Why Over 40% of Agentic AI Projects May Fail

More than 40% of agentic AI projects may be cancelled by 2027, with cost, unclear value, and weak controls driving many failures. The central challenge lies in turning promising demonstrations into accountable, scalable operating models.

This episode explores how organisations can grant autonomy progressively while protecting performance, economics, and governance.

TLDR / At a Glance

• Organisational readiness over model capability
 • Demo-to-operating-model gap
 • Authority, access, and accountability
 • Five-stage Authority Ladder
 • Evidence-based increases in autonomy
 • Proportionate human oversight

A flashy agentic AI demo can make almost any workflow look solved, right up until it hits real data, real users, real risk, and real cost. We dig into Gartner’s headline prediction that more than 40% of agentic AI projects may be cancelled by 2027 and explain why that number is less interesting than the mechanisms behind it: escalating spend, fuzzy business value, and controls that never kept pace with the authority being granted.

Agentic AI succeeds when leaders choose suitable workflows, establish clear ownership, and expand authority only when performance and controls justify it.

The central shift is moving from “can the agent act?” to “may it act?” That question forces an operating model conversation: permissions and least privilege access, monitoring and logging, roll-back paths, escalation rules, and a named human owner who is accountable when the system takes action. We also challenge the sloppy use of the word “agentic”, where assistants and scripted automation get sold as autonomy, leaving teams to pay an autonomy premium while inheriting governance risk they did not design for.

To make this practical, we introduce the authority ladder: observe, advise, act with approval, act within limits, then higher autonomy under continuous monitoring. The goal is not maximum autonomy; it is the right level of authority for the workflow, earned through evidence that performance holds, controls hold, and unit economics work at volume. Along the way, we look at the kinds of workflows where agents already succeed, and why “human in the loop” only counts when the human has the information and power to say no in time.

If you’re building an agentic AI strategy, listen for the tests that kill weak projects early and the governance patterns that let strong ones scale. 

Subscribe for more, share the episode with a colleague who owns AI delivery, and leave a review: what workflow are you most tempted to automate, and that rung of authority has it truly earned?

We design and deploy autonomous agents that operate inside defined workflows. Built around your processes, integrated with your systems, and governed for reliability, they deliver operational leverage without increasing headcount.  Learn more here - Autonomous AI Agents - Kieran Gilmurray

Transcript

Gartner’s 40% Cancellation Warning

SPEAKER_00 0:00

Why over forty percent of agentic AI projects may fail. Gartner has put a number to a growing unease. It predicts that more than forty percent of agentic AI projects will be cancelled by the end of 2027. It is a forecast rather than a body count, and the reasons it gives escalating cost, unclear value, and weak controls are not failures of the technology. They are failures of the organization around it. This article answers the question of why so many agentic projects will stall. It argues that the gap is not between a model and a task, but between a demonstration and an operating model, and that the real question is not whether an agent can act, but whether the organization has decided it may. Blaming the agent. When an agentic AI project stalls, the blame usually lands on the agent. The model was not clever enough, the reasoning wandered, the tool calls misfired, so the answer must be a better agent or a different platform. It is worth reading Gartner's forecast closely, because the reasons it gives for the coming wave of cancellations point somewhere else. The projects it expects to be cut will fall to escalating cost, unclear business value, and inadequate controls. None of those is a property of the model. All of them are properties of the organization that deployed it. It also helps to be careful about what cancellation means. Stopping a project with poor economics, uncontrolled risk, or a simpler alternative is not a failure of nerve. It is portfolio governance working as intended. In many of the cases, Gartner anticipates the failure will not be the cancellation, but everything that led to it. A project chosen for its technology rather than its outcome, launched without a value case and left running until the cost became undeniable. The fair reading of the forecast is not that agents do not work. It is that organizations are approving autonomy before they are ready to operate it. This article takes that as its starting point. The interesting question is not whether an agent can complete a task, because in a growing number of settings it plainly can. The question is whether the organization has built the conditions under which it should be allowed to, repeatedly, economically and accountably, inside real work. That is a different problem, and it is the one most projects have not solved.

Demos Versus Operating Models

SPEAKER_00 2:27

The demo and the operating model. A demonstration and an operating model are not the same thing, and the distance between them is where most agentic value is won or lost. A demo proves that an agent can complete a task under conditions someone chose. Clean inputs, a bounded scenario, a friendly path through the systems. An operating model proves something much harder, that the organization can let the agent act again and again across messy real cases at a cost it can afford, within limits it can enforce with someone accountable when it acts. The first is a technical achievement, the second is an organizational one. Most projects are funded on the strength of the first and quietly assume the second will follow. It does not follow on its own. Between the impressive demo and a durable result, sit the workflow, the agent must run inside, the data and systems it must reach, the economics of every action it takes, the permissions that bound it, and the person who owns the outcome. Skip those and the demo never becomes an operating model, however good the agent looked on the day. What agentic quietly skips.

What “Agentic” Skips In Practice

SPEAKER_00 3:40

Part of the problem starts at the point of purchase, with a word that has lost its edges. There is no settled definition of an agent, and Gartner warns that a great deal of ordinary software, assistance, chatbots, and scripted automation is now being relabeled as agentic. It estimates that only a small fraction of the vendors marketing agentic AI have anything substantive behind the label. An organization that buys a category rather than defining a requirement can pay an autonomy premium for a system that is really an assistant and hand a governance team something that can act while it is presented as something that merely answers. The deeper issue is what the leap from can act to may act actually requires. An assistant that returns a wrong answer wastes a little time. An agent with system access can amend a record, initiate a payment, send a message to a customer, or trigger another workflow, and a wrong action there has consequences that do not reverse themselves. Allowing that safely is not a model setting. It calls for a redesigned workflow, reliable data and system access, identity and least privilege permissions, monitoring, logging, a way to roll an action back, and a named human owner accountable for the result. This is why it helps to state the reframe plainly. Autonomy is not a product feature. It is a decision about authority, access, accountability, and risk, and it belongs to the organization, not the vendor. An agent does not become safe to trust because it performed well in a demonstration. It becomes safe to trust when someone has decided what it is permitted to do, on what data, within what limits, and who answers for it when it acts. There is a simple test that filters a surprising number of weak projects before they start. What must this system be allowed to decide or do that an assistant, an analytics model, or a deterministic automation could not accomplish? If that cannot be answered precisely, the case for autonomy is weak, and the sensible move is to build the simpler thing. The authority ladder.

Climbing The Authority Ladder

SPEAKER_00 5:56

It helps to picture permitted autonomy as a ladder because autonomy is not a switch that is either off or on. On the lowest rung, an agent observes. It gathers and organizes information, but neither recommends nor acts. A rung up, it advises, putting a recommendation in front of a human who decides. Higher still, it acts with approval, preparing an action that a person must authorize. Higher again, it acts within limits, completing defined classes of action on its own while escalating anything outside them. At the top, it acts autonomously, running substantial workflows without routine approval, inside fixed permissions and continuous monitoring. Call it the authority ladder. The point of the ladder is that a system earns each rung. It is not placed on the top one because the demo was impressive. Autonomy rises only as evidence accumulates that the agent performs, that the controls hold, and that the economics work at real volume. Which rung is right for a given workflow is set by the stakes, the cost of an error, how easily it can be reversed, and what a regulator, customer, or board would expect. The most successful deployments in the independent evidence tend to sit a rung below the top, with the agent handling the great majority of cases and people reserved for the exceptions, and they outperform designs that demand human approval for everything. Read this way. The latter reframes the whole project. The question stops being can we make it autonomous and becomes how much authority has this system earned and what would let it earn more? That is a question about evidence and control, which are organizational rather than about model capability, which is increasingly a commodity. Where

Where Agents Actually Work Today

SPEAKER_00 7:42

autonomy earns its place. The pattern is clearest where agents are already working. Heathrow's passenger support agent answers questions about directions, facilities, waiting times, and flight status connected to live data and a knowledge base and resolves the large majority of chats without a handoff. The workflow is bounded, the answers are checkable, the volume is high, and an unresolved case can be passed to a person safely. It is a low rung of the latter used well, and that is precisely why it works. Higher autonomy cases share the same shape rather than a bolder model. A regional supermarket chain, in Stanford's sample of successful deployments, let an autonomous system combine inventory, sales, and supplier data to decide what to buy and when, and reported sharply lower waste and stockouts. The process was repetitive, measurable, connected to systems of record and forgiving of recoverable error. So the feedback loop was built into the work itself. Finance shows the same logic in cash application and invoice matching, where agents assemble the evidence and route the exceptions while approval and audit stay with people for anything material. What these have in common is not full autonomy but fitness for it, a defined start and end, measurable outcomes, reliable data, recoverable mistakes, and a named owner. The independent evidence is blunt about where the difficulty sits. In one careful study of successful deployments, the great majority of the hardest problems were not technical at all, but matters of change management, data quality and process redesign, and the durable advantage lived in the orchestration around the model rather than the model itself. Two cautions belong here. Several of these accounts are reported by the vendors or the companies involved rather than independently audited, so they are better read as evidence of what is possible than as market averages. And the strongest published productivity figures come from samples deliberately chosen for success, which shows that a gentic AI can excel in the right workflow without proving that it usually will. The forecast and its limits. The 40% figure deserves the same skepticism as any other single number. It is a prediction, not an observed rate, and the public forecast does not disclose a sample, a confidence interval, or a precise definition of an agentic project, which could mean anything from a departmental experiment to a multi-year program. The broad failure mechanisms Gartner names are more reliable than the exact proportion, and cancellation, as already noted, is often the right decision rather than proof of a broken agent. Symmetry matters in the other direction too. Sometimes the model really is the constraint, and no amount of operating model design will rescue a task it cannot do reliably. Sometimes a deterministic automation, a rules engine, or a retrieval assistant is simply the better answer, more predictable and far cheaper than an agent. Choosing the simpler tool is not a defeat for a gentic AAI. It is architecture following the problem, which is what good judgment looks like.

Oversight That Prevents Real Harm

SPEAKER_00 11:03

Accountability does not automate. As agents take on more execution, one might expect human involvement to fade. The evidence points the other way. The share of organizations that require a person to validate agent output has climbed steeply in a single year, and most heavy AI users still treat the system's output as a starting point and keep responsibility for the thinking. The near-term operating model is human-led and agent enabled, with autonomy applied where it is earned rather than everywhere at once. The catch is that human oversight can be real or merely nominal. A reviewer who is overloaded, underinformed, and given neither the time nor the authority to say no is not a control. They are a signature. Meaningful oversight means the person reviewing has the knowledge, the standing and the information to challenge what the agent proposes and a real route to stop it. Human in the loop is not a governance model until it specifies who the human is, what they see, what they may overrule, and whether they can intervene in time. Governance in turn should be proportionate rather than uniform. An advisory agent that drafts a summary and a transactional agent that can move money do not warrant the same control burden, and applying the heaviest regime to everything is its own way to kill useful projects under cost and delay. The discipline is to match the weight of control to the height of the rung. Light, touch, low down, then tight access, monitoring, limits and roll back as authority rises. Governance designed this way is not the enemy of autonomy. It is the thing that makes greater autonomy possible.

How Leaders Build Durable Autonomy

SPEAKER_00 12:49

What this means for leaders. The first shift is for many shallow pilots to a few deep ones. Executive pressure and platform marketing both push toward launching agents everywhere, which produces a portfolio of demonstrations and very little operating capability. The stronger move is to start from a workflow worth changing, bounded, measurable, data connected, and recoverable, with a named owner accountable for the outcome. Establish a baseline before building, compare the agent fairly against simpler alternatives, and decide in advance how freed capacity turns into throughput, service, or revenue. Then climb the authority ladder deliberately, granting each wrong only when evidence of performance, control, and economics justifies it, and building the controls before production rather than after the first incident. None of this is an argument against agents. It is an argument for granting them authority the way a serious organization grants it to anyone, in proportion to demonstrated competence inside clear limits, with someone accountable when the system acts. This concludes the article. You can also read this article on my LinkedIn page where I share regular insights on AI, strategy, and emerging technologies.