The Dilemma of AI Implementation in the Retail E-commerce Industry

To explore the reasons behind the failure of AI and machine learning projects, the research team of RAND Corporation in the United States interviewed 65 data scientists and engineers with at least five years of experience in building AI models in industry or academia. These respondents came from enterprises of different scales and across various industries to ensure the conclusions are broadly representative.

Eventually, the interview content and relevant data formed a research report titled "Root Causes of Artificial Intelligence Project Failures and Pathways to Success". It is estimated that over 80% of AI projects end up in failure, either failing to move beyond the pilot phase, or being deployed online but failing to deliver measurable business value.

Similarly, a joint study by MIT and McKinsey shows that 90% of generative AI experimental projects never make it out of the pilot phase. Even more concerning, a 2025 S&P Global survey indicates that 42% of enterprises admitted abandoning most of their AI initiatives, compared to only 17% in 2024. In just one year, the abandonment rate of AI projects surged by 147%.

According to another survey, 84% of business leaders believe AI will have a significant impact on their operations, and 97% of them agree that the urgency of deploying AI technology is on the rise.

However, the same survey shows that only 14% of organizations consider themselves fully prepared to integrate AI into their business operations. Two years ago, only 48% of AI prototype projects could eventually move into production, and the average time from project initiation to production launch was 8 months.

It is obvious that enterprise managers and supervisors are under enormous pressure to "do something with AI" to prove to their superiors that they are keeping up with technological development, yet too many managers do not have a clear idea of how to turn their aspirations into action.

"Zhuang Shuai Retail E-Commerce Channel" hopes to extract five environmental root causes of AI project failure through this report, so as to provide reference for retail e-commerce enterprises to implement AI, avoid common pitfalls and improve project success rates.

Five Root Causes of Failure

Cause 1. Misunderstanding or Miscommunication of Problems by the Leadership

This is the most common cause of AI project failure. Industry stakeholders often misunderstand or miscommunicate what problems need to be solved with AI, and 84% of respondents cited leadership failure as the primary cause.

Specifically, after the trained AI model is deployed, it is found that the metrics it optimizes are incorrect, or the model cannot be integrated into the overall business process and context at all.

Some respondents stated directly that enterprise leaders believe they have a large amount of data because they receive sales reports every week, but these data may not meet new targets.

This root cause of failure is particularly prominent in the retail industry. Many retailers rush to deploy AI customer service or recommendation systems, but fail to clearly define success criteria, such as whether to increase conversion rate, raise average order value, or improve user retention. When the goals are vague, the metrics optimized by the AI model are often disconnected from real business needs.

Two years ago, Target launched an AI chatbot for internal efficiency improvement, but employees generally complained that the tool was incomplete and basically useless. For example, when asked how to deal with impolite customers, the robot only gave suggestions to stay calm and communicate politely.

Cause 2. Lack of High-Quality Training Data

Many AI projects fail because organizations lack the necessary data to fully train effective AI models. After failures driven by leadership, data-driven failures are the second most common cause.

The difficulty lies not only in the volume of data, but also in the cost of acquiring, cleaning and exploring organizational data. Respondents also mentioned the issue of talent availability: 7 out of 50 respondents said talent shortage is a major difficulty, and another 19 said that while there is no overall shortage of talent, high-quality professional talents are lacking.

The "Zhuang Shuai Retail E-Commerce Channel" research finds that in the retail industry, enterprises have massive transaction data, but owning data does not equal to owning available high-quality data. Various types of data are scattered in multiple silos such as POS systems, e-commerce platforms, membership systems and supply chain management systems, with inconsistent formats and disparate standards.

A retail industry data analyst may spend 80% of their time on data cleaning and integration, leaving only 20% for model training and optimization.

Take inventory management as an example, which is one of the most promising application scenarios for retail AI. In reality, problems such as inconsistent product coding, lagging inventory data updates, and asynchronous online and offline inventory will be infinitely amplified in AI systems that follow the "garbage in, garbage out" rule.

Cause 3. Chasing the Latest Technology Instead of Solving Problems

Some AI projects fail because organizations focus more on using the latest and most cutting-edge technology rather than solving practical problems for intended users. The report clearly points out that chasing the latest AI progress itself is one of the most common failure paths.

Successful projects should focus on the problem to be solved, not the technology used to solve it. Many managers are confused by technology hype and forget to examine the real business problems.

After ChatGPT drew global attention, countless retailers rushed to integrate large language models into their own businesses, and the phenomenon of "finding a nail after getting a hammer" became widespread.

In October 2025, Walmart announced its high-profile cooperation with OpenAI to launch an instant checkout function in ChatGPT, allowing users to complete shopping through conversations. However, only five months later, Walmart terminated this cooperation.

What went wrong?

It is understood that this function has serious accuracy problems and cannot match Walmart's internal shopping tools, resulting in the order conversion rate through the OpenAI channel being far lower than that of Walmart's own channels.

Walmart later embedded its own shopping assistant Sparky into ChatGPT and Google Gemini platforms. Early tests show that the purchase completion rate through Sparky reaches about 70% of the direct purchase rate through Walmart.com. Although it is still lower than its own channels, it is far higher than the performance of OpenAI's instant checkout function.

A Walmart spokesperson said that through this cooperation they learned that customers expect a consistent experience at every touchpoint.

This case perfectly illustrates the failure path of chasing the latest technology instead of solving problems: the technology itself is large language models and conversational shopping, but the solution fails to integrate into Walmart's existing shopping processes and user experience, leading to total failure in the end.

Cause 4. Lack of Deployment and Management Capabilities

Organizations may not have sufficient infrastructure to manage data and deploy completed AI models, which significantly increases the possibility of project failure.

The deployment of AI relies far more on the surrounding infrastructure architecture than the model itself. From data governance to model deployment, and subsequent maintenance and iteration, insufficient infrastructure in any link may lead to the failure of the entire project.

The report suggests that upfront investment in data governance and model deployment infrastructure can greatly reduce the time required to complete AI projects and increase the amount of valid data available for training.

According to the "Zhuang Shuai Retail E-Commerce Channel", a large retailer operating hundreds of stores, multiple e-commerce platforms, a complex supply chain network and a huge membership system, it is already a huge project to connect, standardize and synchronize data from these systems in real time.

Theoretically, AI can automatically adjust product prices based on factors such as real-time inventory, competitor prices, weather and holidays.

In practice, this means that the AI model needs to be connected in real time with the POS system, e-commerce platform, supply chain management system and even the in-store electronic price tag system. Delay or error in any link may lead to pricing errors, resulting in profit loss or customer dissatisfaction.

Starbucks launched an AI-powered inventory counting tool, but it was shut down only 9 months after launch.

According to Starbucks' internal technical audit report, the system has an error recognition rate of up to 23% for products with similar packaging, a missing detection rate of over 35% in strong or low light environments, and a recognition accuracy of less than 60% for cylindrical syrup products.

The root cause of the problem is not that the AI algorithm itself is not advanced enough, but that the store lighting conditions and product packaging diversity in the deployment environment far exceed the adaptability of the AI model, and the data collection environment and edge computing capabilities in the infrastructure fail to provide sufficient support for AI deployment.

Cause 5. Applying AI to Scenarios It Cannot Solve

The report emphasizes that AI is not a magic wand that can make any challenging problem disappear. In some cases, even the most advanced AI model cannot automatically complete a difficult task.

Back in 2021, McDonald's cooperated with IBM to launch an AI voice ordering system, trying to replace human order takers with AI. However, the system frequently made errors in actual operation. Videos showed that the system added hundreds of dollars worth of Chicken McNuggets to customers' orders, and some customers reported that bacon was incorrectly added to their ice cream.

Three years later, McDonald's announced the termination of this AI ordering trial.

The problem is that in the drive-thru ordering scenario, car engine noise, rain and wind noise, and various external noises from adjacent lanes seriously interfere with speech recognition, and AI struggles to accurately understand customers' ordering instructions in noisy environments.

This is not a problem of AI not being good enough, but that this specific problem itself exceeds the capability boundary of current AI.

Another interesting case comes from an experiment by AI company Anthropic, where they let the large language model Claude act as an AI store manager to independently operate a snack vending shop, which ended up in a loss.

The AI store manager made frequent mistakes in product selection, pricing and inventory management, and even sold products at a loss due to excessive generosity.

This vividly illustrates that even the most advanced large language models cannot replace human judgment in complex business decision-making, as the problem itself is beyond the capability scope of AI.

The failure of AI projects is not just a simple superposition of the above five problems, but a systematic malfunction of organizational culture, processes and talent structure.

The 10-20-70 rule proposed by Boston Consulting Group holds that the success of AI projects depends 10% on algorithms, 20% on technology and data infrastructure, and 70% on people, processes and cultural transformation.

However, most organizations completely reverse this proportion, investing 70% of their energy in technology procurement and deployment, and only 10% in focusing on personnel and cultural transformation.

As a typical labor-intensive industry, retail e-commerce enterprises are highly dependent on human experience and judgment in every link, from store employees to supply chain managers, from customer service staff to procurement specialists. When organizations try to use AI to replace or assist these roles, if they ignore personnel training, process reengineering and cultural transformation, AI projects are almost doomed to fail.

In addition, the report also finds that the internet-based agile development methodology may have inherent conflicts with AI projects.

10 out of 50 respondents believe that AI projects require an initial stage of data exploration and experimentation whose duration is unpredictable. As one respondent put it, much work has to be repeatedly reopened in the next task, or split into extremely tiny and meaningless tasks to fit into a one-to-two-week task cycle.

For retail e-commerce enterprises, this means that managing AI projects in the way of managing traditional software projects, such as fixed cycles, fixed scope and fixed deliverables, plants the seeds of failure from the very beginning. AI projects require greater flexibility, longer exploration cycles and tolerance for failed experiments.

How to Apply AI Successfully?

First, ensure that technical personnel understand the project purpose and domain background, and can interact effectively with business experts.

In retail scenarios, data analysts need to understand the real situation of retail business, including the rhythm of store operations, nuances of customer behavior and constraints of the supply chain. Simply taking increasing conversion rate as the goal is far from enough; they also need to deeply understand what kind of recommendation is effective for what kind of customers at what time and in what scenario.

Second, AI projects require time and patience to complete.

Before launching any AI project, leaders should prepare to let each product team commit to solving a specific problem for at least one year.

Retail enterprises are often accustomed to quarterly performance assessment and rapid trial and error, but AI projects, especially those involving core businesses such as supply chain optimization, demand forecasting and personalized recommendation, require sufficient time to accumulate data, iterate models and verify effects.

Third, focus on problems rather than technology.

Before launching any AI project, ask three questions: What specific business problem needs to be solved? How to measure success? If there is no AI, is this problem still worth solving?

If the answers to these three questions are not clear enough, this AI project is likely to be on the wrong path from the very beginning.

Fourth, invest in infrastructure.

Upfront investment in data governance and model deployment infrastructure can greatly reduce the time required to complete AI projects and increase the volume of data available for training effective models.

For retail enterprises, this means that before launching AI projects, they need to solve the problem of data silos first, connect systems such as POS, e-commerce, supply chain and membership, and establish unified data standards and a data governance system. This seemingly unglamorous work is precisely the prerequisite for AI success.

Fifth, understand the limitations of AI.

Despite all the hype around AI, leaders need to clearly recognize the capability boundary of AI and involve technical experts for feasibility assessment when considering potential AI projects.

AI has significant advantages in data processing and pattern recognition, but still has huge limitations in understanding human emotions, handling complex social interactions and responding to unexpected incidents.

Retail enterprises should avoid deploying AI in scenarios that are highly dependent on human judgment and emotional interaction, at least under current technical conditions.

Driven by both huge competitive pressure and technology hype, retailers are rushing to deploy AI, including AI customer service, AI recommendation, AI pricing and AI inventory management. However, too many projects fall into the trap of "doing AI for AI's sake", forgetting that AI is only a tool, and business is the end goal.

For any organization that is considering or has already launched an AI project, before using new technologies, make sure that the problem is correctly defined, the data is ready, the infrastructure is solid enough, and the personnel and culture are in place.

This article is from the WeChat Official Account "Zhuang Shuai Retail E-Commerce Channel", written by the Zhuang Shuai Research Team, and authorized for release by 36Kr.