Retail AI has become a language problem. An associate tool, a merchant copilot, a delivery-routing system, and a shopping agent can all be called agentic. The label makes the products sound related while hiding a practical difference: each one holds a different amount of authority.
An associate tool answers a question. A merchant copilot assembles evidence. A routing system acts within a set of operating rules. A shopping agent may eventually make a recommendation directly to a customer. The operating question is simple: what may the system decide, and what must stay with a person?
That question changes the conversation when a demo becomes an operating plan. It explains why one use of AI can feel routine while another triggers concern, even when both systems draw on similar technology.
What the word hides
Target’s Store Companion is a useful place to start. Target rolled it out to team members across nearly 2,000 stores to answer process questions on handheld devices: how to sign up a guest for a Circle Card, for example, or how to restart a register during an outage. The scale is significant, and the scope is deliberately narrow.
Store Companion makes institutional knowledge easier to reach. It can shorten the time between a question and an answer, help a seasonal employee learn a process, and leave more time for a guest interaction. The associate still handles an exception and makes promises to the customer. The system supports the interaction; it does not own it.
Walmart’s merchant assistant, Wally, moves closer to the commercial decision. Walmart built it to work with proprietary merchandising data: to surface insights from large datasets, identify likely reasons an item is under- or overperforming, answer operational questions, and automate calculations and predictions. A merchant who wants to understand demand for protein-based bread can compare stores, markets, channels, items, and brands without assembling reports by hand.
That changes the cost of getting informed. The merchant still decides what to buy, where to place it, how much inventory risk to carry, and which supplier relationship deserves attention. Information arrives faster. Commercial accountability remains with the person in the role.
The moment a recommendation becomes an action
Walmart’s own roadmap shows where the boundary begins to move. The company says it is working toward letting Wally act on a merchant’s behalf within configurable guardrails. Elsewhere, it describes systems that help decide where inventory lives, coordinate delivery, and adapt to conditions across the network. It also says its Trend-to-Product capability can compress traditional fashion-production timelines by as much as 18 weeks.
These systems participate in execution. Delivery routing, inventory positioning, and production timing can be bounded by service levels, capacity, geography, approval bands, and exception rules. A retailer can define the objective, set an escalation point, and keep a person responsible for stopping the system.
The customer side shows the same tension. Walmart’s 2025 Retail Rewired survey found that 27% of respondents preferred AI recommendations to influencer recommendations, and 69% said shopping speed mattered when deciding where to shop. The same survey found that 46% were somewhat or very unlikely to let a digital assistant handle an entire shopping trip. People will use help with comparison, availability, and convenience. Many still want the final choice.
The same boundary appears inside a retail organization. A model may route an order well and still lack the mandate to change a price, shift a seasonal buy, deny a return from a high-value customer, or make a claim that affects a brand relationship. Capability is one test. The business also has to decide whether it will accept the consequence of delegation.
The work is deciding where authority belongs
Governance belongs in the operating plan. The National Institute of Standards and Technology’s AI Risk Management Framework places it across the lifecycle and asks organizations to define roles, human oversight, risk tolerance, and executive responsibility for deployment decisions.
A retail leadership team can make that concrete. Name the decision being delegated. Name the boundary around it. Decide what evidence will show that the system is performing as intended. Assign the exception when it fails.
There is no universal answer. Grocery replenishment and a luxury-client relationship carry different costs of error. A decision that can be reversed in minutes differs from one that changes margin, trust, inventory exposure, or a supplier relationship for a season. The point of a framework is to make those distinctions explicit before a system receives authority.
Retail has already used machine learning in forecasting, recommendations, fraud detection, and supply-chain work for years. Generative AI brings a new set of interfaces, and agentic systems bring a new possibility: software that completes parts of the work. That makes authority the central design decision.
A retailer can describe a mature deployment in plain language: what the system decides, what conditions limit it, and who owns the result. That answer will matter longer than an impressive demo.