Progressive Automation: A Tiered Delegation Framework from Human Review to Full Autonomy

"Review Everything or Trust Everything" Is the Most Expensive False Choice

The most common trap when a company adopts AI agents is treating "human does it" and "AI does it" as an either/or switch: either every single action needs human approval, or the whole thing is handed off to run on its own. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, driven mainly by runaway costs, unclear business value, and inadequate risk controls, with most projects still stuck in hype-driven proofs of concept that never reach production. The flip side is clear: reviewing everything strips automation of its value, while trusting everything takes on uncontrollable risk before the system has been proven. Both extremes stall the project. The workable approach is to design delegation as a graded, verifiable, reversible staircase—starting with full human review, loosening the reins tier by tier as performance clears the bar, while keeping a brake that can drop back a level at any time.

Supervisory Postures and the Roots of Tiered Delegation

Before discussing tiers, settle the vocabulary. The relationship between humans and automation is usually split into three postures; IBM and aerospace human-AI teaming research both use these three layers to describe the intensity of oversight:

Notably, Article 14 of the EU AI Act requires high-risk AI to be "effectively overseen" by humans—but it explicitly does not mean a human must review every decision. It requires that humans be able to monitor, understand, intervene in, and halt the system at any time, and stay alert to "automation bias," the tendency to over-rely on AI output. That is exactly what tiered delegation solves: granting humans different intervention strengths for different actions.

Slicing autonomy into tiers is not new. As far back as 1978, Sheridan and Verplank proposed a ten-level automation scale for undersea teleoperators, ranging from "the human makes every decision" all the way to "the machine acts fully autonomously and does not even notify the human." In 2000, Parasuraman, Sheridan, and Wickens mapped degrees of automation onto four stages of information processing—acquisition, analysis, decision, and action—noting that the same system can have different autonomy at different stages. Post-2025 AI agent frameworks inherited this idea directly: a widely cited paper, "Levels of Autonomy for AI Agents," defines five tiers by the role the user plays—operator, collaborator, consultant, approver, and observer—transitioning from human-led to human-as-bystander. What enterprises should borrow is precisely this "same agent, different actions, different tiers" layered thinking.

An Actionable Five-Tier Delegation Ladder

Distilling the theory above into five tiers a company can apply directly. The core principle: the same AI agent can sit at different tiers for different actions—a read-only query may already be Level 4, while an action involving money or irreversibility stays at Level 1.

This ladder breaks the fuzzy question "should we trust the AI" into a series of verifiable, revocable small decisions—loosening the reins tier by tier is the cumulative result of a long build-up, not a single leap to full autonomy.

Promotion Criteria: Quantified Thresholds Decide When to Let Go

Promotion cannot rest on gut feel; it must be bound to measurable metrics. Define three types of threshold for each tier in advance, and allow a move up only when all are met:

• Accuracy threshold: a pass rate over a sufficient sample—for example, two consecutive weeks with no fewer than 200 cases—meeting that tier's requirement (Level 1 to Level 2 is often set at 98%, rising tier by tier to above 99.5% for Levels 3 and 4).
• Observation-period threshold: the target must hold over a stable stretch of time, so a single day's pretty numbers cannot buy delegation, and no major incident may occur during that window.
• Risk-ceiling threshold: set a ceiling on amount, scope of impact, or degree of irreversibility for each tier; actions above the ceiling stay at a lower tier and do not rise with the overall promotion.

Writing thresholds as explicit numbers rather than statements of principle turns promotion into an auditable decision: who, on what data basis, pushed which class of action from which tier to which—all logged—so decision-makers can justify the delegation to the board or auditors.

Fallback: The Circuit Breaker, Not Just the Kill Switch

A ladder's real sense of safety comes from its ability to go down, and here two things must be distinguished. A kill switch is manual: a human notices something wrong and terminates it. A circuit breaker fires automatically at the infrastructure layer—when an agent calls tools more than N times in M minutes, the same call fails three times in a row, or a single session exceeds a cost cap, it pauses and alerts automatically. That is exactly where the two diverge: if an agent falls into a retry loop while unattended and fires off a large volume of failing calls overnight, only an automatically triggered circuit breaker can stop the bleeding in time—the kill switch is useless because no one is awake. Set three automatic-downgrade triggers: a metric falling below threshold drops one tier and raises the review ratio; a single major incident triggers immediate downgrade and freezes that action class pending human review; an action touching an amount or impact ceiling is automatically routed to human review. Delegation is thus a two-way ladder you can climb up or down, not a one-shot gamble.

Putting It to Work: Delegate by Action Type

The most common mistake in practice is treating the entire AI agent as a single tier and delegating all at once. The correct approach delegates by "action type": read-only queries, drafting, external sending, writing to systems, and actions involving money or irreversibility each keep an independent tier and threshold, with an audit trail for every action recording the tier, confidence score, and whether a human intervened. Salesforce's Agentforce implements this through its Trust Layer: data masking, per-action audit logs, and approval and escalation paths set by risk and sensitivity. Governance maturity remains a widespread gap—a 2026 Deloitte survey found only about 21% of enterprises have mature governance for autonomous agents, while 74% expect to adopt agentic AI within two years. Build the delegation ladder and circuit breaker first, and you can race to full autonomy on low-risk actions while firmly binding high-risk actions to human review—balancing efficiency and safety.

How Nerdtechnic Can Help

Nerdtechnic helps enterprises embed a tiered delegation framework into existing processes: from inventorying action types and risk levels, defining each tier's quantified thresholds and observation periods, to building sampling review, audit trails, and automatic-downgrade circuit breakers, then accompanying the company in loosening the reins by the data after go-live. If your AI system is stuck in the dilemma of "afraid to fully trust it, yet unwilling to review everything," talk to Nerdtechnic about designing an automation ladder you can climb up or down.

References

• Gartner (via Forbes), "Why 40% of Agentic AI Projects May Be Canceled by 2027," 2026. Source
• European Union, "EU AI Act - Article 14: Human Oversight," 2024. Source
• MIT OpenCourseWare (citing Sheridan & Verplank, 1978), "16.422 Human Supervisory Control." Source
• Parasuraman, R., Sheridan, T. B., & Wickens, C. D., "A Model for Types and Levels of Human Interaction with Automation," IEEE Transactions on Systems, Man, and Cybernetics, 2000. Source
• Feng, K. J. et al., "Levels of Autonomy for AI Agents," arXiv:2506.12469, 2025. Source
• Salesforce, "Trust Layer | Agentforce Developer Guide," accessed 2026. Source
• Deloitte, "Agentic AI Is Scaling Faster Than Guardrails," 2026. Source