Every tech executive wants to talk about their frontier models. They brag about parameter counts, context windows, and benchmark scores like sports stats. But if you ask Siobhán Mc Feeney, Senior Vice President at Target, raw models aren't where enterprise defensibility lives. Speaking at VB Transform 2026, Mc Feeney laid out a clear reality: the models running inside retail giant Target aren't what gives the company its edge. Everything built around them is.
The real moat is the operational scaffolding—the taxonomy, data governance, registration protocols, and guardrails that let thousands of autonomous software components run without burning down the enterprise.
What Is an Enterprise AI Platform Beyond the Model?
To understand Target's strategy, you first have to ask a fundamental question: what is an enterprise ai platform?
Many leadership teams confuse an enterprise AI platform with an API gateway connected to OpenAI or Anthropic. That's a mistake. An enterprise AI platform is the unified operational and governance foundation that bridges raw machine learning models with real-world enterprise infrastructure. It wraps intelligence in strict data lineages, security boundaries, registration processes, and evaluation harnesses. Without this substrate, an AI model is just a statistical prediction engine firing into a vacuum.
At Target, that platform takes the form of agentic systems that sit underneath supply chains, inventory management, and demand forecasting. As Mc Feeney pointed out, retail still boils down to its classic promise: getting the right product into the right place at the right time. Doing that at scale with autonomous software requires an architectural layer capable of deciding whether a problem even warrants an AI agent in the first place.
Before writing a single line of code, Target forces development teams through a series of uncomfortable questions. Does this business problem actually require an autonomous agent? If it does, what kind? Are you building a master orchestrator, a domain-specific agent, or a super agent? Or is the tool you're calling an "agent" really just a script with fancy marketing?
Defining those boundaries upfront prevents waste. Target mandates that every agent must be formally registered and certified before deployment. If another team already built a solution for inventory rebalancing, you don't get to build a duplicate simply because you wanted to play with a new LLM framework.
How Agentic Data Platforms Enforce Governance and Lineage
When autonomous software starts making decisions across complex business systems, failure is inevitable. The difference between a minor incident and a disaster comes down to auditability.
Mc Feeney highlighted Target's insistence on full lineage tracking, starting from an agent's initial creation through to its live execution in production environments. "We're trying to make sure we have lineage from the very beginning — the birthing of this agent, all the way through — because at 2 a.m. one morning, when something goes sideways, we want to make sure we understand everything that happened," she explained.
If an agent triggers an automated purchase order at midnight, engineers need to know the exact catalyst. Was it triggered by a system event, an automated timer, or a manual user command? What specific database tables did it read? Which internal tools did it invoke?
This level of operational transparency requires robust agentic data platforms that maintain metadata integrity across every step. Target doesn't stop at measuring latency or uptime. They continuously evaluate agents against intentionality, calibration, and long-term trajectory. If an agent's predictions start to drift from its original intent, the governance layer flags it immediately. Security guidelines and lineage tracking mean that when failure occurs, recovery is fast and predictable.
The Four-Level Ladder: Earning Autonomy in Production
One of the most striking parts of Target's model is how autonomy is granted. Human employees don't get unlimited authority on their first day; they earn trust by demonstrating consistency over time. Target applies the exact same logic to software.
Target structures agent autonomy across a four-level ladder:
- Level 1: Passive Observation. The agent monitors system data and generates internal insights without taking any external action or making user recommendations.
- Level 2: Guided Suggestion. The agent formulates action plans and presents recommendations to human operators, requiring explicit human approval before execution.
- Level 3: Guardrailed Action. The agent executes tasks autonomously, but strictly within predefined operational parameters and strict system guardrails.
- Level 4: End-to-End Execution with Oversight. The agent runs complex workflows independently, while maintaining continuous human oversight and real-time monitoring.
"The autonomy levels for the agents are super important," Mc Feeney said. "They earn them, and they can lose them if they don't perform as expected."
If an agent exhibits model drift or violates guardrails, its autonomy is downgraded immediately. It gets pulled from production or pushed back down to Level 2 where human approval is required for every decision. This dynamic revocability keeps teams moving fast. Developers can push ambitious updates because they know the platform's security envelope will catch out-of-bounds behavior before it impacts customers.
Proven in the Aisle: Digital Twins and Model Cost Gradients
Theoretical governance sounds great in executive meetings, but how does this play out in live retail operations? Mc Feeney shared a concrete example involving inventory forecasting for men's shorts across three Target locations in Long Beach, California.
A digital twin simulation recommended sending six to seven times more inventory to one specific store compared to the other two. Human inventory analysts reflexively rejected the output, assuming the model had hallucinated. But when they looked closer at the spatial data, the math held up. That single store was located less than two miles from the oceanfront. The other two stores were 10 to 12 miles inland. The system had picked up on localized micro-geography and seasonal traffic patterns that human planners overlooked.
The human analysts let the recommendation stand. The inventory arrived, and the stock sold through completely.
"This is science. This is mathematically more significant and more confidence-filling than humans doing it," Mc Feeney noted. Outperforming human intuition in rigorous trials is precisely how agents unlock higher rungs on Target's autonomy ladder.
Target also balances intelligence against economic reality by using different model "gradients." Not every task requires a high-cost frontier LLM. Heavy merchandising supply chain calculations demand massive reasoning power, justifying top-tier models. Simple intent classification or basic workflow routing, however, can run on smaller, specialized models at a fraction of the compute cost. Enterprise platforms must balance model capabilities against unit economics so that scaling AI doesn't sink corporate margins.
The Human Side of Engineering Autonomous Workflows
Building this infrastructure changes the nature of software engineering itself. When software begins observing, deciding, and acting independently, the role of human developers shifts from writing procedural code to managing and architecting resilient agentic systems.
As Mc Feeney pointed out, teams are operating at speeds that were unimaginable just a few years ago. This requires establishing strict evaluation harnesses while cultivating a new culture of oversight. Engineers now find festivities acting as system architects, agent supervisors, and mentors to junior builders all at once.
"You're a builder. You're observing agents building, and you're also coaching humans observing agents building," Mc Feeney remarked. "The level of nuance is pretty special."
The lesson for enterprise tech leaders is unambiguous. Stop obsessing over which foundational model provider won this week's benchmark test. Models will continue to commoditize. The real enterprise moat lies in building the governance, registration, lineage, and observability infrastructure that turns chaotic model outputs into reliable enterprise action.
According to reporting by VentureBeat, Target's approach demonstrates that long-term competitive advantage belongs to organizations that invest in the architecture surrounding the model, not just the model itself.
