Enterprise implementation of AI is turning out to be a rocky road. In May 2026, Starbucks abandoned its high-profile AI-based inventory system, after nine months of failing to get it to work. The “Automated Counting” system was deployed at 11,000 Starbuck stores, rolling out from September 2025. Staff were provided with handheld tablets to scan shelves and refrigerators in the stores, and the software was supposed to identify inventory items, count them and flag any shortages of stock. The system utilised computer vision, spatial mapping and a shelf scanning technology based on LiDAR (Light Detection and Ranging), the technology used in autonomous vehicles to detect obstacles.
The system regularly confused similar looking objects, sometimes missed items entirely, and gave inaccurate counts. Starbucks should have had a clue about potential problems when their own promotional video for the technology clearly failed to recognise a bottle of peppermint syrup when scanning nearby items, but this was ignored. It turned out that the problem was more challenging than expected, since some stock-keeping units look similar, some objects may be partly obscured when on a shelf, and it transpired that reflective packaging confused the system. Starbucks reverted to manual inventory counting after staff gave up using it and started reverting to manual inventory themselves.
This project is an example of where cutting-edge technology is deployed because it is fashionable rather than because it is well suited. The long-established RFID (radio frequency identification) tag technology does not require line of sight from readers, and works up to hundreds of metres. This is what is used in hotel key cards and, you guessed it, inventory systems, in many companies. RFID works just fine for stock control, so why did Starbucks go for an unproven AI technology? While only the executives involved can say for sure, it seems possible that this is an example of “fear of missing out” when applied to AI. There is such momentum of publicity around AI that business leaders want to be seen to be embracing it, and in some cases seem to be leaving their critical faculties unused when making decisions. The costs of this fiasco are unknown, but an 11,000-store rollout will not have been cheap.
Starbucks are not alone in being let down by AI. Pizza Hut deployed a new delivery system across several US states, and are now being sued by a franchisee (Chaac Pizza Northwest, who operate a hundred stores) for a $100 million loss. The lawsuit claims that deliveries used to arrive with 30 minutes at least 90% of the time, but that the new AI system “led to slower delivery times, colder product (caused by delays), and reduced customer satisfaction.” Pizza delivery times went from 30 minutes to 45 minutes, customer satisfaction fell and sales dropped. The new system was a logistics optimisation engine (“Dragontail”) that used historical data and order information to decide when to release orders to the kitchen, when to assign drivers and when to group orders together. The system was reportedly supposed to be a real-time closed loop system, using actual store throughput data, driver timestamps and live preparation times.
Other earlier well documented AI project failures include the Air Canada chatbot that provided incorrect refund policy advice and ended up in a court case. Years earlier, IBM’s Watson AI Health service had been abandoned, with costs estimated at $4 billion. In particular, large language models are probabilistic in nature, producing output one token at a time based on their weightings and their training data. Such “token slot machines” can be very useful, but applying them to business situations that require rules based, deterministic outcomes is inherently risky. Any situation where consistency, accuracy and auditability are important is unlikely to be a good match for large language models.
Of course, technology projects failing is nothing new and has a long history well before AI became such a recent bandwagon. However in aggregate it appears that AI projects have an unusually high failure rate, with 95% failing to deliver value reported by both MIT and Boston Consulting Group in a 2025 report. Historically, IT projects have often struggled. One academic study found that 19% of IT projects in 2017 failed outright, while a further 50% were “challenged”. The study compared failure rates over a ten-year period and found them essentially unchanged. However, the rush to implement AI in the enterprise has so far led to an even higher rate of failure.
There is much speculation as to the reasons for this. Causes may include a lack of measurable success criteria, poor data quality, lack of integration into existing production systems, poor monitoring of models in production and the basic overestimating of AI capability. This overestimation may be due to the sheer level of AI news in recent years, with executives feeling pressure to implement what is perceived to be a leading-edge technology for competitive reasons. There is a well-established gap between executives’ belief in AI compared to the employees on the ground. An Axios study in 2025 found that 73% of executives believed that AI adoption was “strategic and controlled” yet only 49% of their employees agreed. Other studies on this subject with similar results include studies by Enterprise Risk Management Initiative, and a study from Sopra Steria reporting a “stark gap” between senior leaders and frontline managers on AI maturity and skills.
AI investment continues apace, but we are now at the stage where the rubber hits the road for AI projects in enterprises. Three and a half years after the ChatGPT launch in November 2022 we should soon expect to see more and more documented case studies of enterprise AI projects, both successes and failures. It is to be hoped that the huge investment in AI that has occurred over the last three years starts to yield better results than those of Starbucks. This will only happen if project outcomes are rigorously documented and studied, and lessons learned and applied going forward. As philosopher George Santayana noted as far back as 1905, those who do not learn from history are doomed to repeat it.
