Key takeaways
- Gartner projects more than 40% of agentic AI projects will be canceled by 2027, citing escalating costs, unclear business value, and inadequate risk controls.
- None of the three causes is model quality. The failures are governance and ownership failures.
- Every cause has a specific, buildable countermeasure: one workflow with one owner, cost attribution from day one, and a governance path designed before the pilot rather than after.
- The same market data shows record spending and the strongest growth in regulated industries — the cancellation figure describes how projects are run, not whether the technology works.
The number, and what it actually says
Gartner’s projection — that more than 40% of agentic AI projects will be canceled by end-2027 — has become the category’s most-quoted statistic. It is usually cited as evidence of a hype cycle collapsing. Read the stated causes, though, and it says something more specific and more useful:
- Escalating costs. Agent costs are hard to attribute and easier to lose. A shared API key, per-seat tool licenses, retries nobody counts, human review time nobody budgets — the invoice grows while the connection to business value stays fuzzy.
- Unclear business value. Pilots demonstrate capability (“look, it does the task”) without demonstrating outcome (“the workflow’s cycle time dropped 30% at equal or better quality”). Capability is not value; value needs a named workflow, an owner, and a baseline.
- Inadequate risk controls. The agent reaches systems nobody scoped, acts under credentials nobody audited, and produces no evidence trail a risk committee can accept. At review time, the honest answer to “who allowed this and how do you stop it?” is silence — and silence does not clear production.
Notice what is not on the list: model quality. No one is canceling agents because the models got worse. The technology has been good enough for the entire period these projects ran. What failed is everything around it.
Why pilots die at the governance review
The pattern in the canceled 40% is consistent enough to describe from the inside. A team demonstrates an agent on real data. It works. The demo earns a wider pilot. Somewhere between the pilot and production, someone asks the questions that production requires:
- Which agent did this? (There is no per-agent identity — a shared service account answers nothing.)
- What is it allowed to touch? (Whatever its developer could touch.)
- Who approves the consequential actions? (There is no gate; the agent just does them.)
- How do you stop it? (A deployment rollback, which stops code, not the agent.)
- Show me the record of what it did. (Application logs, unattributable and unapproved.)
None of these questions is answerable by the pilot, because none of them was designed for. The pilot proved the agent could do the work. The review asks whether the organization can own the agent doing the work. Those are different proofs, and the second one was never built.
The failures compound in a specific order: ungoverned agents make value unmeasurable (which cause was the agent’s contribution?), unmeasurable value makes cost unjustifiable, and unjustifiable cost plus unacceptable risk makes cancellation the rational decision. Gartner’s three causes are one cause seen three ways: nobody built the control layer.
The three countermeasures, in the order they pay off
One: anchor to a workflow, not to a technology. The projects that survive start with a named business workflow — KYC onboarding, claims triage, reconciliation breaks — with an accountable owner who feels the queue size, and one measurable outcome agreed before any build. “Agent for document review” has no baseline. “Cut claims-triage cycle time from 6 days to 2 at equal or better accuracy, measured against the same 200 cases” does. The workflow anchor is also what makes the value legible to the people who fund production.
Two: attribute cost from the first day. Cost observability is a governance feature, not a finance nicety. When every model call, tool invocation and human-review minute is stamped with the agent and the workflow that incurred it, cost per completed case is computable — and the budget conversation becomes a tradeoff discussion (“more authority for the agent in exchange for fewer review minutes”) instead of a renewal debate. Per-agent budgets written into the keys, with calls refused past the cap, keep the spend bounded by construction.
Three: design the governance path before the pilot, not after it. This is the counterintuitive one. Teams defer governance because it feels like friction, then discover that retrofitting identity, approvals and evidence onto a working pilot is the most expensive way to buy them. Built in from the start — the agent registers with its own identity, its calls pass a policy gateway, consequential actions gate to a named human, every action leaves a tamper-evident record — the governance path is simply how the agent works, and the production review reads evidence instead of halting at questions. The pilot-to-production transition becomes a formality because the proof was constructed along the way.
What this means for regulated enterprises specifically
Banks, insurers and payors face a version of this with higher stakes, because their review bar is not internal enthusiasm — it is regulators who are converging on identical expectations: identifiable agents, attributable actions, tested stop controls, auditable evidence. SR 26-2 leaves agentic AI formally out of scope while signaling an RFI; DORA treats agent activity as part of operational resilience; the EU AI Act imposes oversight and documentation duties on high-risk systems. The enterprises that instrument governance now will answer those regimes with evidence. The ones that deferred will answer them with projects that look a lot like Gartner’s 40%.
The market context sharpens the point: agentic AI spend in financial services is projected to grow from $7.8B in 2026 to $43.5B by 2031, and only 15% of enterprises have scaled multi-agent adoption. The money is committed and the scaling is rare. The 40% is what happens in between when governance is absent — and its mirror image is the opportunity: the projects that build the control layer are drawing from the same budget pool as the ones that die.
What we would do
Start with one workflow and its owner. Diagnose how it runs today — handoffs, exceptions, decisions, current numbers. Define the agent-human split and the controls the workflow needs before production: identity, approvals, evidence, stop conditions. Then build the smallest version that proves the outcome, with the governance path included from the first commit. That sequence is the whole difference between the canceled 40% and the scaled 15%. Our risk and control framework page maps the six control layers the review will test, and how an engagement runs shows the sequence from discovery to handover.
Sources
- Gartner, projection on agentic AI project cancellations by 2027 (causes: escalating costs, unclear business value, inadequate risk controls), 2026 — reported in Gartner press releases on agentic AI.
- KPMG, global agentic AI spend (~$50B in 2025), 2026.
- Mordor Intelligence, Agentic AI in Financial Services ($7.78B 2026 → $43.5B 2031, 41% CAGR), January 2026.
- Deloitte, survey of 501 US leaders (15% scaled multi-agent adoption; 16% processes prepared), August 2026.
- Federal Reserve/OCC/FDIC, SR 26-2 (April 2026); EU Regulation 2022/2554 (DORA); EU AI Act (Regulation 2024/1689).
Frequently asked questions
Why will 40% of agentic AI projects be canceled?
Gartner cites three causes: escalating costs, unclear business value, and inadequate risk controls. All three are governance and ownership problems rather than model capability problems — agents get canceled when no accountable owner tied them to a measurable outcome, when spend is unattributed, or when risk review cannot clear them for production.
How do you keep an agentic AI project out of the canceled 40%?
Anchor it to one named workflow with an accountable owner and a measurable outcome before building; attribute cost to the workflow from the first day; and design the governance path to production — identity, approvals, evidence, stop controls — as part of the pilot, not after it. Projects that define these three things early survive the review that kills the rest.
Does the 40% figure mean enterprises should not invest in agentic AI?
No. The same market data shows record spending — roughly $50B globally in 2025 — and the largest growth in regulated industries. The figure is a warning about how projects are run, not about whether the technology works. The projects that fail are the ones that skipped owner, outcome and governance, which are exactly the things that can be fixed early and cheaply.
Governing Autonomous AI Agents in Production
Moring provides the runtime control layer, identity verification, human-in-the-loop approvals, and audit records needed to deploy production-grade AI agents safely.
About the author
Balaji Nagaraj KumarChief AI Officer, Moring AI, Moring AISweden
Leads platform and delivery engineering at Moring AI, turning each engagement into a substrate the customer’s own team operates at handoff. His background spans developer-platform leadership at Volvo Cars and engineering at CERN. He writes about governance for agents that take actions in production systems.
LinkedIn profile