Target's agent autonomy ladder: earn it, don't grant it

Target's SVP of technology for core retail services laid out how the company decides how much rope an AI agent gets, and it's the most copyable thing we've read this month: a four-level autonomy ladder where agents start out only watching and have to earn each promotion. The agent autonomy model matters more than the model choice, and unlike a $200M data platform, you can implement the ladder on a whiteboard this week.

What actually happened

Speaking at VB Transform 2026, Siobhán Mc Feeney described Target's real moat as everything built around the models — architecture, taxonomy, a data governance layer, and continuous observability — rather than the models themselves. Her filter on scope was blunt: "Every enterprise wants AI agents, but not everything needs one."

The ladder has four rungs: observe only, suggest actions a human approves, act inside defined guardrails, then run end to end with human oversight. Agents climb by demonstrating performance and get demoted if they drift. Target measures what each agent was supposed to do, how well-calibrated it was, and the path it took to get there.

The example that sells it: a digital-twin simulation forecast men's shorts inventory for three Long Beach stores and called for six to seven times more stock at one of them. Inventory analysts assumed it was broken. The model had picked up something they hadn't weighted — that store sits less than two miles from the beach, while the other two are 10 to 12 miles inland. They let it run. It sold through.

Why it matters for your business

Read that example again, because the lesson isn't "trust the AI." It's that the analysts could let it run at all. There was a defined rung the agent occupied, a measurable outcome attached to it, and a way to walk the decision back if the shorts had rotted in the stockroom. That's what made a weird-looking call a cheap experiment instead of a bet on a black box.

Most small businesses do the opposite. An agent gets read-write access to the CRM, the inventory system, or the payment processor on day one because the integration was easier that way, and nobody defines what "working" means. When it does something strange, there's no measurement to argue with and no rung to demote it to — so it gets ripped out entirely, and you learn nothing.

Set the ladder up front. Rung one: the agent writes recommendations to a Slack channel or a table and touches nothing. Rung two: it drafts, a human clicks send. Rung three: it acts, but only under explicit limits — reorders under $500, refunds under $50, appointments only in open slots. Rung four is rare and should stay rare. Log every action with the input, the decision, and the outcome from day one, because promotion should be a number, not a vibe. And write down the demotion trigger before you need it.

Sources: VentureBeat.