One statistic has done more work in agentic AI board presentations than any other. In June 2025, Gartner predicted that more than 40 percent of agentic AI projects would be canceled by the end of 2027. The figure gets quoted constantly, usually as a warning about hype. Almost nobody reads the three reasons Gartner gave for it.
Those reasons were escalating costs, unclear business value, and inadequate risk controls. Read that list again and notice what is missing. Model capability is not on it. Not one of the three named failure modes would be solved by a smarter model, a longer context window, or a better reasoning benchmark.
That absence is the whole story, and the industry keeps walking past it.
The Three Failure Modes Have Nothing to Do With Intelligence
Costs escalate when a project has no defined scope and the integration work turns out to be larger than the pilot suggested. Business value stays unclear when nobody agreed in advance what the agent was supposed to improve or how it would be measured. Risk controls stay inadequate when an autonomous system is given authority to act before anyone decides who can shut it down.
Every one of those is an operating discipline problem. They live in scoping, ownership, and data, and none of them is a machine learning problem.
Gartner made a second observation in the same research that deserves more attention than it got. Of the thousands of vendors claiming agentic capabilities, the firm estimated only around 130 were building anything that genuinely met the definition. Buyers are being sold reasoning, but the thing that determines whether reasoning produces a good decision sits somewhere else entirely.
An Agent Is Only as Good as the Questions It Can Answer About Your Environment
Think about what an agent has to establish before it can safely take action inside an enterprise IT estate. It needs to know what the service actually depends on, who owns it, what changed in the environment over the last 24 hours, and what breaks downstream if the action turns out to be wrong.
Those are not hard questions for a model. They are hard questions for an organization.
In most enterprises, that context is scattered. Some of it sits in a configuration database that the team stopped trusting two reorganizations ago. Some of it sits in the free text of old tickets, written by people who have since left. A meaningful amount of it exists only in the memory of the three engineers everyone calls when something serious breaks at 2am.
An agent cannot query the memory of your senior engineers. It can only query what you have written down, and most enterprises have written down far less than they think.
The Readiness Numbers Trail the Investment Numbers Badly
The data on this gap is unusually consistent across independent studies. Fivetran’s 2026 Agentic AI Readiness Index, based on a survey of 400 data professionals, found that only 15 percent of organizations are fully prepared to support agentic AI in production. Nearly 60 percent of the same group reported investing millions to tens of millions in the technology.
The most cited barrier in that study was not model performance, and it was not a shortage of talent. It was data quality and lineage, named by 42 percent of respondents.
The picture holds when you widen the lens beyond agents. A March 2026 study by Harvard Business Review Analytic Services with Cloudera found that only 7 percent of enterprises considered their data completely ready for AI adoption. A separate Cloudera survey of 1,270 IT leaders, published in April 2026, found that close to 80 percent say their AI initiatives are constrained by limited data access across environments.
Spending is running years ahead of the foundation it depends on. That is not a technology gap. It is a sequencing mistake, and it is being made at scale.
Confident Reasoning Over Stale Context Is More Dangerous Than No Automation
Here is the part that should worry operations leaders more than the cancellation rate.
When a human engineer makes a decision on bad information, the decision moves at human speed and usually carries visible hesitation. Somebody pauses to check the plan, or turns to a colleague before committing to it. That hesitation is itself a control, and it slows the mistake down long enough for another person to catch it.
An agent working from the same bad information behaves very differently. It acts in seconds, it acts across every system it has been given access to, and it produces a clear, well-argued explanation for why the action was correct. The confidence is not evidence of correctness. It is a property of the interface.
Autonomy multiplies the quality of the context you already have. If that context is accurate, autonomy compounds the value. If it is stale, autonomy compounds the damage, and it does so faster than your change advisory process can convene.
Human in the Loop Is Being Used as a Substitute for Good Context
The standard answer to this concern is to keep a human in the loop. I think that answer is doing far more reassurance work than it deserves.
Ask what the human in the loop is actually being asked to do. If the reviewer has to reconstruct the context from scratch to judge whether the agent’s recommendation is sound, then no time has been saved and no risk has been removed. A step has been added, and steps that feel redundant do not stay rigorous. Within a quarter, that approval turns into a click.
Approval is only a control when the approver has better information than the system asking for approval. In most agentic deployments today, they are both reading from the same incomplete record. Governance that depends on a tired engineer disagreeing with a confident machine at 3am is not governance.
The Unglamorous Prerequisites Nobody Wants to Fund
The work that actually determines whether agentic operations succeed is boring, slow, and almost impossible to demo. It is also easy enough to describe.
A system of record that is maintained rather than merely populated, with ownership fields that name a person who still works at the company. Telemetry correlated to services rather than to hosts, so that an agent reasoning about impact is reasoning about the business and not about a server rack. Change history that a machine can query rather than a spreadsheet a person has to interpret.
Rollback authority has to be defined in advance, with a named human on the override switch. A measured baseline has to be taken before deployment, so that six months later somebody can say whether the agent helped or simply generated activity.
None of that is exciting. All of it gets cut first when a project is sold on the strength of a demo, because the demo runs on clean sample data and the production environment does not.
I will be blunt about what this costs. Doing this properly delays your agentic rollout by two or three quarters while a competitor announces theirs. That is a genuinely uncomfortable position for any CIO to defend in a board meeting, and I do not want to pretend otherwise.
Where the Autonomy Line Should Sit Right Now
Given all of that, the sensible posture in 2026 is narrower than most roadmaps assume.
The starting point should be read-only agents. Let them investigate, correlate, summarize, and recommend, and let them do it for a few months while the team checks their reasoning against what actually happened. This is the cheapest possible way to audit the quality of your context, because every wrong recommendation is a free diagnostic of a gap in your data rather than an incident in production.
Move to recommend-then-act only where the action is reversible and the blast radius is bounded. Restarting a stateless service is a different category of decision from modifying a firewall rule, andtreating them as one class of automation is how organizations end up in the 40 percent.
Reserve full autonomy for the narrow set of actions where the required context is machine verifiable at the moment of execution. That set is smaller than vendors suggest and larger than sceptics admit, and the honest work is figuring out where its edges sit in your specific environment.
The Next Two Years Will Reward Boring Work
The cancellation wave that Gartner forecast is already underway, and it will be read as a verdict on the technology. That reading will be wrong. The agents that get switched off in 2027 will not have failed because they could not reason. They will have failed because they were reasoning about an environment nobody had bothered to describe accurately.
The organizations still running agents in 2028 will not be the ones that bought the best model. They will be the ones that spent 2026 doing the unglamorous work of making their own environment legible to a machine. That work has no launch event and no press release attached to it.
It is, however, the only part of this that a competitor cannot buy their way past.
