Artificial intelligence is neither the risk nor the promise in supply chains. By itself, the technology does what evidence shows it can do: improve forecasts, tighten inventory, and lower logistics costs. When AI fails, the cause is rarely the technology alone. It is about what surrounds it: the wrong pilot chosen for the wrong reason; data too scattered or stale for a model to trust; an operating model and workforce not built to act on what AI recommends; and no clear accountability for the decisions it shapes.
The gap between AI’s promises and actual business results lies largely in areas supply chain leaders control directly. This paper examines where AI already delivers in supply chains, why the payoff stalls even when the technology works, where the real risk sits, and what a company needs in place before it scales AI beyond a pilot.
Where AI Already Delivers
AI’s case in supply chains does not rest on projections alone. McKinsey’s review of distribution operations puts the achievable range at inventory reductions of 20% to 30%, logistics cost reductions of 5% to 20%, and procurement spend reductions of 5% to 15%, a range drawn from companies that have already deployed AI in these areas, not a guarantee for the next one. [1] The same research also highlights check on how unprepared many companies are to capture that range: 95% of distributors it surveyed are exploring AI use cases, fewer than 1 in 10 have a road map that prioritizes which to pursue first, and only about 3 in 10 say they have the internal talent to take a pilot past its first deployment. [1] The technology is not the constraint. Readiness is.
Unilever’s more recent work with Walmart Mexico shows what picking the right pilot looks like in practice. In 2022, the two companies built a joint, AI-driven planning model under what they call a “One Supply Chain” approach, linking real-time point-of-sale and inventory data so both organizations could plan from a single source of truth rather than two separate forecasts. [2] They started with a single category, Unilever’s nutrition business, before extending the model across Unilever’s full range of products sold through Walmart Mexico. On-shelf availability rose above 98%, inventory levels came down, and fill rates and forecast accuracy improved, results the two companies reported within less than a year of the pilot. [2] The pilot design mattered as much as the technology itself: a shared, measurable outcome that both companies could see and be held to, rather than an AI initiative run inside Unilever alone.
DHL ran a comparable test at its largest distribution center in Latin America, in Louveira, Brazil, though the technology there is simulation, not AI. Warehouse managers had been planning each shift manually in a spreadsheet, estimating how many pickers a day would need. DHL replaced that guesswork with a simulation-powered digital twin, a live model of the warehouse, and used it to plan shifts with 98% accuracy. [3] The result was fewer resourcing gaps and more reliable on-time delivery. The example belongs here because the lesson extends beyond AI: any model a company puts between a decision and the person who used to make it, simulation or machine learning, only pays off if the data behind it is clean and the people running it are trained to act on what it says.
These are not isolated wins. Forecast accuracy, inventory position, and logistics cost show up again and again as the levers that move when AI, or a comparable model, is pointed at a well-defined problem with good data behind it. That is the part of the story that should ground everything else in this paper.
Where the Payoff Stalls
The question is no longer whether AI works. It is that most companies have not built the conditions—in data, people, and process—for it to pay off at scale.
A June 2025 survey of 120 supply chain leaders who had already deployed AI found that only 23% had a formal AI strategy. [4] The rest were running AI project by project, chasing short-term wins with no plan behind them, and, per the McKinsey data above, often without the internal talent or the road map to know which project to run next. [1] In other words, roughly three out of four companies are inesting in AI without a clear vision of what they are building or an operating model capable of scaling success beyond pilot.
The timeline makes this worse. A 2025 survey of more than 1,800 executives across Europe and the Middle East found that AI typically pays back in two to four years, three to four times longer than the seven to twelve months a conventional technology investment takes. Only 6% of companies saw a return in under a year. [5] Even so, the same survey found that 85% of companies increased their AI spending over the past year, and 91% plan to increase it again. Most supply chain leaders are ready to defend next quarter’s AI budget. Few are prepared to justify investments that may take two to four years to generate measurable returns, particularly when success metrics were never defined upfront.
Where the Real Risk Sits
The risk is not job loss. It is decision-making without effective oversight. The risk magnifies when no one is positioned, and empowered, to catch the model when it is wrong.
Zillow’s home-buying business is the clearest public example. In 2021, Zillow shut the business down after unprecedented swings in home prices exposed a forecasting model that had been built for calmer markets, leading to more than $500 million in inventory write-downs and cuts to roughly a quarter of its workforce. [6] The cause was not simply an algorithm left to run unchecked. Reporting on the shutdown shows Zillow’s own staff bid above the model’s price estimates to compete for homes, and pricing teams were discouraged from questioning its numbers, meaning the humans in the loop pushed in the same direction as the model’s errors instead of catching them. [6] The lesson is not that human oversight was absent. A human in the loop is not a safeguard unless that person is empowered, to challenge the model , and accountable for doing so.
A version of this risk sits inside most AI-run supply chains today. As AI takes on more forecasting, ordering, and exception-handling decisions, the reasoning behind those decisions often is not recorded anywhere. In many cases, no one can explain what data, assumptions, or rules produced a specific recommendation. That leaves no record to check when a decision is later questioned by an auditor, a regulator, or a customer.
A separate risk sits at the edge of the company, with suppliers and vendors. The World Economic Forum’s Global Cybersecurity Outlook 2026 found that roughly one-third of organizations still have no process to validate an AI tool’s security before it is deployed, 40% review a deployed AI tool on a regular schedule, and the remainder do a single check at launch and nothing after. [7] The total share doing any kind of assessment did nearly double year over year, from 37% to 64%, so the trend is toward more scrutiny. But close to a third of companies are still starting from zero on AI tools already running in production. [7]
The same report found that third-party and supply chain weaknesses are now the single biggest resilience concern among large companies specifically, cited by 65%, up from 54% a year earlier. [7]
Confidence is also moving the wrong way. Akkodis’ 2025 survey of company leaders found that confidence in their own AI strategy fell 11 points in a single year, from 69% to 58%. Among CEOs, the share who called themselves very confident in their AI strategy dropped from 82% to 49%. [8] Spending is going up. Confidence in the plan behind that spending is going down. That combination is a warning sign about readiness, not about the technology.
Why the Same Technology Gets Different Results
The answer is not more advanced technology. It is what a company puts in place before, during, and after implementation: selecting high-value use case, ensuring data quality, and building an operating model that can act on AI-generated recommends, and establishing clear accountablility for outcomes. The difference between success and failure is rarely the model itself. It is organizational readiness. Companies that scale AI successfully treat governance, workforce capability, data quality, and measurable business outcomes as part of the implementation, not as afterthoughts.
Recommendations
DSCI recommends a four-step pathway before any company scales an AI-driven decision system beyond a pilot.
1. Assess readiness before committing to a pilot.
Take stock of whether the necessary data, talent, governance, and decision processes are in place before selecting a use case. A company that skips this step is choosing its pilot blind.
2. Select the highest-value pilot, not the easiest one.
Prioritize use cases based on business value and data readiness, not convenience or organizational politics. Start in one category or one site, the way Unilever and Walmart Mexico began with a single product category, and prove the model before extending it.
3. Define clear KPIs and human accountability before scaling.
Define success metrics before launch and assign a single accountable executive for high-stakes AI-driven decisions. Otherwise, the company cannot prove the pilot worked, and it has no way to defend a decision later when an auditor, a regulator, or a customer asks why it was made.
4. Monitor, learn, and scale deliberately.
Evaluate the model’s performance at least twice a year, and treat every pilot’s results as a decision point: scale it, adjust it, or stop it, rather than letting it run indefinitely on the strength of its first good quarter.
NET NET
AI is already creating measurable value in supply chains. The companies pulling ahead are not necessarily those with the most sophisticated models, but those with the strongest foundation for deploying them. Success depends on readiness: clean data, capable people, clear accountability, and measurable outcomes. Governance is part of that equation, but readiness is what determines whether a pilot becomes an enterprise capability or an expensive experiment.
DSCI’s AI Supply Chain Readiness Index and pilot program were designed around this principle. They help organizations assess where they stand, identify gaps, and move from isolated pilots to AI-enabled decisions that leaders, boards, partners, and customers can trust.
The Digital Supply Chain Institute (DSCI) helps supply chain leaders and companies move to and navigate the future through applied research and project-based learning. Contact Dravida Seetharam at [email protected] or Jeffery Caine at [email protected] to learn more about impact-driven collaboration with DSCI.
Footnotes
[1] McKinsey & Company, Harnessing the Power of AI in Distribution Operations (November 15, 2024). Findings drawn from a McKinsey AI in Distributor Operations sentiment survey (September 2024, n=40) and a McKinsey survey of distributor operations (December 2022, n=74). Figures presented are ranges of reported impact from companies that have deployed AI in these areas, not outcomes guaranteed across the full sample.
[2] Consumer Goods Technology, Unilever Expanding Real-Time Supply Chain Model After Boosting Walmart OSA Rates (October 2024). Case study on Unilever and Walmart Mexico’s “One Supply Chain” AI-driven planning model, launched 2022.
[3] Simul8 Corporation and DHL Group, Using a Simulation-Powered Digital Twin to Transform Day-to-Day Decision-Making in Fast Fashion Logistics, case study on DHL’s distribution center in Louveira, São Paulo State, Brazil. The technology described is a simulation-powered digital twin, not artificial intelligence or machine learning.
[4] Gartner, Gartner Survey Shows Just 23% of Supply Chain Organizations Have a Formal AI Strategy (June 11, 2025). Based on a survey of 120 supply chain leaders who had deployed AI within the prior 12 months, conducted December 2024 to January 2025.
[5] Deloitte, AI ROI: The Paradox of Rising Investment and Elusive Returns (2025). Based on a survey of 1,854 executives across Europe and the Middle East, supported by 24 in-depth interviews.
[6] Zillow Group, Q3 2021 earnings announcement and investor call (November 2, 2021), announcing the wind-down of Zillow Offers, and subsequent reporting and analysis, including Stanford Graduate School of Business, on the pricing decisions made as the business scaled, including staff bidding above the algorithm’s estimates to compete for homes.
[7] World Economic Forum, Global Cybersecurity Outlook 2026 (January 2026). Findings on AI security assessment practices and supply chain risk drawn from the report’s global executive and CISO survey data; the year-over-year comparison (37% to 64%) is as reported within the 2026 report itself.
[8] Akkodis (The Adecco Group), What CTOs Think: Using Digital Transformation to Scale Skills and Unlock Enterprise Potential (2025). Based on a survey of 2,000 global executives, including 500 chief technology officers.
