Three Reasons Agentic AI Projects Get Canceled

In June 2025, Gartner predicted that more than 40 percent of agentic AI projects would be canceled by the end of 2027. The three causes it named were escalating costs, unclear business value, and inadequate risk controls.

None of those is a model problem. You cannot fix any of them with a better prompt or a bigger context window. They are the problems a company has when it deploys workers without managing them.

Read the three causes as management failures, and the fix becomes obvious: treat the AI agent as an AI employee, with a job, a scorecard, and a manager. This post takes each cause in turn.

The scale of the problem

Agent adoption is broad and shallow at the same time. McKinsey's 2025 State of AI survey of nearly 2,000 respondents found that 23 percent were scaling an agentic system somewhere in the enterprise and another 39 percent were experimenting, but no more than 10 percent were scaling agents in any single business function.

Capgemini's 2025 study of 1,500 executives put it more starkly: 2 percent had deployed agents at scale, 12 percent at partial scale, and the rest were piloting or exploring. Lots of agents. Very few jobs.

Cause one: escalating costs

Costs escalate when nobody owns the outcome. An agent allowed to run without a defined scope will keep running: more conversations, more retries, more tokens, more integrations bolted on for the edge cases nobody planned for. The bill grows, and nothing in the budget review says what the spend produced.

An employee does not have this problem, because an employee has a job with boundaries and a manager who notices when effort is not turning into results. Give the agent the same structure: a role, a set of approved actions, and a cost-per-outcome number that someone reviews every week. Cost stops escalating the moment it is tied to an outcome.

Cause two: unclear business value

Value is unclear when the project was scoped as a capability instead of a result. "Deploy an agent for customer inquiries" has no number in it. "Resolve routine account requests end to end and cut time to resolution in half" does.

The difference shows up in the research. MIT's 2025 GenAI Divide study found 95 percent of enterprise generative AI pilots delivered no measurable profit-and-loss impact, and that the successful minority chose tools that integrated deeply into specific workflows and put line managers, not only central AI teams, in charge of adoption. McKinsey found its high performers were nearly three times as likely to have fundamentally redesigned workflows around AI and pointed to KPI tracking as part of what separated them.

Take the AI Lead Conversion Agent. Its job is to respond to new leads in seconds, qualify fit and intent, and book the meeting or hand the sales-ready lead to a person. Speed to lead, qualification rate, and meetings booked are the KPIs. The value is clear because the job was written with the number in it.

Cause three: inadequate risk controls

Risk controls are inadequate when an agent has access without rules. It can read the CRM, so it reads everything. It can send messages, so it sends them at 11 p.m. It can take actions, and nobody can say afterward which ones it took or why.

This is why trust is falling even as adoption rises. Capgemini found that executive trust in fully autonomous agents dropped from 43 percent to 27 percent in a single year, with lack of transparency and limited understanding of agent capabilities among the reasons.

The answer is the one you already use for people: defined access, approval boundaries, a record of every action, and a clear handoff path for anything out of policy. The NIST AI Risk Management Framework organizes the discipline into four functions: Govern, Map, Measure, and Manage. An AI employee with role-based access, logged actions, and a manager reviewing its work covers all four without a separate compliance project.

All three causes have the same cure

Put the three fixes side by side: an owner who reviews cost against outcomes, a job written with a number in it, and access rules with a record of every action. That is not three initiatives. That is what it means to manage a workforce.

The companies in Gartner's surviving 60 percent will not be the ones with the most sophisticated agents. They will be the ones that gave their agents jobs and built the AI workforce management layer to run them: observability into what each agent did, outcome tracking against its KPIs, and controls that make the risk legible.

How Vida manages agents as a workforce

Vida is where businesses hire, manage, and measure AI agents as digital employees. Each agent is built around one of six jobs, given its own browser and scoped access to your systems, and assigned outcome KPIs. The AI workforce management layer gives the manager on your team observability across every conversation and action, version history as skills expand, and access controls, so cost, value, and risk are visible from the first week.

If you have an agent project that is drifting, bring it. We will help rewrite it as a job with a number and prove the outcome in a paid pilot on live traffic. Design Your AI Agent.