Target AI agents rely on governance, not just models, SVP says

Target AI agents are not a competitive edge by themselves, according to Siobhán Mc Feeney, a senior vice president at the retailer. Speaking at VB Transform 2026, Mc Feeney said Target’s advantage comes from the operating system around the models: architecture, taxonomy, data governance, security, observability and rules for how much freedom an agent gets.

Her comments cut against the current enterprise pattern of treating agent deployment as the milestone. Mc Feeney said companies need to decide first whether a problem needs an agent at all, or whether a conventional tool would do the job. The company did not disclose spending, revenue impact or savings tied to the work.

How does Target use AI agents?

Target is applying agents inside core retail systems, including supply chain, replenishment and demand forecasting, Mc Feeney said. The aim is familiar retail execution: having inventory in the right location at the right moment, supported by systems that can connect data signals and recommend or take action.

That work starts with classification. Target asks what problem is being addressed, what kind of agent is needed, and whether the proposed system should be an orchestrator, a broader agent, a domain-specific agent or something less autonomous. Mc Feeney said agents are registered and certified so teams can avoid duplicating tools that already exist.

Target also defines what starts an agent’s work, whether that is automation, a person, a schedule or another system trigger. Mc Feeney said the retailer wants a clear record from the creation of an agent through production use, so the company can see what happened if a system fails during off-hours.

Access is also scoped. Target evaluates what data, systems, tables and databases an agent can use. The company then monitors more than runtime and latency, Mc Feeney said, including whether the agent is doing what it was built to do, how well calibrated it is and whether performance is changing.

Why does Target say agents must earn autonomy?

Mc Feeney described Target’s autonomy model as a four-step ladder. Agents begin by observing without taking action. They can then move to making recommendations that require approval, then to acting inside set limits. At the highest level Target currently uses, an agent can run a process end to end, but a human remains involved.

Performance can move an agent up or down that ladder, Mc Feeney said. If a model drifts or stops meeting expectations, Target can reduce its permissions or remove it from service. The point is to make autonomy measurable rather than assumed.

Mc Feeney gave one inventory example from Long Beach. A digital-twin simulation forecast demand for men’s shorts across three Target stores and indicated that one location needed six to seven times more inventory than the other two. Analysts initially doubted the result, she said, but the store was less than two miles from the beach, while the other two were 10 to 12 miles inland. The recommendation was kept, and the inventory sold through.

The example is also a cost argument. Mc Feeney said different models fit different jobs, and frontier models can make sense for complex tasks involving very large data sets, such as merchandising supply chains. In other cases, she said, they may be too expensive for the expected benefit.

What changes for builders?

Mc Feeney said engineering teams need new habits as agents become part of software delivery and operations. Evaluation systems, registration and tracking become more important when teams are moving faster and when agents are participating in workflows.

She also said accountability remains with builders even when an agent has broad autonomy. Engineers are being asked to manage a mixed environment in which people supervise agents, agents produce work, and other people coach teams that are overseeing those systems. For Target, the claimed moat is that management layer around AI, rather than exclusive access to any one model.

This story draws on original reporting from VentureBeat.