Company Success at Risk Gartner Warns Against Uniform Governance for All AI Agents

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

If organizations apply the same governance rules to all AI agents, regardless of their level of autonomy and area of deployment, this can lead to the failure of enterprise-wide AI agent initiatives, according to market research and advisory firm Gartner, Inc.

Gartner predicts that around 40 percent of companies will downgrade or decommission autonomous AI agents by 2027 because governance gaps will only become apparent after incidents in productive operation. Problems arise in particular when companies do not make a clear distinction between an agent's autonomy of action and the scope of its access rights.

"Many organizations view the governance of AI agents as an either-or proposition: either strictly regulated or fully trusted. This is the root cause of failure," said Shiva Varma, Senior Director Analyst at Gartner. "AI agents operate at different levels of autonomy and within different trust boundaries. Applying the same control mechanisms across the board to all agents often leads to two problems: Either simple agents are overly constrained, slowing implementation and encouraging shadow development, or more autonomous agents are not adequately controlled, leading to correspondingly higher operational, security and compliance risks."

Adapt Governance to the Level of AI Agents

To mitigate these risks, Gartner recommends applying a proportionate governance approach where AI agents are categorized into different levels of autonomy, with each level representing a different trust boundary and corresponding governance requirements.

Stage 1: Observe

Here, monitoring agents only have read access to defined data sources. The results are only visible to the requesting user. Common use cases include summarizing documents, retrieving data or knowledge, and explaining code. "At this level, governance should focus on basic controls such as limited data access, user authentication, usage logging, and basic functional and security testing," says Varma. "Since the risk is primarily limited to data disclosure and accuracy of results, controls should remain lean and focused."

Stage 2: Advice

Consulting agents generate recommendations, drafts or proposals for action, while humans review the results and take action manually. These agents have read-only access without write access and are used, for example, to draft emails, generate reports or code and support decisions. Although humans execute the decisions, advisory agents can influence judgment and cause downstream risks. "Governance for consultative agents should include all Level 1 controls and extend to quality of results and influence on decisions through accuracy and hallucination testing, domain-specific quality assessments, and user training on appropriate confidence levels," Varma said.

Level 3: Acting with authorization

This allows agents to perform tasks such as writing data, sending messages or changing configurations—but only after explicit approval by a human for each individual action. "At this level, human review is only effective if it remains a meaningful control measure," says Varma. "Without rigorous security testing, clear approval workflows with audit trails and agent-specific incident response procedures, approvals can lose quality under time pressure or due to approval fatigue. This would give a false sense of security while increasing the attack surface."

Level 4: Autonomous action

At the highest level of autonomy, agents carry out actions independently within defined framework conditions. Humans only check exceptions, test protocols and aggregated results instead of individual decisions. "When agents act autonomously, actions are performed at a scale and speed that can exceed human oversight," warns Varma. "Because responsibility for outcomes remains with the organization, this level requires the most rigorous governance, including continuous monitoring, enforced safety limits, mechanisms for rapid resets, protection circuits that halt agent operations when thresholds are exceeded, and clear accountability for agent behavior."

Smart-Factory

Access solutions for the Smart Factory - Industry 4.0