The companies that succeed with AI agents are the ones that limit their autonomy

Gartner anticipates that more than 40% of ongoing agent AI projects will not make it to 2028; deployments that hold up deliberately restrict what the agent can do alone.

The dominant narrative is that the competitive advantage goes back to who deploys the most autonomous agents. An analysis published on August 22, 2026 defends the opposite: it is the organizations that limit the scope of action of their agents that achieve lasting results, governance having become the discriminating factor rather than the gross power.

The figures are severe. Gartner estimates that more than 40% of active AI projects will not survive until 2028. The average maturity of responsible AI devices is up to 2.3 out of 4, and only about 30% of organizations reach Level 3 or above on governance controls. Out of thousands of products sold as "agentics", nearly 130 would actually have autonomous capabilities. Almost two thirds of organizations cite security and risk as their main obstacle.

Four recurring reasons

The deployments that take place share four features: staff with a narrow perimeter assuming one responsibility rather than general assistants; human checkpoints placed before high stake actions and not after; full traceability of integrated decisions from the design stage; and sovereignty of data thought as an active mechanism of containment. The areas mentioned — financial reconciliation, compliance, industrial quality control, clinical documentation — are precisely those where autonomy creates a problem of responsibility.