Target's Competitive Edge Lies in Infrastructure, Not Models

The Real Differentiation in Enterprise AI

When a digital twin recommended that one Target store in Long Beach stock six to seven times more men's shorts than two nearby locations, inventory analysts pushed back. The number looked wrong. But the system had identified a variable human planners missed: the store sat less than two miles from the beach, while the others were ten to twelve miles inland. The recommendation stood. The inventory moved.

That outcome, according to Target, illustrates why the retailer's advantage in AI has little to do with which models it runs. Siobhán Mc Feeney, senior vice president overseeing AI systems at Target, argues that the infrastructure surrounding models - the taxonomy, security layers, governance frameworks, and autonomy protocols - creates defensible value in ways that model selection alone cannot.

At DailyTechWire, we've tracked dozens of enterprise AI deployments across Asia and North America over the past eighteen months. The pattern Mc Feeney describes is increasingly common among operators who have moved past pilot phase: the hard problems in production aren't about parameter counts or training runs. They're about orchestration, observability, and knowing when not to deploy an agent.

Not Everything Needs an Agent

Target begins every potential agent deployment with a single question: what problem are we solving? That leads immediately to a second: does this problem require an agent at all?

Mc Feeney describes the discipline as "controversial" within an industry that has largely embraced agents as a default architecture. But Target has built a formal intake process. Teams must define the problem, identify whether an existing solution already addresses it, and specify what type of agent - if any - fits the use case. Is it an orchestrator? A domain-specific tool? A super agent? Or is the proposal actually just a workflow automation being mislabeled?

Once a team determines an agent is justified, they must register and certify it. That registration captures trigger conditions (automation, engineer input, timer), data access scope (which tables, which systems, which databases), and autonomy level. New agents start at the lowest tier and earn expanded permissions over time based on measurable performance.

The rationale is operational, not philosophical. When something breaks at two in the morning, Target wants full lineage from the agent's creation through every action it has taken. That traceability requires architecture decisions made at the beginning, not retrofitted after deployment.

A Four-Level Ladder for Autonomy

Target structures agent autonomy as a graduated ladder with four rungs. At the base level, an agent observes and logs but takes no action. At the second level, it suggests actions and waits for human approval. At the third, it acts within predefined guardrails. At the highest level currently in use, agents execute end-to-end workflows - but still with a human in the loop.

Autonomy is not granted by default. It is earned through consistent performance and can be revoked if an agent drifts or underperforms. Mc Feeney draws a parallel to human career progression: no one receives additional responsibility simply because they ask for it. They must demonstrate competence over time. Agents, she notes, have the advantage of being scientifically measurable. Accuracy, drift, goal attainment, latency - all can be quantified and tracked.

This structure allows builders to move faster because the boundaries are clear. If a team follows security guidelines, registers the agent, and operates within established guardrails, they can iterate without waiting for ad hoc approvals. And if something does go wrong, the lineage trail makes recovery faster and more reliable.

Models that drift are taken offline. Agents that fail to meet performance thresholds lose autonomy or are decommissioned. The system is designed to surface problems early, before they cascade into supply chain disruptions or customer-facing failures.

Choosing the Right Model for the Job

Mc Feeney emphasizes that different tasks require different "gradients" of models. Frontier models excel at complex, data-intensive problems - merchandising supply chains that involve billions of signals, for example. But those models carry higher inference costs, and in many scenarios a lighter-weight model performs adequately at a fraction of the expense.

The decision framework is straightforward: match the model's capability to the task's complexity, and ensure the cost-benefit calculation justifies the choice. Target's infrastructure allows teams to swap models in and out as requirements evolve, because the surrounding architecture - data pipelines, monitoring systems, governance layers - remains consistent.

This modularity is where the "moat" Mc Feeney describes becomes concrete. Models are increasingly commoditized; every enterprise has access to the same frontier LLMs from OpenAI, Anthropic, Google, and others. But the ability to route tasks to the appropriate model, track performance in production, and iterate on autonomy levels without rebuilding infrastructure from scratch - that remains scarce.

The Oldest Promise in Retail, at Scale

Target's AI systems increasingly connect signals across supply chain, replenishment, and demand forecasting. Mc Feeney frames this as delivering on retail's oldest promise: the right product, in the right place, at the right time. The difference now is scale and speed.

Agents are becoming part of the underlying architecture rather than one-off tools. They trigger replenishment orders, adjust forecasts based on weather and event data, and flag anomalies in inventory distribution. The Long Beach example - where proximity to the beach drove a sevenfold increase in shorts inventory at one location - illustrates the kind of signal that human analysts might miss but that a well-instrumented system surfaces automatically.

Mc Feeney describes these outcomes as "science" - mathematically significant and confidence-filling in ways that manual analysis cannot match. But she is careful to note that confidence is built over time, not assumed at launch. The first time an agent makes a counterintuitive recommendation, analysts question it. After the recommendation proves correct repeatedly, trust accumulates and autonomy expands.

Observability and Cultural Shift

Target measures every dimension of agent performance: intended function, calibration, trajectory, runtime, latency, and deviation from expected behavior. This creates transparency and allows continuous tuning. Agents that perform well earn more responsibility. Agents that drift or underperform are constrained or retired.

The monitoring layer also serves a cultural function. As teams work at speeds that were unimaginable even two years ago, accountability structures must keep pace. Builders are now managing hybrid workflows where human workers and AI systems operate side by side. The skill set required is different from traditional software engineering or operations management.

Mc Feeney describes the shift as "super exciting" but acknowledges the nuance involved. Builders must observe agents, coach humans who are observing agents, and maintain guardrails that allow both to function effectively. It is a layered responsibility that requires new training and new career pathways.

Why Infrastructure Becomes the Differentiator

The argument Mc Feeney makes is straightforward but runs counter to much of the current discourse around enterprise AI. The industry has focused heavily on model capabilities - context windows, reasoning benchmarks, multimodal performance. Those capabilities matter, but they are widely available. Every retailer, logistics operator, and supply chain manager can access the same models.

What is not widely available is the discipline to deploy agents selectively, the infrastructure to track them rigorously, and the governance frameworks to scale them safely. Target has invested in taxonomy, data governance, security layers, and autonomy protocols. That investment creates optionality: the ability to choose the right model for each task, to iterate quickly without breaking production systems, and to recover cleanly when something does go wrong.

For enterprises watching the AI deployment cycle mature across retail, logistics, and manufacturing, the lesson is clear. The moat is not in the models. It is in everything built around them.