Consider the seemingly straightforward legal requirement in California: a landlord has precisely 21 days after a resident vacates to return a security deposit or furnish an itemized list of deductions. Miss that deadline in bad faith, and the penalty can escalate to twice the deposit amount. This critical window of compliance doesn’t differentiate between a letter composed by a human administrator or one generated by a sophisticated piece of software. In either scenario, accountability remains paramount, and someone—a human—will ultimately be held responsible for the outcome. This stark reality serves as a potent illustration of the operational challenges that companies embracing AI agents must confront head-on.

This scenario is far from abstract for me. At Get Covered, my company, we’ve integrated 16 distinct AI agents into our operations, overseeing more than 3 million rental units. When we embarked on building these agents, the technical hurdles, surprisingly, proved to be the more tractable problems. The truly complex and persistent question for each agent was, and continues to be: "Who does it report to?" This question delves into the heart of organizational structure, accountability, and operational governance in an age of increasing automation.

Most companies are on the cusp of grappling with this very same question, and alarmingly, few appear adequately prepared. Deloitte’s "2026 State of AI" survey casts a revealing light on this impending shift, reporting that 23% of companies are already moderately utilizing agentic AI, with a staggering 74% anticipating adoption within the next two years. This surge in adoption, while promising for efficiency and innovation, simultaneously introduces unprecedented operational complexities.

Compounding this optimistic outlook with a dose of caution, Gartner has issued a sobering prediction: over 40% of agentic AI projects are projected to be canceled by the end of 2027. This high failure rate isn’t necessarily a condemnation of AI technology itself, but rather a stark indictment of the lack of robust operational frameworks, clear accountability, and effective risk management strategies. Projects falter not because the AI isn’t intelligent enough, but because the human organizations deploying it lack the foresight and structure to manage its intelligence responsibly.

These compelling statistics collectively build an irrefutable case for managing AI agents with the same rigor and discipline applied to human staff in highly regulated positions. This entails establishing clear, written rules of engagement, meticulously defining their scope of action, assigning a human manager for oversight, and instituting regular performance reviews and audit trails. The operational blueprint for AI agents should mirror the best practices of human resource management, adapted for the unique characteristics of artificial intelligence.

Drawing from our journey at Get Covered, and my extensive experience in financial and insurance industries, I offer the following advice for decision-makers aiming to scale their AI agent initiatives successfully and sustainably.

Start with Work That Already Has a Rulebook

The most effective starting point for deploying AI agents involves tasks where the governing rules are not subjective or implicit, but explicitly codified. Think of statutes, contractual clauses, regulatory mandates, or precise insurance requirements embedded within vendor agreements. These tasks offer a clear, objective framework against which an agent’s performance can be measured and audited.

Consider the laborious task of managing vendor certificates of insurance (COIs). Across the property management industry, teams often dedicate approximately three hours per property each week to the tedious, yet critical, process of chasing these documents via email. This work is inherently repetitive, rule-bound, and completely specified. It’s a perfect candidate for an AI agent. The agent can be programmed to understand the exact requirements, identify missing or expired certificates, and initiate follow-up communications, all while adhering strictly to predefined parameters.

Assignments of this nature provide decision-makers with two significant advantages. Firstly, the AI agent possesses an objective benchmark against which to check its own output, minimizing subjective interpretations and potential errors. Secondly, your human team can audit the agent’s results without becoming entangled in debates over discretionary judgment calls. The outcome is either compliant with the rulebook or it isn’t, simplifying error detection and correction.

McKinsey’s most recent "State of AI" research reinforces this point, highlighting that the small cohort of companies realizing significant bottom-line returns from AI – roughly 6% of respondents – were nearly three times more likely to have fundamentally redesigned their workflows. This "redesign" isn’t merely about automating existing inefficiencies; it’s about re-evaluating and formalizing processes to make them AI-ready. In my experience, this foundational redesign commences with the act of writing down every rule. If a new human hire couldn’t follow your description of a decision on their very first day, an AI agent certainly won’t fare any better. Clear, unambiguous instructions are paramount.

Set Autonomy by Workflow

A common misconception among tech executives is to view AI system autonomy as a fixed attribute of the product. In practice, the appropriate level of autonomy for an AI agent should be fluid and dynamic, meticulously configured based on the specific workflow it interacts with and the potential impact of its actions.

At Get Covered, we meticulously configure each agent’s autonomy based on the nature of the task. For example, an AI agent sending a proactive reminder to a vendor whose certificate of insurance is set to expire next month can operate with full automation. The risk of error is low, and the consequences are easily rectifiable. Conversely, an AI agent tasked with drafting a response to a resident who has requested an emotional support animal is never permitted to operate autonomously. Such a task involves nuanced interpretation, potential legal implications, and significant emotional sensitivity, necessitating direct human review and approval.

A highly valuable exercise for any leadership team is to enumerate every workflow an AI agent will touch and, for each, pose a critical question: "If this action is wrong, how long until we detect the error, and what will be the cumulative cost of that error by the time it’s discovered?" Workflows characterized by quick detection and low potential cost can afford the agent a greater degree of operational freedom. Conversely, workflows with high potential cost or delayed detection demand a "human-in-the-loop" approach, with varying levels of oversight, from automated pre-approvals to mandatory human review before execution. This structured risk assessment allows for intelligent delegation, optimizing efficiency without sacrificing control or compliance.

Give Every Agent an Owner with a Name

Earlier in my career, particularly in financial services, a fundamental lesson I learned was that every automated credit decision, without exception, eventually required explanation—either to a customer seeking clarification or to a regulator scrutinizing compliance. The explanation "the model did it" was, and remains, an utterly unacceptable response.

This principle holds even greater weight in the era of AI agents, which now operate with more agency and across a broader spectrum of systems than any traditional underwriting model ever did. When an AI agent makes a mistake, the question isn’t if someone will be held accountable, but who.

Consequently, every single AI agent within Get Covered’s ecosystem is assigned a human owner. This individual bears ultimate responsibility, much like a manager overseeing a human team or a budget holder for a specific department. This designated owner is responsible for reviewing the agent’s error logs, making informed decisions on when to expand or restrict its autonomy, and ultimately, answering for its performance and any resulting outcomes. This human oversight ensures that the "buck stops somewhere," fostering a culture of accountability and continuous improvement for AI operations.

Furthermore, assigning clear ownership actively preempts a more insidious and quieter problem: the unintended consequences of uncoordinated AI deployments. When multiple teams independently launch AI agents without a centralized oversight mechanism, it’s highly probable that these agents will, at some point, contradict one another, leading to operational chaos, conflicting directives, and potentially damaging outcomes. A human owner, with a holistic view of their agent’s interactions, can identify and resolve such potential conflicts before they escalate. This ensures a cohesive and harmonized AI workforce, preventing internal friction and maximizing organizational effectiveness.

Where the Constraint Actually Sits

In the dynamic and rapidly evolving marketplace of today, it’s an undeniable truth that AI models will continue to advance, becoming ever more sophisticated and capable. However, a company’s internal clarity regarding who decides what, who is accountable, and how processes are governed will not improve organically. This crucial operational clarity requires deliberate, strategic intervention. Gartner’s forecast of significant AI project cancellations directly reflects this reality: the vast majority of agent projects that fail will do so not due to technological shortcomings, but because nobody adequately defined the value proposition, established clear metrics for success, or, critically, controlled the inherent risks. Both of these are fundamentally management problems, not purely technical ones.

For decision-makers navigating the complexities leading into 2027 and beyond, my advice is unequivocal: treat each AI agent with the same seriousness and operational diligence that you would afford a new employee stepping into a regulated and critical role. This means proactively writing down its operational rules, meticulously limiting its scope of action to prevent unintended consequences, and assigning it a dedicated human manager who is tasked with checking its work, discussing its performance, and taking ultimate responsibility for its actions.

Companies that embrace this proactive, operationally-centric approach will be uniquely positioned to progressively grant their AI agents greater responsibility and autonomy over time. This increased trust will be built upon a foundation of tangible evidence, robust audit trails, and a clear understanding of the agent’s capabilities and limitations, thereby justifying its expanded role. Conversely, businesses that choose to bypass this crucial operational rigor will inevitably spend their time in 2027 and beyond in a reactive state, scrambling to explain why their software executed an action that no human ever explicitly approved. The future of AI success hinges not just on technological prowess, but on astute operational leadership.

Rick Folgmann is chief operating officer at Get Covered, an insurance technology company that provides cost-effective property coverage with just a few clicks. He has more than 25 years of executive leadership experience in the financial and insurance industries.

Illustration: Dom Guzman