Supply Chain Analytics Pulse

Who Approved That Order? Decision Rights for AI in Supply Chain Planning 

October 2026 | Supply Chain Analytics Pulse

In early 2021, Zillow told the pricing experts who checked its home-buying algorithm to stop questioning it. Within months, the company was buying homes far faster than it could sell them. In the third quarter of 2021 alone, Zillow Offers bought 9,680 homes and sold 3,032. By the end of the year it had written down $407.9 million in inventory, shut the business down, and cut about a quarter of its workforce. 

The algorithm got the blame. The researchers who later studied the case pointed somewhere else: the company had removed the human guardrails around a model it was still learning to trust. Zillow Offers was, at its core, an inventory business, and the checkpoint between an AI estimate and a purchase had been taken out. 

Supply chains are now handing similar decisions to AI. Gartner expects spending on supply chain management software with agentic AI capabilities to grow from under $2 billion in 2025 to $53 billion by 2030. Two thirds of supply chain digital investment already goes to AI, yet 55% of chief supply chain officers say they are unclear on the return. And ISG’s latest research finds that more than half of enterprises still keep a human in the loop, rising to more than two thirds for consequential decisions. 

That brings us back to one of the themes from Pulse’s September live session with AI expert Michael Alf: Accountability when AI gets a decision wrong. 

This edition takes that theme into daily operations. It maps where AI decision authority sits in supply chain planning today, where accountability breaks first in data, inventory, and planner overrides, what bounded autonomy looks like at companies that set it up well, and how to assign decision rights that let AI move fast without leaving the outcome unowned. 

01 From Recommending to Deciding: Where AI Authority Sits in 2026

02 Where Accountability Breaks First: Data, Inventory & the Override

When an AI-driven decision goes wrong, the failure rarely starts inside the model. It tends to start in one of three places: the data the model consumes, the moment a recommendation turns into a commitment of stock or cash, and the override, where a human steps in or quietly steps away. 

The decision nobody knows they are making

Target Canada was not an AI failure. It ran on automated forecasting and replenishment, which is exactly why it is worth rereading now. 

Rushing toward its 2013 launch, the company had merchandisers enter data for roughly 75,000 products into SAP on a rigid schedule. An internal investigation later found dimensions entered in inches instead of centimetres, dimensions in the wrong order, the wrong currency, and missing fields. The team estimated the data was accurate about 30% of the time, compared with 98% to 99% in Target’s U.S. systems. At that point the system had no way to flag entry errors. 

The replenishment logic depended on exact product and shelf dimensions. With bad data, it could not work out when a shelf needed restocking. Shoppers found empty shelves while distribution centres overflowed, and the company switched auto-replenishment off at its first stores so staff could restock by hand. 

Target Canada filed for creditor protection in January 2015 and closed all 133 stores by April. 

Nobody at Target Canada decided that the replenishment system should run on data that was right a third of the time. Nobody was assigned to stop it either. The fix, when it finally arrived in 2014, was a verification step that blocked bad data before it entered the system. That is a decision right in its own form: who may write data that an automated decision will consume, and what gets checked before it does. 

When a recommendation becomes a commitment

Zillow’s original process had three checkpoints. The algorithm estimated a home’s value, pricing experts vetted and adjusted the bid, and renovation teams subtracted expected repair costs. 

Project Ketchup, launched in early 2021 to speed up purchasing, collapsed that chain. Zillow began using its Zestimate directly as the cash offer for qualifying homes, stopped pricing experts from modifying the algorithm’s estimates, and sometimes raised bids by thousands of dollars above the algorithmic price to win more sellers. Purchaces jumped from 3,805 homes in the second quarter of 2021 to 9,680 in the third, more than the previous five quarters combined. 

The researchers who studied the case argue that Zillow grew volume without treating its human experts as guardrails for a price model that was still maturing. Their closing question could headline this whole edition: “When AI-based business models fail, who is accountable?” 

The same pattern exists in supply chain every day, in purchase orders, stock transfers, and allocation decisions. The risk grows with three things: how much value one decision commits, how hard it is to reverse, and how many of them run before anyone looks. 

The human in the loop is also a risk point

Keeping a person in the loop is not automatically safe. A study of more than 60,000 forecasts from four supply chain companies found that planner adjustments improved accuracy on average in three of them. Small adjustments, though, often made it worse, and upward adjustments were more often wrong, a sign of systematic optimism.

Target Canada had an override problem too. Analysts judged on in-stock rates switched off automatic replenishment alerts for some products, so the system stopped reporting them as out of stock. On paper, the numbers looked good. On the shelves, they did not.

A human in the loop only works when the loop is visible. Every override needs an owner, a reason, and a record, or the AI ends up blamed for decisions it never made.

03 Bounded Autonomy in Practice: Who Is Getting It Right

04 Organizational Readiness Check

Before approving the next AI feature or agent, it is worth an honest look at whether the organization knows who is allowed to decide what. The first dimension matters most. Without a decision inventory, every other control is guesswork. 

05 Key Implementation Steps: Choose Your Perspective 

Click on the perspective you want to analyze – executives or analysts.

06 Conclusion

Every case in this edition has the same shape. Zillow’s losses grew after it removed the checkpoint between its algorithm and its purchases. Target Canada’s replenishment system followed its data, and nobody owned the data. Each time, the larger problem was a decision nobody had clearly been given the right, or the duty, to make.

What held up was bounded authority, decided in advance. Walmart let AI commit it to supplier terms because people had already defined every term it could offer. With ISG expecting only a small share of agentic processes to run fully autonomously through 2027, most AI decisions will be shared between people and systems for years. That shared zone is where accountability goes missing unless someone designs it.

When the AI gets a decision wrong, the person on the hook should be whoever owned that decision class before the AI touched it, backed by a record of what the AI did, which data it used, and who reviewed it. Organizations that can produce that record within a day are ready for more autonomy. The rest will find out the expensive way.