Why do so many AI projects fail and what does effective scaled use look like?

There is so much talk today about AI and how it can revolutionise the supply chain, but a 2024 International Data Corporation (IDC) study found that out of 33 observed AI proof-of-concepts (POCs), only four progressed to production, meaning 88% of POCs failed to reach wide-scale deployment. Additionally, Gartner reported earlier this year that, by the end of 2025, at least 50% of generative AI (GenAI) projects had been abandoned after the proof-of-concept stage, citing poor data quality, inadequate risk controls, escalating costs or unclear business value.

At ALSC Global 2026, Nitin Sethi, vice president and head of IT at Autoliv Americas, emphasised how statistics like these demonstrate a common issue in AI deployment, especially within the supply chain: a disconnect between AI solutions and the problems they should be solving. 

Read more: A unified IT and supply chain team to drive digitalisation at Autoliv 

Autoliv Americas' vice-president and head of IT, Nitin Sethi argues that IT’s value proposition has changed – and that companies failing to unify IT, business and supply chain management teams risk leaving competitive advantage on the table.

"We are sitting at the crossroads of a major transformation and I am very confident that AI is going to be pervasive in every form or function of what we are going to do in the next few years," said Sethi. "But as with any other form of technology, we have to make sure that this technology assists us and our organisations to move forward and solve problems."

According to Sethi, AI applications "only make sense when you have a problem statement and you are trying to solve the problem"; it can't be deployed just because it is the latest technology available. "Start with the decision, not the technology,” he advised.

This sentiment was shared by Jared Belt, vice president of commercialisation at ArcBest | Vaux, who further emphasised that excitement for AI must be tempered and the technology should only be deployed where it can deliver real value. "The first step to any successful automation journey is understanding your baselines," he said.

A pilot project without the correct foundation will never be successful, Belt reiterated. He argued that even before ensuring the correct technological foundation for an AI application, firms must first have a comprehensive understanding of their operations, otherwise the outcome of the initiative will be "fictional" and provide little to no real value in the real world.

Data and operational context

Once a relevant operational problem has been identified – be it reducing inventory, improving quality, speeding up decision-making or any number of challenges that a company may wish to overcome – and a real operational baseline has been established, the next step is ensuring data quality. Without the right data, any outcome from the project will again be a "hallucination" rather than actionable insights.

"You can't just use your existing infrastructure, put AI on top of it, and expect that it's going to work," said Julia Fuchs, general manager of material and transport control, delivery Assurance at BMW Manufacturing in Spartanburg, South Carolina. "That's not realistic."

Adrian Jennings, chief product officer at Cognosos, highlighted two different ways that inadequate data quality can inhibit the success of an AI pilot.

First of all, he discussed what he referred to as "data debris" – gaps, repetitions and inconsistencies in data. He noted that in the process of reading and interpreting data, there is more human error-proofing involved than one would perhaps expect. And while an experienced member of staff can easily discard what is nonsensical and focus on what's important, AI cannot make those same judgements. Often, it can treat every data point as equally meaningful, leading to anomalies being over-indexed.

That leads onto the second problem. Jennings noted that even if data is interpreted perfectly, a lack of "domain knowledge" – the day-to-day understanding of how an operation actually runs on the ground – is essential to ensure AI systems are given the data alongside the appropriate context. Without that context, AI can produce confident but incorrect answers because it lacks the operational rules and domain knowledge that experienced workers would naturally apply.

"I don't want AI to invent an answer where there isn't one," said Jennings, emphasising the need to contextualise data. "I want it to tell me if a question is unanswerable at this point, and explain why."

Governance and ownership

With the data and its operational context established, the next challenge is determining ownership of the AI application and the outcome it is intended to deliver. On stage at ALSC Global, panellists repeatedly returned to governance as a prerequisite for scaling AI, particularly as applications begin to emerge organically across different parts of an organisation.

Jennings shared his belief that the responsibility for training AI and ensuring that "tribal knowledge" does not leave a business when an individual does should be shared between a company's operations and IT departments. This would involve IT documenting operations teams' processes to build up "guardrails" and reinforce the right signals.

Watch: Toyota's Jason Stone on leveraging AI and data visibility to make proactive decisions

On the Red Sofa at ALSC Global 2026, Jason Stone, general manager for supply chain development and optimisation at Toyota North America, discussed how AI, data and control towers can improve visibility and shift decision-making from reactive to predictive.

However, Sethi challenged this division of responsibility. He argued that "businesses knows their data best", and that organisations should give operations teams the tools to identify, label and validate their own data rather than expecting IT to determine how it should be interpreted. At Autoliv, he shared how "digital citizens" and "digital champions" have been encouraged, with IT providing a self-service platform while business and engineering teams help validate the information being used.

During a panel discussion on digital transformation in logistics, Thomas Cook, senior manager of inbound domestic transportation and packaging engineering in the US at Nissan North America, made a similar case for operational ownership at Nissan, arguing that the supply chain organisation should ideally own its digitalisation strategy because it is closest to the problems the technology is supposed to solve.

"If we're not doing it, then I don't know that we can be clear on if we're solving the right problem or applying the technology in the right way," Cook said.

The distinction is important as AI moves beyond individual pilots. IT still has a central role in security, governance and providing the technical infrastructure, but experts like Sethi and Cook suggested that effective ownership cannot sit entirely within the technology function. From their perspective, the people who best understand the operation need to have an active role in defining how AI is used, what its outputs mean and where its limits lie.

"You can't have AI applications running some of these critical operations unattended," Sethi said. "There has to be a human in the loop, and there has to be a clear accountability and ownership for training the data."

Designing a pilot to scale

One of the fundamental problems preventing AI pilots from being scaled up is that often they are not designed to scale. As Jennings identified, applications are often built around the specific data, processes and circumstances of a single location or use case, meaning they can work effectively in isolation but struggle when applied elsewhere.

Jennings illustrated the problem as teaching an AI system to respond to a specific scenario without giving it the wider rules needed to understand similar situations. The system may learn that a particular input requires a particular response, but without the underlying operational context it cannot necessarily apply that learning when the circumstances change. Designing for scale therefore requires organisations to identify the underlying principles behind a successful application rather than simply replicating the original pilot.

Belt similarly warned against relying on individual expertise to make pilots work. While a project can be made successful through the efforts of the people directly involved, that approach can be difficult to reproduce across an organisation. "To earn the right to scale, you have to adopt the systems and you have to put the right processes in place," he said.

Scaling AI is not simply a matter of deploying the same technology to more sites. Organisations need to establish the processes, data foundations, governance and human involvement required to make an application repeatable, while ensuring that the underlying business problem and operational value remain clear.

That also means considering the wider network rather than optimising individual points within it. In the digital transformation panel discussion at ALSC Global, Guru Rao, founder and CEO at nuVizz, noted that an automotive logistics operation can perform well in isolation while a failure elsewhere – such as in linehaul, cross-docking or final delivery – still affects the end customer. "The network is only as efficient as the least efficient entity in the network," he said.

For Sethi, that value ultimately needs to be measurable. AI can help speed up decision-making processes, avoid additional costs associated with premium freight and excess inventory, automate repetitive tasks and improve consistency, but there must be a clear connection between the application and the operational or financial outcome it is intended to improve.

"AI is not the KPI; AI enables improvement in KPIs," said Belt, capturing the panel's broader argument that successful AI deployment should be measured by the improvement it delivers to the wider operation, rather than by the adoption of the technology itself.