Retail’s boldest AI experiments have a habit of failing where everyone can see them. Drive-thru orders mangled beyond recognition. A cashierless store quietly switched off. A pricing algorithm that lost hundreds of millions. It is tempting to file these under “AI is not ready.”
That is the wrong lesson. In almost every case the technology did roughly what it was built to do. What failed was the readiness around it: the testing, the data, the honesty about how mature it really was, and whether its goal included the customer.
AI did not let these retailers down. Their readiness did. And that is exactly why the stumbles are worth studying, because every retailer is now making the same kinds of decisions.
The stumbles, and what they teach
| Case | What happened | The lesson |
|---|---|---|
| McDonald’s and IBM drive-thru (2024) | Ended a multi-year AI voice-ordering trial at 100-plus US drive-thrus after it misread real, messy orders | AI has to survive contact with real conditions, not just demos |
| Amazon “Just Walk Out” (2024) | Pulled the cashierless technology from its US grocery stores; it still relied on significant human review behind the scenes | Automation that quietly depends on people is not as ready as it looks |
| Zillow home-buying algorithm (2021) | A pricing model overpaid for homes, leading to a write-down of around US$304 million and the unit’s closure | A confident model on shaky data fails fast and at scale |
| Coles smart anti-theft gates (2023 to 2024) | Automated gates trapped and frustrated honest shoppers | AI aimed at the business can backfire when it punishes good customers |
They are readiness failures, not AI failures
Two of these come down to the same thing. AI met the real world, and the real world was messier than the model.
McDonald’s ran its IBM voice-ordering trial for three years before ending it in 2024, after the system kept misreading orders under the noise, accents, and improvisation of an actual drive-thru. Amazon pulled its “Just Walk Out” technology from its US grocery stores the same year. The system worked, but it still relied on significant human review behind the scenes to confirm transactions. Amazon has said those people were training the model rather than watching shoppers live, but either way it is a sign of a capability being sold as finished while it was still learning. Neither was a foolish idea. Both were deployed as complete before they were.
Zillow’s case is the sharpest, and it is about data. Its home-buying arm used a pricing algorithm to decide what to pay for houses. When the model’s assumptions drifted from reality, it overpaid at scale, and the company took a write-down of around US$304 million and shut the unit down. It is a real-estate example, but the lesson is pure retail: a confident model sitting on shaky data does not fail quietly. It fails fast, and expensively, because it makes the same wrong call thousands of times before anyone catches it.
When AI optimises against the customer
The Coles example is a different kind of stumble, and an instructive one for anyone putting AI in-store. Coles introduced automated anti-theft gates, and honest shoppers reported being trapped and frustrated by them, as widely covered in the Australian press through 2023 and 2024.
The technology did what it was told. The problem was what it was told to value. AI pointed purely at a business goal, here loss prevention, can quietly punish the majority of customers who did nothing wrong. It can hit its target and still cost you trust, and that trade almost never appears in the business case.
The common thread: AI amplifies your operations
Across all four, the AI mostly did its job. What let it down was the readiness around it: real-world testing, an honest read on maturity, trustworthy data, and a goal that included the customer. AI does not rise above weak foundations. It amplifies whatever is beneath it, at speed and at scale.
That is the useful part for any retailer weighing its own AI plans. Before the model matters, the operation underneath it matters more. Where retail AI touches stock, orders, and pricing, it is only ever as reliable as that data and those workflows. Krisp Systems helps retailers connect POS, orders, inventory, and fulfilment into one operational view, so the information AI draws on reflects what is actually happening across the business. It will not write your drive-thru script, but it gives the AI that touches your core operations something solid to stand on.
Before your next AI initiative, ask four questions
Each public stumble maps to a question worth answering before you deploy, not after:
- Reality: has it been tested against messy, real-world conditions, not just a controlled demo?
- Maturity: is it genuinely ready, or quietly propped up by manual effort?
- Data: is the information underneath it accurate and connected enough to trust?
- Customer: does its goal account for the customer experience, or only a business metric?
If any answer is uncertain, that is where your AI will stumble first.
The practical takeaway
The retailers who got AI wrong in public were not reckless. They were early, and they moved before the groundwork was ready. The lesson is not to avoid AI. It is to be honest about what is underneath it, and to fix that first.
AI will find every weakness in your operation and broadcast it at scale. Build the foundation, and it will do the same with your strengths.
FAQs
Why do retail AI projects fail publicly?
Usually not because the AI is bad, but because it was deployed before the operations, data, and testing around it were ready. The technology then amplifies those gaps at scale.
What can retailers learn from AI failures like McDonald’s or Zillow?
That AI has to survive real conditions and rest on trustworthy data. McDonald’s voice AI struggled with messy real orders, and Zillow’s pricing model overpaid when its data drifted from reality.
Does connected data prevent AI failures?
It does not fix every kind of failure, but where AI touches stock, orders, and pricing, connected and accurate data is what lets it give reliable answers. AI amplifies the quality of the data beneath it.
What questions should retailers ask before deploying AI?
Whether it has been tested against real conditions, whether it is genuinely mature or propped up by manual effort, whether the underlying data is trustworthy, and whether its goal accounts for the customer, not just the business.
Is AI worth the risk in retail?
Yes, when the groundwork is ready. The lesson from public stumbles is not to avoid AI, but to build the operational and data foundation it depends on before scaling it.
Want AI that stands on a foundation it can trust? Talk to Krisp Systems about connecting POS, inventory, orders, and fulfilment into one operational view.
