not because it’s a reason to stop, but because it’s a map of exactly where the landmines are buried. This piece walks through where that number comes from, why it keeps showing up across completely independent research, what actually kills these projects in practice, and what the organizations avoiding that fate are doing differently.
The Short Answer
Gartner’s prediction isn’t really about AI capability falling short. Every major study that followed it points to the same root causes: companies scoping projects too broadly, chasing hype instead of a defined workflow, and deploying agents with real access to real systems before governance, accountability, and rollback controls exist. The agents that survive past 2027 will not be the most technically impressive ones. They’ll be the ones with a narrow job, a clear owner, and a way to shut them off the moment something goes wrong.
The Prediction, in Plain English
Here’s exactly what Gartner said, stripped of the hedging. In a June 2025 press release, Gartner Senior Director Analyst Anushree Verma explained that most agentic AI projects today are still early-stage experiments, often driven by hype rather than a genuine business need, and that this hype tends to hide the real cost and complexity of running AI agents at scale. That gap between the pilot and the production version of the same project is exactly where most of these initiatives will stall out and eventually get shut down.
The prediction was built on a January 2025 poll of over 3,400 webinar attendees. Only 19% said their organization had made significant investment in agentic AI. Another 42% described their investment as conservative. The remaining 39% had made no investment at all or were still taking a wait-and-see approach. In other words, even the prediction itself was made before most companies had gotten far enough in to fail yet. The failures were still ahead of them.
Why “Agent” Might Not Mean What You Think It Means
Part of the coming failure wave has nothing to do with the technology and everything to do with what’s being sold as the technology. Gartner has been blunt about this: of the thousands of companies currently marketing themselves as agentic AI vendors, the firm estimates only around 130 offer something that genuinely qualifies. The rest are engaged in what the industry now calls agent washing, rebranding a chatbot, a rules-based automation script, or a simple AI assistant as an “agent” because the label sells better right now.
This matters more than it sounds like it should, because a project built on a rebranded chatbot will hit a ceiling almost immediately. It cannot plan multiple steps, cannot adapt to a changing situation, and cannot recover gracefully when something outside its narrow script happens. When that ceiling gets mistaken for a technology failure instead of a procurement failure, the project gets canceled for the wrong reason, and the lesson never actually gets learned.
The Data Says the Same Thing From Every Angle
Gartner’s number would be easy to dismiss as one analyst firm’s opinion if it stood alone. It doesn’t. Nearly every major study published since has landed on some version of the same story, even when the methodology, the sample, and the researchers had nothing to do with each other.
| Source | Finding |
|---|
| Gartner (June 2025) | Over 40% of agentic AI projects will be canceled by end of 2027 |
| MIT Project NANDA (2025) | 95% of enterprise generative AI pilots show no measurable financial impact |
| BCG and MIT Sloan (Nov 2025, 2,100 organizations) | Agentic AI adoption hit 35% in two years, the fastest of any AI wave, yet most deployments remain stuck at the pilot stage |
| Deloitte, State of AI in the Enterprise (2026) | Only about 1 in 5 organizations has a mature governance model for autonomous agents |
| Enterprise technology leader survey (March 2026, 650 organizations) | 78% have at least one agent pilot running, only 14% have scaled one to full production |
| Multiple 2026 security surveys (Cloud Security Alliance, AvePoint, Gravitee) | Between 65% and 88% of organizations report at least one AI-agent-related security incident in the past year |
Read across that table and a pattern jumps out immediately. Adoption is happening fast. Production, and the return on investment that’s supposed to come with it, is not keeping pace at anywhere near the same speed. MIT’s researchers, drawing on roughly 150 executive interviews and an analysis of 300 public AI deployments, described this as a “learning gap”: companies are buying tools that individuals can use well, but failing to build the institutional processes that let those tools actually change how the organization operates.
A Real One Already Happened
Back to McDonald’s, because the case is unusually well documented and unusually instructive. The partnership with IBM began in 2021, building on voice recognition technology McDonald’s had acquired two years earlier. Early internal figures reportedly put order accuracy around 85%, with roughly 1 in 5 orders needing a human to step in and fix something. That’s a number that looks almost acceptable on a slide in a steering committee meeting. It is not a number that survives contact with a drive-thru at rush hour.
What actually broke the system tells you more than the failure rate does. The AI was a standalone tool, disconnected from order history, inventory context, and the conversational thread of what had already been said in the same interaction. It couldn’t tell which car in a busy lot was actually speaking. Background noise, regional accents, and multiple people talking over each other, all normal conditions for a real drive-thru, were conditions the system had never been meaningfully tested against. And critically, the deployment tried to replace human judgment entirely rather than support it, so when the system got confused, there was no graceful fallback, just a wrong order and a viral video.
Every one of Gartner’s three named causes shows up in that story. Costs escalated as the pilot expanded to over 100 locations before the core problems were solved. Business value became unclear once staff had to spend as much time correcting the AI as they would have spent just taking the order themselves. And risk controls were inadequate in the most literal sense, there was no reliable way to catch an order going wrong before the customer drove away with it.
The Real Reasons Enterprise AI Agent Projects Die
Strip away the specifics of any one case study and the failure patterns across the research converge on a short, repeatable list.
Escalating costs that were never modeled honestly
Agentic AI projects routinely go over budget because the pilot is priced like a proof of concept and production is priced like enterprise software, and those are two very different price tags. A narrow pilot connecting to one clean data source looks nothing like the same agent connecting to a dozen real, messy systems of record at scale. Teams that don’t model that jump in advance get blindsided by it midway through the project, right around the point where canceling looks cheaper than continuing.
Unclear business value, dressed up as innovation
A striking amount of AI agent budget still goes toward sales and marketing use cases, the most visible and the easiest to demo in a boardroom, even though independent research has consistently found the strongest, most measurable returns showing up in back-office automation instead: the unglamorous work of processing, routing, and reconciling that never makes it into a keynote. Projects chasing visibility over measurable value tend to look impressive right up until someone asks what it actually saved or earned, and there’s no clean answer.
Risk controls that exist on paper, not in practice
This is where the security data becomes genuinely alarming. Multiple 2026 surveys, run independently by different research organizations, converge on the same uncomfortable range: somewhere between 65% and 88% of organizations have experienced at least one security incident tied to an AI agent in the past year, with data exposure and agents exceeding their intended permissions among the most common causes.
One widely cited 2026 study found that only around 1 in 5 organizations treats its AI agents as independent, identity-bearing entities at all, with most instead running agents on shared credentials meant for human employees. Another found that only around 1 in 14 organizations has a single named individual formally accountable for what an agent does. When something goes wrong, and the data says it usually does, there’s often no clear owner and no clean way to prove what happened.
A data problem hiding underneath all three
Underneath the cost overruns, the unclear ROI, and the governance gaps sits a more boring, more fixable issue: most organizations simply don’t have their data in a state that’s ready for an agent to reliably act on. Pilots tend to run against a clean, curated slice of data built specifically for testing. Production means messy, inconsistent, real-world data the agent was never actually validated against. That mismatch is consistently cited as the single biggest reason a working pilot fails to survive the jump to real deployment.
The 60% That Won’t Fail: What They’re Doing Differently
Gartner’s number implies something easy to miss in all the bad news: a majority of these projects will not be canceled. The organizations landing in that group tend to share a specific set of habits, and none of them require a bigger model or a bigger budget.
They scope narrow, on purpose. The agents that make it to production reliably started as a single, well-defined task with a measurable output, something like a document classifier or a routing agent, rather than an open-ended assistant meant to handle anything thrown at it. Broader, more ambitious agents consistently failed at scale because their edge cases multiplied faster than anyone could test for them.
They set a bar before they build, not after. Successful teams evaluate a candidate use case against a short, concrete checklist before writing a line of code: is the volume high enough to matter, does it follow a consistent, demonstrable pattern most of the time, and is there a clear metric to prove it worked. A use case that fails any of those three tends to produce an agent that demos well and delivers nothing measurable.
They treat scaling as a separate project from piloting. A pilot that works on 100 test cases with a 95% success rate is not the same system running against 10,000 real cases with every messy edge case included. Organizations that plan for that gap in advance, instead of assuming the pilot’s success rate will simply hold at scale, are the ones that avoid the expensive surprise later.
They put a name on the override switch. Every surviving deployment we found in the research had a specific, accountable owner for what the agent is allowed to do, paired with a working way to intervene or shut it down the moment it drifts outside that scope. Governance written into a policy document that nobody operationalizes doesn’t count.
They buy before they build, wherever the workflow isn’t genuinely unique. Research following real enterprise deployments has found that purchased, vendor-integrated agent tools succeed roughly twice as often as fully custom-built systems, largely because vendors have already absorbed the hard lessons that internal teams are still discovering for the first time. Custom development still has a real place, particularly for anything built on proprietary data or a truly differentiating workflow, but it should be a deliberate choice, not a default. (For a deeper breakdown of exactly when building beats buying, see our guide on custom AI agents versus off-the-shelf AI tools.)
They expand scope only after stability, not enthusiasm. A narrow agent that has run cleanly for roughly 90 days is a reasonable candidate for a wider mandate. A narrow agent that looked good in its first two weeks is not. The successful organizations resist the pressure to expand scope the moment early results look promising, because that’s exactly the point where hidden edge cases haven’t shown up yet.
A Practical Pre-Launch Checklist
Before an AI agent touches a real system, real customer, or real dollar, run it through this list.
- Define the single job this agent does, in one sentence, with no “and also.”
- Confirm the volume and pattern consistency of the workflow, high enough to matter, consistent enough that most cases follow a demonstrable pattern.
- Name the metric that proves success, and confirm someone can actually measure it before launch, not invent it afterward.
- Audit the data the agent will actually touch in production, not the clean sample it was tested against.
- Name one accountable owner for the agent’s behavior, in writing, not “the AI team” as a group.
- Build the override and rollback path first, before the agent goes live, not after an incident forces the issue.
- Set the 90-day stability bar before any conversation about expanding scope happens.
- Model the production-scale cost, not the pilot-scale cost, before presenting a budget to leadership.
Signs Your Project Is Already Heading for the 40%
- Nobody can answer, in one sentence, what specific problem this agent solves.
- The vendor calls it “agentic” but can’t explain what happens when a step fails midway through a task.
- Success metrics were never defined before the pilot started.
- The agent runs on a shared login rather than its own identity and permission set.
- There’s no single person who owns what happens if the agent does something it shouldn’t.
- The team is already talking about scaling before the pilot has run cleanly for more than a few weeks.
If two or more of these are true right now, that’s not a reason to panic. It’s a reason to pause before the costs escalate any further.
The Bottom Line
Gartner’s 40% isn’t a verdict on whether AI agents work. It’s a verdict on how most organizations are currently choosing to deploy them: too broad, too fast, and without the accountability structure that any other piece of business-critical infrastructure would be required to have before going live. The technology McDonald’s and IBM used in 2021 wasn’t fundamentally different from what a well-scoped, well-governed deployment uses today. What was missing was the discipline around it.
The organizations that avoid becoming part of that 40% won’t be the ones with the most advanced model. They’ll be the ones who scoped a real problem, named a real owner, and built a real way to catch it before the customer, the regulator, or the internet does.
