There is a quiet heresy running through Target SVP Siobhán Mc Feeney's message at VB Transform 2026, and it is the most refreshing thing we have heard from a retail technology leader in years. She is not claiming her team built a better model than OpenAI or Google. She is not pretending the algorithm is the product. Instead, she is pointing at the unglamorous scaffolding around the models and calling it the moat. That is a profound admission, because it flips the entire industry script. The market has spent two years obsessing over who has the smartest weights, while Target has been quietly building the boring, hard, essential layers of governance, lineage, and evaluation that make those weights useful. We should all be taking notes, because Mc Feeney is telling us that the future belongs to the disciplined, not the flashy.
Her most pointed challenge is aimed at the herd mentality sweeping every enterprise boardroom. She says bluntly that not everything needs an agent, and that autonomy must be earned, not granted. We love this because it cuts against the grain of every vendor pitch deck promising a "super agent" for every workflow. The practical takeaway for our readers is simple: before you build, ask whether the problem even requires an agent, or whether you are just renaming an API call. Target's four-level autonomy ladder, starting with observation and only moving to end-to-end execution after measurable proof, is a template for avoiding the graveyard of half-deployed pilots. Scale AWS Server Deployments Effortlessly with Stateless Model Context Protocol shows how even infrastructure choices are becoming more modular and stateless, which pairs naturally with Target's insistence that agents be registered, certified, and tracked from birth. The infrastructure conversation is no longer just about compute; it is about control.
The real insight here, though, is that Target is treating AI like a scientific instrument, not a magic trick. The Long Beach shorts story is the perfect illustration. A digital twin predicted one store needed seven times more inventory, the human analysts balked, and the system was right because it had factored in proximity to the beach. That is not hype. That is a measurable, defensible win. But Mc Feeney is careful to frame it as earned trust, not blind faith. She measures calibration, trajectory, and drift, not just latency. This is exactly the kind of rigor that separates a useful tool from a costly toy. In Bridging Retrieval and Action: A New Approach to AI Tasks, we see how connecting retrieval to action creates more complex workflows, and Target's approach of asking what type of agent is needed, orchestrator versus domain-specific, is the right first question. And when something does go wrong at 2 a.m., as she puts it, full lineage means they can recover instead of scramble.
What we would tell any reader asking about this is to stop shopping for models and start shopping for management layers. The moat is not the intelligence; it is the evaluation harness, the security framework, and the willingness to demote an agent that drifts. Mc Feeney's point about agents losing autonomy if they underperform is the single most important governance principle we have heard this year. It treats AI like an employee who has to pass a performance review, not a god you pray to. The one detail we will be watching is how Target scales this human-and-agent coaching skill set across its workforce, because Mc Feeney's description of builders observing agents building while also coaching humans is a genuinely new job description. If Target can make that cultural shift stick, it will not need a better model to win. It will just need to keep doing the unglamorous work better than everyone else. The takeaway to quote is simple: the model is a commodity; the discipline around it is the product. Watch whether your own organization is investing in that discipline, or just buying another API key.