40% of enterprise apps will have embedded AI agents by the end of 2026. Gartner’s forecast has been widely shared as a signal of inevitable adoption - a technology wave cresting fast. What gets mentioned less often is the second Gartner number from the same research period: more than 40% of agentic AI projects will be canceled before the end of 2027.
Two 40% figures. No contradiction. This is exactly what a technology cycle at peak hype looks like.
Why 40% of Projects Get Canceled
Gartner isn’t forecasting technical failure. They’re forecasting organizational failure.
Three patterns dominate the cancellation picture:
Over-scoping from day one. Organizations try to automate complex, judgment-intensive processes before they have sufficient data quality or infrastructure. The agent produces unreliable output, trust collapses, project shuts down.
Skipping guardrails investment. Deploying agents without human escalation paths, quality monitoring, or output evaluation frameworks means errors surface too late to course-correct cheaply.
Brittle legacy integrations. Agents perform well in sandbox environments but break against real systems. Poor API quality and undocumented internal tools are a quiet killer behind many canceled projects.
IDC data puts a hard number on this: only 11-14% of enterprise AI agent pilots currently reach production at scale. The remaining 86-89% fail to generate durable value.
171% ROI - but only for some
Organizations that deploy agentic AI successfully report an average ROI of 171% (IDC research). U.S. enterprises average 192% - three times better than traditional automation approaches.
But this number belongs to a specific group: the ones who got scope and infrastructure right before scaling.
High-ROI implementations share three characteristics:
- Right scope: repetitive, high-volume tasks where correct/incorrect output is measurable
- Tool quality: reliable API integrations, clean data, clear documentation
- Ongoing operations discipline: treating agents as products with release cycles, monitoring, and regular improvement - not one-time deliverables
Projects that skip any of these, or treat deployment as the finish line, consistently report the opposite results.
86.4% of Marketers Use AI - but Most Are Still on the Surface
HubSpot’s 2026 State of Marketing report shows 86.4% of marketing teams are using AI in at least some areas. Content creation leads at 42.5% extensive use, followed by media creation (37.2%) and ad automation (34.1%).
These numbers look like high adoption. But look at what kind of AI is being used.
Content generation, ad optimization, and media creation are AI-as-tool use cases. The agent still needs human direction for every strategic decision. True agentic AI - systems that set their own sub-goals, plan multi-step sequences, execute across platforms, and self-adjust without continuous human input - is a different category.
The gap between “using AI tools” and “deploying production-ready AI agents” is the biggest gap being underestimated right now.
The Vietnam and APAC Angle
Asia Pacific is the fastest-growing region in the agentic AI market, driven by government AI initiatives, enterprise deployments in BFSI and telecom, and rapid cloud infrastructure expansion. Vietnam sits in the middle of this trajectory.
The strategic advantage for late-moving markets is timing. Vietnam’s tech and marketing sector is entering the agentic AI adoption curve with a full map of what failure looks like. There’s no requirement to repeat the 86-89% failure pattern that early-adopting enterprises went through.
But APAC markets including Vietnam face specific constraints that global failure benchmarks don’t fully price in:
- Data quality is inconsistent across many organizations - a dirty data pipeline is the single biggest hidden risk for agent performance
- Legacy systems are common in established businesses - integration complexity is a real cost, not a theoretical one
- AI operations talent is scarce - deploying an agent is far easier than maintaining and improving one over 12 months
All of these factors point in the same direction: scope smaller, measure earlier, don’t treat deployment as done.
How to Stay Out of the 40%
No formula is absolute. But Gartner and IDC converge on a framework:
- Define the right task first: repetitive, high-volume, with measurable ground truth. Success has to be objectively checkable.
- Build infrastructure before scaling: logging, monitoring, human-in-the-loop for edge cases. Slower deployment but sustainable operations.
- Treat as product, not project: an AI agent is not a deliverable. It needs ongoing development, iteration, and occasional rollback - the same discipline as any production software.
Gartner’s warning extends beyond 2026: organizations heading into 2027 without a clear AI risk management framework - regardless of how many agents they’ve deployed - will form the bulk of the canceled cohort.
40% of apps will have AI agents. 40% of projects will be canceled. Where you land in those two numbers is a question that needs answering before the next pilot kicks off.
NateCue's Take
The failure pattern I see most often - in Vietnam and across Asia - is organizations trying to automate too much, too fast, too early. An agency wants AI to "fully replace" account managers. A SaaS startup wants to "agentify the entire pipeline" in Q1. These are exactly the over-scoped projects Gartner flags as cancellation-bound. The 171% ROI benchmark is real. But it belongs exclusively to teams that scoped precisely: repetitive, high-volume, measurable tasks with clean data pipelines. For markets like Vietnam where data quality is still inconsistent and legacy systems are common, the right starting scope is even smaller than global benchmarks suggest. The winner in the agentic AI race won't be whoever deploys the most agents. It'll be whoever defines the right task first.