Sam's Club Moved 65,000 One-Hour Orders. Now the Hard Part Starts

Sam's Club switched on one-hour delivery across 600 clubs this past year and pushed 65,000 of those orders out the door in a matter of weeks. Uber Eats started sending drivers to collect retail returns off customers' porches. Those numbers will be stale by next quarter, but the shift they signal won't be: fast delivery has gone from nice-to-have to expected, even in categories nobody ever planned to move in under an hour.

The expectation is food apps' doing. They spent a decade training consumers to expect on-demand speed, and that demand has spread to paint, lumber and bulk groceries. Underneath sits software making thousands of small calls an hour — deciding who picks up each order and how it reaches the door.

The fix everyone reaches for is the AI agent. Most of what's sold today isn't ready. As of June 2025, Gartner counts only about 130 legitimate vendors among the thousands now selling agentic AI, and the firm expects more than 40% of agentic AI projects to be canceled by the end of 2027.

What should the software be allowed to decide?

Salman Habib, cofounder of delivery-network company Burq, argues the industry fixates on the wrong question. "Everyone wants to know what the AI can do," he writes. His narrower version: "What has this software earned the right to decide in your operation?"

He flags two vendor traps. The first is what Gartner calls "agent washing" — a chatbot or rules engine relabeled an AI agent and priced like one. The second is the full-autonomy pitch, software that supposedly runs the whole operation. "That one should scare any operations lead who answers for missed deliveries," Habib writes.

Both traps treat autonomy as something you install and flip on. Habib's counter: autonomy is trust, granted by an operator, one type of decision at a time. No vendor can install it for you. He sees buyers catching on — a year ago the first meeting question was "Do you have AI?" Now it's "What does AI decide on its own, and what does it hand to a person?"

How do you grade the machine?

Delivery, Habib argues, is where the machine should prove itself fastest. When an order stalls with a provider while the delivery window closes, someone has to decide when waiting becomes the riskier call and the order moves elsewhere. The verdict arrives within the hour: the customer has the order, or they don't. No three quarters of internal debate required.

His method runs on a record most operations don't keep. Every operation tracks what moved and what it cost. Almost nobody tracks the third one: why each order went the way it did. That third record is what lets you grade the machine.

The playbook:

• Run the machine's calls next to your dispatchers' calls for a month, with reasons on both sides.
• Count how often they agree, each side's on-time rate, and how often a person steps in and is right.
• Where the machine wins, stop reviewing every decision. Where it doesn't, it recommends and a person approves.
• Break a rule you set, and the authority comes back to a person on the spot.

One decision might earn its way through in 30 days, another in a year. If the dispatcher still sees something the machine doesn't, the dispatcher keeps the call. Once the machine has the better record, paying someone to rubber-stamp it is wasted motion. The constant: a decision earns autonomy on evidence from your own operation, not a vendor demo.

What stays with people?

Assume the models keep improving. Assume routing, batching and ETAs all get solved. Two parts of the business still run on people, Habib writes.

The first is physical. A person still has to pick up the order and get it to the door — whether that's a can of paint on a porch or a new phone at a gated community where nobody answers the callbox and the driver burns through the delivery window. "Whatever the doorstep robots turn out to be good at, they aren't carrying catering trays up a stairwell anytime soon," Habib writes. For most retail delivery, the last 50 feet is a person.

The second is accountability. Software optimizes against whatever goals you hand it, so the real work is deciding what those goals should be for a particular retailer serving particular customers. Habib's example: a $3,000 washer and dryer going to a third-floor walk-up needs a two-person crew who'll haul out the old pair, hook up the new ones and run a cycle before leaving. No model knows that until a person tells it. And that judgment work gets heavier, not lighter, as software improves — fewer decisions made by hand means each one covers more ground.

The line is already moving

None of this slows down, Habib writes. There's another retailer next quarter, another promise to keep, and the software keeps earning its way into more of the work whether anyone's watching or not. His closing instruction: ask whoever runs your dispatch floor which decisions the software makes without a person in the loop, and what it did to earn each one. If the answer isn't written down anywhere, the line already moved and nobody decided to move it.

Vendors will keep describing what AI can do, and increasingly they'll be right, he concedes. What none of them can tell you is what it should be deciding in your operation, on your accounts, against the promises you've made. That part was never theirs to sell.