Siobhán Mc Feeney, Senior Vice President at Target, told VB Transform 2026 that while AI models are important, the company’s competitive edge comes from the architecture, governance and operational systems built around those models. "There's a lot in it. That to us is the moat," she said.
Mc Feeney argued that Target approaches agents (AI systems with some degree of autonomy) deliberately: not every problem needs an agent, and autonomy is earned rather than granted by default.
When should an agent be built?
Target begins by asking whether an agent is the right solution for a given problem. Follow-up questions include:
- Is the problem suitable for an agent at all, or is it simply a tool?
- If an agent is needed, what type: an orchestrator, a super agent, or a domain-specific agent?
- Does an existing solution already solve the problem, avoiding duplication?
Agents must be registered and certified before deployment, Mc Feeney said, as part of a process that prevents unnecessary or redundant development.
Design, triggers and lineage
Agent design also answers operational questions: what triggers an agent (automation, an engineer, a timer), what needs to be tracked, and what the agent can access in terms of data, systems and databases. From the very beginning, Target insists on lineage — a traceable history from the agent’s creation through execution — so incidents can be diagnosed if something goes wrong.
Autonomy is staged: new agents typically start with base autonomy and gain more over time if they perform well. Continuous monitoring and observability are required because agents will not improve or reliably solve problems without ongoing evaluation.
"We measure everything: what it was intended to do, its calibration, its trajectory, not just runtime and latency," Mc Feeney said. That measurement creates transparency and enables gradual tuning.
Layers, cost-benefit and model choice
Target’s "layers of autonomy" create the foundation to scale and to invest in the right models for the right problems. Mc Feeney noted models have different "gradients": frontier models excel at complex tasks requiring vast data crunching — for example, heavy merchandising supply chains — but can be cost-prohibitive. Decisions always weigh cost versus benefit.
Example: digital twin and stocking autonomy
A digital-twin simulation predicted men’s shorts inventory across three Target stores in Long Beach this summer; one store returned a forecast calling for six to seven times more stock than the others. Analysts initially doubted the result, but the model had identified a factor they had not: that store was less than two miles from the beach, while the others were 10 to 12 miles inland. Analysts approved the recommendation and the extra inventory sold through.
Mc Feeney used this example to illustrate how agentic systems can be more mathematically significant and confidence-inspiring than human intuition alone — and how strong results allow agents to earn greater autonomy.
Four-level autonomy ladder
Target structures agent autonomy in four levels:
- Observation only — the agent collects data but takes no action.
- Suggestion — the agent recommends actions and waits for human approval.
- Action within defined guardrails — the agent acts but under preset constraints.
- End-to-end operation — the agent runs processes end-to-end, still with a human in the loop.
Agents can gain or lose autonomy based on performance; models that drift are removed from service, Mc Feeney said.
Cultural shifts and new skills
Even with high agent autonomy, builders remain accountable for failures. Teams are working at previously unimaginable speed, requiring evaluation harnesses, registration and tracking of agents. The workforce is being reshaped: engineers and builders need new skills to manage humans and AI systems side by side.
Mc Feeney described the evolving role as exciting: "You're a builder. You're observing agents building, and you're also coaching humans observing agents building. The level of nuance is pretty special."
Conclusion
For Target, agent success is not about a single model but the confluence of architecture, taxonomy, autonomy levels, security and observability. Models matter, but the competitive advantage arises from the governance and systems built around them.
