The current enterprise AI gold rush is defined by a frantic pursuit of autonomy. Every CTO is chasing the dream of the autonomous agent—a system that doesn't just suggest a solution but executes it without human oversight. Yet, for most organizations, this ambition hits a wall of risk. The fear of a rogue agent hallucinating a million-dollar procurement error or collapsing a supply chain keeps the most powerful tools locked in a permanent state of beta. This tension between the desire for speed and the necessity of control is where the industry is currently stuck.

The Four-Step Ladder of Earned Autonomy

At VB Transform 2026, Siobhán Mc Feeney, SVP at Target, provided a blueprint for breaking this deadlock. Rather than treating autonomy as a default setting, Target treats it as a performance-based reward. The company has implemented a rigorous four-stage ladder system where AI agents must prove their reliability before gaining more power. This framework begins at stage one with simple observation, where the agent monitors data without intervening. Once a baseline of accuracy is established, the agent moves to stage two: action suggestion, where it proposes a move to a human operator.

Only after consistent success does an agent reach stage three, allowing it to execute tasks within strictly defined guardrails. The final peak is stage four, end-to-end execution, though even here, human intervention remains a critical fail-safe. This is not a one-way street. If an agent's performance dips or the system experiences model drift—the gradual degradation of model accuracy over time—Target simply revokes the autonomy or shuts the service down entirely.

This systemic caution produces tangible business results. During a recent deployment in the Long Beach region, Target utilized a digital twin to predict inventory needs for men's shorts across three stores. The system flagged a massive anomaly, predicting that one specific store required six to seven times more inventory than its neighbors. While human analysts were initially skeptical of such a drastic variance, the AI had identified a critical geographic variable: that specific store was two miles closer to the beach than the others. The analysts trusted the system, the inventory was shipped, and every single unit was sold.

To prevent the chaos of overlapping AI tools, Target has also standardized the onboarding process. An engineer cannot simply request an agent; they must first define the problem and determine if the solution requires a complex orchestrator, a domain-specific agent, or if a simple tool would suffice. This registration process includes a mandatory check for existing solutions to prevent redundancy and requires a predefined trigger—whether it be an automated event, a manual engineer prompt, or a timer—along with a clear lineage path for every action the agent takes.

Shifting the Moat from Models to Architecture

While the broader market is obsessed with which frontier model is the most capable, Target is pivoting its focus. The company argues that the true competitive advantage, or moat, does not lie in the model itself but in the architecture, taxonomy, and data governance layers surrounding it. A frontier model is a powerful engine, but without a sophisticated chassis and steering mechanism, it is a liability in a complex merchandising supply chain that processes billions of data points.

This realization has led Target to adopt a gradient approach to model selection. Instead of deploying the most expensive, high-parameter model for every task, they match the model's complexity to the task's requirements. This optimizes the cost-to-efficiency ratio, ensuring that massive compute resources are reserved for the most intricate problems while simpler tasks are handled by leaner, faster models.

This philosophy extends into how Target views observability. Most companies measure AI success through latency or response speed, but Target tracks the trajectory of the agent. They monitor calibration states and whether the agent is adhering to its original intent over time. This level of transparency is not just about optimization; it is about recovery. When a system fails at 2 AM, the governance layer allows engineers to trace the agent's entire lineage—from its creation and configuration to the exact moment of execution—to pinpoint the failure and restore service rapidly.

Target is effectively moving from a phase of unconditional adoption to one of verified expansion. The core insight is that while every business wants agents, not every problem requires one. By building guardrails through security guidelines and registration protocols first, Target allows its builders to move fast without risking the stability of the core supply chain.

This shift redefines the role of the AI engineer. The builder is no longer just a coder writing prompts or fine-tuning weights; they have become a multi-layered manager. They must observe the agent, and then observe the human operator who is observing the agent. In an environment where AI and humans work in tandem, the ability to manage these subtle contextual nuances is becoming the most valuable technical skill in the enterprise.

For practitioners, the lesson is clear: the metric of success is shifting from runtime performance to intent achievement and drift measurement. By treating autonomy as a privilege to be earned rather than a feature to be toggled, companies can extract real business value from AI while keeping the risks firmly under control.