In 2026, AI is no longer being piloted — it is being deployed. Deloitte measured that employee access to AI rose by 50% in 2025 across leading-edge organizations, according to its State of AI in the Enterprise 2026 survey, published on January 21, 2026 at Davos, covering 3,235 leaders across 24 countries. In the same wave, McKinsey published its AI trust maturity survey: trust is progressing, but gaps persist in strategy, governance, and risk management.
That is this year's paradox: the faster everyone accelerates, the more projects die. Gartner, in its 2026 Hype Cycle for Agentic AI released this summer, says it bluntly: the rapid progress of agentic AI is exceeded by hype and confusion. And Deloitte, on agentic deployments: close to three-quarters of enterprises plan to deploy them within two years, but only 21% say they have a mature model for agent governance.
Between massive acceleration and rare maturity lies a graveyard. BCG, in the fourth edition of its annual survey published on June 3, 2026, sums it up: only about 5% of organizations manage to reap substantial financial gains from AI. One organization in twenty, the same as the year before.
These 2026 figures confirm what the 2024 and 2025 editions were already pointing to: AI projects almost never die because of the technology. They die because of decisions — or decisions not taken — made before the first line of code. Those failed decisions leave traces. Here are the five signals that predict the failure of an AI project, visible from the very first meeting.
Signal 1: The phantom sponsor
The symptom. The project has the approval of "management", but nobody can name the executive who will put their budget and reputation behind it. The official sponsor attends steering committees, nods, then delegates everything.
Why it's fatal. BCG, in June 2026, says it plainly: executive engagement is one of the strongest predictors of AI maturity, and companies that make it a senior leadership priority, rather than a mere technology project, scale faster and generate more value. A phantom sponsor is the absence of the strongest predictor of maturity. At the first serious obstacle — and one will come — nobody will defend the project. BCG's "future-built" companies know this: their executives embody the daily use of AI instead of merely funding it.
The test question. "Who signs the decision to kill this project if it must be killed, and who pays for its success?" If two different names come up, or if the answer is silence, the signal is on.
The reflex. Require a named sponsor, with a written mandate and real budget authority, before any other spending. This is a framing-phase deliverable, not an end-of-project formality.
Signal 2: The fuzzy metric
The symptom. The project will "improve efficiency", "unlock the value of data", "accelerate time to market". Everyone nods. Nobody can write the sentence: by December 31, if X does not move by Y, the project has failed.
Why it's fatal. Gartner settled it in its June 25, 2025 press release, and the prediction remains the reference for agentic projects in 2026: over 40% of agentic AI projects will be canceled by the end of 2027, for three reasons — escalating costs, unclear business value, and inadequate risk controls. McKinsey, in its 2026 survey, points to the same gap on the strategy side, naming it first among the weaknesses. A project without a baseline metric cannot fail, and therefore cannot be arbitrated either: it drifts until its budget runs out. It is the most expensive failure because it is the slowest.
The test question. "Which profitability indicator, or which operational indicator, moves by how much, measured how, by when?" If the answer takes more than one sentence, it does not exist.
The reflex. A one-sentence definition of success, quantified, dated, with a baseline measured before launch. If the baseline does not exist, taking it is the project's first expense.
Signal 3: The unaudited data
The symptom. "We have plenty of data" is the argument that justifies the project. Nobody has checked whether that data is accessible, complete, current, and legally usable.
Why it's fatal. This is the best-documented cause, and Gartner's official forecast from February 2025 remains central for 2026: 60% of AI projects unsupported by AI-ready data will be abandoned by the end of 2026, according to Roxane Edjlali, senior principal analyst at Gartner. The math is simple: a project whose raw material is unusable is condemned the day it starts, for reasons a few weeks of auditing would have revealed. Deloitte 2026 confirms it from another angle: the gap between leading-edge organizations and the rest widens on data first.
The test question. "Who has inspected a real sample of this data, hands in the mud, this quarter?" A data audit is not a report: it is someone who opened the files.
The reflex. An availability and quality audit before any build commitment: access, completeness, freshness, compliance. If there is no ready data, the right project for the quarter is not a model — it is data preparation.
Signal 4: The team without the business
The symptom. The project lives in the data and IT department. The business is "involved" through a monthly steering committee. End users will discover the tool in training, three weeks before deployment.
Why it's fatal. BCG, in June 2026, is unambiguous: roughly 10% of the value derived from AI comes from the algorithms themselves, and another 20% from the technologies and data that enable them. The remaining 70% comes from rethinking the people component. When BCG's "future-built" companies plan to upskill more than 50% of their employees in AI, against 20% for laggards, they are stating a precise conviction: AI deployment is first a transformation of people, not a software rollout. Without business ownership, AI produces brilliant demos and deserted deployments.
The test question. "Who, in the core team, carries the pain of the current process, and who can shut the project down without escalation?" If the answer is "nobody, it's a data project", the signal is on.
The reflex. A mixed team with real parity: the process owner decides as much as the technical lead. The steering committee validates; it does not replace ownership.
Signal 5: The missing kill criteria
The symptom. The project has a roadmap, milestones, a budget. It has no written condition under which it stops. Ask the team: "In what scenario does this project stop?" The answer will be a wavering silence.
Why it's fatal. Deloitte 2026 made arbitrage discipline its central theme, summed up as "from ambition to activation". Its finding: leading-edge organizations do not have more projects — they stop more of them, and earlier. Gartner's 2026 Hype Cycle for Agentic AI recalls that rapid progress is exceeded by hype and confusion. When half of agentic projects will have been canceled by 2027, the survivors are those whose arbitrage was designed upstream. McKinsey 2026 points at the same flaw: the persistent gaps in strategy, governance, and risk management are precisely what prevents saying stop in time.
The test question. "Show me the page that says in which cases this project stops." One page. Not a paragraph in a slide deck.
The reflex. Written kill criteria before kickoff: performance thresholds, deadlines, costs, adoption. With a mandatory review date. Stopping a project on time is a success that gets counted.
What these five signals have in common
Read them again. None of them is about models, technology, or prompts. They are all about the governance of the decision: who decides, on which measure, with which data, with what mandate, and how far. The 2026 surveys all converge on this point.
McKinsey 2026 finds that maturity gaps are primarily about strategy, governance, and risk management. Deloitte 2026 shows that maturity separates organizations that move from ambition to activation, not the ones that go fastest. BCG 2026 calculates that 70% of the value comes from the people component, and that executive engagement is the number-one predictor of maturity. Gartner, in its 2026 Hype Cycle for Agentic AI, observes that technical maturity now outpaces organizational maturity.
The failure of AI projects is not a technical problem in disguise. It is an unresolved decision-architecture problem.
That is the core of the APV Framework: a five-phase methodology that evaluates an AI project across five axes — value, data, team, governance, and arbitrage capacity — before resources are committed. Each of the signals above maps to an axis that can be verified at framing time. You can run it internally on your own project portfolio, or certify your teams to apply it continuously.
AI projects will keep failing at scale. The only question is whether yours will fail before they cost you.
Sources
- Deloitte, State of AI in the Enterprise 2026, published on January 21, 2026 at Davos. Survey conducted from August to September 2025 among 3,235 leaders across 24 countries. Employee access to AI rose by 50% in 2025; close to three-quarters of enterprises plan to deploy agentic AI within two years, but only 21% report a mature model for agent governance. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
- McKinsey, State of AI trust in 2026: Shifting to the agentic era, 2026 AI Trust Maturity Survey. Verified via archive capture on April 22, 2026. Central thesis: trust maturity is progressing, but gaps persist in strategy, governance, and risk management. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/tech-forward/state-of-ai-trust-in-2026-shifting-to-the-agentic-era
- BCG, AI Transformation Is a Workforce Transformation, published on June 3, 2026, fourth annual survey. About 5% of organizations manage to reap substantial financial gains from AI; 70% of the value comes from rethinking the people component; executive engagement is one of the strongest predictors of maturity. https://www.bcg.com/publications/2026/ai-transformation-is-a-workforce-transformation
- Gartner, 2026 Hype Cycle for Agentic AI, released in the summer of 2026. The rapid progress of agentic AI is exceeded by hype and confusion. https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai
- Gartner, official press release, June 25, 2025: over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls. Still the reference for 2026. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Gartner, official press release, February 26, 2025: 60% of AI projects unsupported by AI-ready data will be abandoned by the end of 2026. Still the reference forecast on data readiness. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
Limits of this verification. The Deloitte, BCG, and Gartner pages were downloaded and their quotes were found word for word in their original language. The detailed body of the McKinsey 2026 report could not be extracted, because McKinsey blocks automated clients; that source is therefore cited for its central thesis only, not for a specific figure.