AI Agents: 40% of Projects Could Be Canceled by 2027

The rise of AI agents is accelerating, but more than 40% of projects could be canceled by the end of 2027, according to Gartner. While 23% of companies are already using these agents and 74% plan to do so within two years (Deloitte), the difference will come down to management. Get Covered, which operates 16 agents across more than 3 million rental units, advocates for written rules, a clear scope, and a designated owner for each agent.

Project Cancellations Expected by 2027 and Governance Challenges

Gartner anticipates that more than 40% of AI agent projects will be canceled by the end of 2027. At the same time, it is reported that models will continue to improve, while internal clarity on decision-making distribution does not progress on its own. The interpretation of this projection is that many failures will stem from undefined value and unmanaged risk, both management issues. Organizations that fail to frame their agents could spend 2027 justifying unapproved software actions.

Significant Increase in Usage, But Needs Regulation

According to Deloitte, 23% of companies are already using AI agents moderately, and 74% plan to adopt them in the next two years. These dynamics, when viewed alongside projected cancellations, support the idea of treating agents like personnel subject to rules, with an explicit scope, an owner, and regular reviews. On the ground, Get Covered operates 16 AI agents across more than 3 million rental units and reports that defining the responsibility of each agent has been more demanding than resolving technical issues.

Start with Already Regulated and Measurable Tasks

Tasks where rules already exist—whether from law, contracts, or insurance requirements—are suitable candidates for agents. In property management, searching for insurance certificates takes about three hours per property per week through email exchanges, a tedious but fully specified task. In such cases, an agent can verify its output against objective criteria, and the team can audit the result without subjective judgment. McKinsey reports that the few companies (about 6% of respondents) that see significant financial gains are nearly three times more likely to have restructured their workflows. The indicated operational experience is that this restructuring begins with writing the rules: if a procedure is not understandable by a new employee on the first day, an agent will not follow it either.

Get Covered Assigns Appropriate Levels of Autonomy to Each Agent

Autonomy is not a fixed property: the same agent must operate with different degrees of freedom depending on the workflows. At Get Covered, autonomy levels are configured by use case: an automatic reminder can be sent to a vendor whose certificate expires next month, while a response regarding an emotional support animal is never automated. Teams are encouraged to inventory the targeted workflows and estimate, for each, the error detection time and its cost, in order to adjust the freedom granted. Practices inherited from automated credit remind us that every decision must be explainable; "the model did it" is not acceptable, especially as agents intervene more frequently and in more systems. Consequently, each agent at Get Covered has an owner who monitors error logs, adjusts autonomy, and assumes accountability, also limiting the risk of contradictions between agents launched by different teams.

Software Does Not Erase Legal Accountability

In California, a landlord has 21 days to return a deposit or detail deductions, under penalty of a sanction that can reach twice the deposit. This deadline applies regardless of whether a letter comes from a person or software. It is reported that, in all cases, a person will have to answer for it. This illustration serves as a reminder that software tools do not exempt companies from their legal responsibilities in executing tasks delegated to agents.