The board approved it last quarter. The use cases are lined up for faster delivery, lower freight cost, automated customs documentation, dock-to-stock optimisation, exception-aware routing. The model partnerships are signed. The pilot timelines are on the slide. Your enterprise is part of a market that MarketsandMarkets projects will hit $40.53 billion by 2030 at 28.2% CAGR. Gartner forecasts that supply chain management software with agentic AI alone will grow from under $2 billion in 2025 to $53 billion by 2030. The investment is real. The momentum is real. The vendor pitches are confident.

The investment you actually made, however, was never in the model.

It was in everything underneath the model the data infrastructure across TMS, WMS, ERP, OMS and carrier systems; the integration wiring that lets AI execute a booking instead of recommending one; the monitoring that keeps a model accurate when port congestion shifts twice in a week; the governance that lets you trace a customs declaration back to the data and decision logic that produced it. In most logistics enterprises, that work wasn’t built. It was assumed.

That assumption is where the GenAI investment either delivers or stalls. And the data on stalling is no longer ambiguous.

The Failure Pattern Is Structural, Not Technological

MIT’s Project NANDA found that 95% of enterprise GenAI pilots delivered no measurable P&L impact in 2025, against an estimated $30–40 billion in funding. Gartner predicts 60% of AI projects will be abandoned through 2026 because of a lack of AI-ready data, and over 40% of agentic AI projects will be cancelled by end of 2027 citing escalating costs, unclear value, and inadequate risk controls. S&P Global’s 2025 enterprise survey shows the share of companies abandoning the majority of AI initiatives before production has surged from 17% to 42% year over year. BCG’s 2025 study of more than 1,250 firms worldwide finds only 5% are achieving AI value at scale.

Read those root causes again: data quality, integration, risk controls, workflow misalignment. None is a model problem. All are foundation problems. RAND’s primary research on AI failure root causes frames it directly: more than 80% of AI projects fail, twice the rate of non-AI IT projects, driven by organisational and data issues, not technical ones.

The pattern holds in logistics specifically. The 2026 MHI Annual Industry Report shows AI usage in supply chain operations rising from 30% to 41% in a single year, with adoption projected to reach 82% within five years and 60% of leaders planning over $1 million in technology investment. The same report names the top barriers workforce talent, forecasting accuracy, and an inability to translate AI capability into a business case. APQC, cited by the World Economic Forum, finds only about a quarter of supply chain leaders believe their digital transformation is complete, and more than 40% of organisations have limited or no visibility into Tier 1 supplier performance. BCI Global’s research is sharper: only 35% of companies have appropriately detailed end-to-end data visibility across their supply chain. McKinsey’s repeated supply chain leader survey finds 90% report a digital talent gap that has not meaningfully improved since 2020.

You cannot run production-grade AI on a supply chain where two-thirds of organisations cannot see their own shipments end to end, and where the people who would build the foundation are not on staff.

The cost of leaving it that way is already on the P&L. Gartner estimates poor data quality costs the average organisation $12.9 million per year. MIT Sloan research puts the revenue impact at 15% to 25% annually. In a logistics business running at single-digit margins, that is not an IT line item. It is a competitive position.

Where the Investment Actually Has to Go

The foundation is not one thing. There are four, and all four are engineering work not slideware.

The data layer is unified, contract-governed shipment, order, inventory and partner data flowing across TMS, WMS, ERP, YMS, OMS, carrier APIs, port and customs systems, and telematics. Without it, the model is reasoning over three contradictory versions of the same shipment, and the freight cost it is trying to optimize is calculated against the wrong baseline. Most logistics enterprises did not build this layer because they grew it through M&A, through carrier onboarding, through emergency Covid-era integrations that were never consolidated. The data is there. It is just not addressable.

The integration and execution layer is the wiring that lets AI generate the booking, file the document, dispatch the load, and raise the exception instead of producing a recommendation a human still has to action. This is the line between advisory AI and operational AI. McKinsey estimates GenAI can reduce documentation lead time by up to 60% in supply chain workflows, but only when the model is connected to the systems that produce the document. Gartner explicitly notes that enterprise deployment of AI-driven SCM “will lag behind general availability” because of the gap between the model and the operating model. The bottleneck is integration, not intelligence.

The monitoring and feedback layer is observability built for a volatile environment. Resilinc’s EventWatchAI logged a 38% year-over-year increase in supply chain disruptions in 2024. McKinsey estimates disruptions lasting a month or longer now occur every 3.7 years and cost the average organisation 45% of a year’s profits over a decade. Model accuracy degrades fast and silently in that environment. IBM’s Global AI Adoption Index found 62% of businesses cite unexpected performance variation and model drift as their top AI management issue. Yet S&P Global reports MLOps tooling adoption at just 27%. The layer that catches drift is being built last, not first.

The governance and traceability layer is auditability built in, not bolted on. From August 2026, the EU AI Act’s high-risk obligations become enforceable, and AI systems used in road traffic and critical supply infrastructure sit explicitly in scope. Vectara’s HHEM benchmark records GenAI hallucination rates between 3% and 27% on summarisation tasks. In customs, hazmat, pharma cold chain and contractual freight, “mostly accurate” is not a position you can defend to a regulator, an insurer, or a customer.

The Companies Winning Are Publishing the Foundation, Not the Model

C.H. Robinson reported over three million shipping tasks executed by generative AI agents in 2025, including more than a million automated price quotes and a million orders and credits GenAI as a key contributor to a 30% productivity uplift across 2023 and 2024. Kuehne+Nagel’s eTouch platform saved 1.27 million hours of manual work in air cargo alone, equivalent to 750 full-time employees and roughly two percentage points of margin. XPO is converting AI-assisted planning and freight flow optimisation into 4% productivity gains against a 1.5% target every point worth approximately $9 million in annual profit. Flexport cut its customs error rate from 1.8% to 0.2% and identified $4 billion in tariff refunds through its AI tooling.

None of these companies won because their models were better. They won because they built the data, integration, monitoring and governance underneath. The contrast is instructive. Maersk and IBM shut down TradeLens, their flagship blockchain-enabled global trade platform, in 2023, citing failure to reach commercial viability across the partner network. In the same window, Maersk doubled down on its reefer data foundation, upgrading more than 80% of its refrigerated container fleet to deliver hourly data into Captain Peter, its visibility tool, with 90% coverage targeted by end of 2023. The platform play that lacked the data network underneath it failed. The foundation that quietly equipped 380,000+ containers with telemetry now powers a real product. McKinsey’s 2025 State of AI survey makes the differentiator explicit AI high performers are nearly three times as likely as their peers to have fundamentally redesigned the underlying workflows.

Foundation Work Is Not Infrastructure Debt It Is the AI Investment

The conversation in most logistics boardrooms still treats AI strategy and data, integration, and platform work as parallel tracks. Approve the AI use case here, sort out the data debt over there. That is how 95% of pilots become PoCs that never go live, and how the vendors of next year’s models will sell the same demo to the same buyer who is still trying to operationalise the last one.

The logistics companies that will define the next eighteen months will stop funding two halves of a project and start funding one. They will not buy a model they will build a foundation engineered for execution: an AI-ready data fabric across operational systems, integration that turns advisory output into operational action, monitoring that keeps models accurate as the supply chain moves, and governance that survives a regulator’s audit. This is what AI execution engineering actually means in a logistics context, not better prompts, but a production-grade base under the prompts.

Before approving the next GenAI use case, the questions are sharp. Can your AI execute a booking end-to-end without a human bridge, or only recommend one? Do you have a single source of truth for shipment state across TMS, WMS and carrier systems, or three versions of the same shipment? If your forecasting accuracy dropped 12% this week, would you know how fast? Can you trace any AI-generated trade document back to its data, model version and decision logic? When the EU AI Act’s high-risk obligations land in August 2026, will the systems that route your loads pass the audit?

If the answers are uncertain, the next conversation is not about which model to buy. It is about who will engineer the foundation underneath it.

The model is not the moat. The foundation is.