More Efficient Does Not Mean More Sustainable
Artificial intelligence is entering supply chains promising less waste. MIT's new report shows that the climate balance sheet depends on who measures, who responds, and what is actually being measured. Nobody dares to say out loud anymore that sustainability doesn't matter. In the 2026 report from the MIT Sustainable Supply Chain Lab, the seventh edition of an annual survey, 83% of surveyed professionals judge it important or very important for long-term success, compared to 80% last year.
Following recent shifts in American climate policy, 59% declare a greater commitment and only about 13% a lesser one. So far, good news. Then the picture cracks: 68% report that uncertainty surrounding tariffs forced them to delay, scale back, or compromise their strategies; about one-third saw trade disruptions increase operational emissions (switching suppliers, lengthening shipping routes); and 84% believe armed conflicts reduce the effectiveness of climate action. Sreedevi Rajagopalan, director of the lab and lead author, summarizes this phase in three moves: solid commitment, expanding measurement, accelerating AI. And she warns that now comes the hard part—turning all of this into precise accountability.
Before trusting the numbers, it helps to know where they come from. There are 1,810 responses from 91 countries, collected between February and April 2026 using targeted and snowball sampling, reaching participants through professional associations, MIT networks, partner lists, and social media. 29% of respondents work in sustainability roles—precisely where interest in the topic is highest. This is not an independent audit of companies: it captures self-reported perceptions and practices. The funders are the foundation of C.H. Robinson, a logistics firm; Breakthrough, a US Venture group company; and CO2 AI with Dassault Systèmes, two tech companies—and each receives a credited quote in the text. Finally, the report itself declares that it was prepared with AI assistance to process data, calculate statistics, generate figures, and draft text, which was subsequently validated by the authors. In an investigation into what AI actually does, that detail belongs in plain sight.
Who Is Accountable for Emissions
The second gap, the report explains, is internal. Among companies with public targets, only 51% have a formal cross-functional team, 37% have a dedicated team with little leverage over the rest of the organization, and 12% entrust everything to a single department. As for oversight, 33% place it with the board of directors, 52% at executive level, and 15% have none at all. Where a cross-functional team exists, 57% say sustainability often or always enters daily decision-making; where responsibility sits in a single department, only 13% say so. A forty-four percentage point gap that the authors read as the return of governance. Yet an association remains: the most committed companies might simply be the ones that build cross-functional teams, and the report speaks of a link, not causation.
The decision-making map is equally telling. Sustainability weighs heaviest where work is immediate: transportation and logistics (39%), procurement (35%), warehousing (34%), product and packaging design (33%). It thins out where decisions span years: capital investments, network design, risk planning. Companies are learning to manage supply chains better, not yet to design them better—and the difference is huge: choosing a carrier changes a quarter, while choosing where to build a facility locks in emissions for decades. The self-reported benefits confirm the tone: regulatory compliance (62%), operational efficiency (61.6%), and supply chain visibility (60.7%) beat cost reduction (51%). In short, the business case holds up through control rather than savings.
Who wins in this scenario? Organizations that built structures capable of surviving shifting political seasons. Who loses? Those that made promises without equipping themselves, because a commitment managed solely at the operational level, the report notes, is easier to drop quietly when conditions tighten. An uncomfortable open question remains: does governance produce seriousness, or does seriousness produce governance?
Local Efficiency, Global Emissions
Artificial intelligence enters this frame with striking figures. Industrial AI adoption—predicting, optimizing, and aiding decision-making—in sustainability initiatives is reported by 78% of respondents in North America, 77% in Europe, and 76% in Asia, excluding China. In the Chinese subgroup, 100% declare adoption—a result the report attributes to sample composition or government push, rather than a population estimate. Among users, 87% report a positive effect, 86% report reduced material waste, 81% report reduced energy or fuel consumption, and 63% use it to estimate missing supplier emissions. As for AI's own carbon footprint, 37% embrace it without reservation, 47% use it while weighing its environmental costs, and 9% avoid or limit it.
The report, however, clarifies the nature of these gains. Those seeing a positive effect attribute it primarily to operational efficiency (57%), followed by emissions reduction (40%), measurement (40%), and, dead last, strategic decisions (27%). Only 25% employ it to simulate future scenarios. AI, therefore, improves the work already being done, but does not yet change which work gets done.
Then comes Figure 31, which flips the script. Among mature users—those with extensive adoption or autonomous, integrated decision-making—53% perceive a reduction in total greenhouse gas emissions (17% strong, 36% slight), 15% see no change, and 29% report an increase (21% slight, 8% strong). The authors clarify that this does not mean AI causes the increases: it means that optimizing a single process and improving net outcomes are two different things. This warrants repeating: these are perceptions, and for this subgroup, the report does not state the sample size.
The idea that optimizing a part does not equal improving the whole has a rich narrative lineage. Ted Chiang, in a 2017 essay on fears surrounding superintelligences, imagined an AI tasked with picking as many strawberries as possible that, to succeed, converts the Earth into strawberry fields; his point was that such a monster resembles corporations pursuing a single metric without regard for anything else. The MIT report invokes no monsters, but its warning is allied: a system squeezing one metric can degrade another if nobody asked it to look beyond. The same risk surfaces in the International Energy Agency's analysis, which cites the rebound effect in transport: autonomous, highly efficient vehicles could draw passengers away from public transit.
Warehouses, or Real-World Factorio
The clearest testing ground is the warehouse, and Factorio provides a fitting analogy: the video game where players build an ever-expanding factory on an alien planet. Every operating machine emits pollution, the cloud attracts hostile fauna, and according to the official documentation, machine pollution accelerates enemy evolution. Automating more means producing more, and until you monitor the pollution meter, success and problem grow hand in hand.
The data—around 368 responses from companies operating warehouses or distribution networks, predominantly large multi-site operations—read as if written by the game's designer. 40% of warehouses operate at an advanced or fully automated level, and among those using AI, 71% allow it to act autonomously, partially or extensively. It is deployed primarily for inventory optimization (64% for industrial AI, 56% for agentic AI capable of executing decisions) and energy management (52% and 45%), while real-time emissions tracking ranks last (36% and 26%). Following AI adoption, 59% of moderately automated warehouses report increased energy use and 54% increased waste; among advanced or fully automated facilities, those figures rise to 78% and 76%. Within this second group, those managing up to ten warehouses report increases in 92% of cases for both categories, whereas those managing more report 65% and 59%, with about a quarter seeing a decrease. The report summarizes: scale appears to soften the toll.
The text is candid about limitations. The data cannot separate two competing explanations: whether autonomous AI drives equipment harder and longer, or whether facilities granting higher autonomy are already the largest and fastest-growing, where energy and waste would have risen regardless. The rebound effect here is a hypothesis, not a proven fact.
The most revealing metric concerns visibility. 88% of warehouses claim to measure sustainability, but only 24% use real-time sensor or AI systems, 49% rely on partially automated software, and another 24% depend on spreadsheets and annual reporting. Note: that 24% figure applies across all warehouses in the sample, not just the most automated ones, despite what press summaries might imply. Consequently, consumption spikes may surface with significant delays. Those reaping gains are operators boosting throughput, as respondents cite floor space and robotics among their strongest economic returns; smaller operators appear most exposed. How much is attributable to AI versus overall growth remains unknown.
Trucks, Suppliers, and Estimates
In transport, AI is not the central protagonist, and the report does not clarify how much optimization actually relies on AI. Among ~348 sector responses, 76% have fully deployed route optimization, while load consolidation, driver programs, and carrier requirements exceed 67%. Yet 43% run fleets entirely or predominantly on diesel, only 19% have crossed the 50% low-emission energy threshold, and just 8% have completed the transition. The report summarizes: phase one extracted efficiency from existing assets; phase two must replace the assets themselves. The divide between Europe and North America comes down to capital: battery electric vehicles are fully deployed in 60% of European companies versus 36% in North America; alternative fuels stand at 63% versus 50%; and 84% of European firms view recent policy impact as positive, compared to 44% in North America. On the other side of the Atlantic, only 16% view policy negatively, while 40% are unsure or see no effect—confusion rather than resistance. Hydrogen is deployed in 6% of cases and excluded by 8%. AI can trim empty miles, but it cannot replace fuel.
In Scope 3—indirect value chain emissions, the largest and hardest part of any footprint—progress is clearer. Participation in joint supplier programs rose from 31% to 58% in a single year; shared supplier agreements grew from 23% to 55%; cross-sector alliances rose from 28% to 47%; and those working entirely alone dropped from 18% to 6%. On tooling, however, the report cautions that 2025 and 2026 questions used different metrics and should not be read as growth: today ~75% routinely use spreadsheets and a similar share dedicated software—spreadsheets haven't gone anywhere.
Here AI plays its most delicate role. 63% use it to estimate missing supplier emissions. Those using AI to track emissions express high confidence in their Scope 3 accuracy in 50.5% of cases, versus 15.3% for non-users; the report establishes no causality. Confidence, however, is not accuracy, and an estimate grounded in sector averages can prioritize action without proving actual reductions. That is our commentary, not a report conclusion. Figures regarding savings from applied AI in the text originate from the CEO of CO2 AI citing BCG and McKinsey estimates: those are sponsor and consultant voices, not survey findings. Winners are those building a clearer picture to target interventions. Unresolved questions remain: who verifies these estimates, and who answers if they are wrong?
The Energy Bill
No evaluation of AI in supply chains can ignore the footprint of AI itself, which the MIT report measures only as an attitude, not a carbon tally. For raw numbers, we turn to the International Energy Agency report from April 2025: in 2024, data centers consumed ~415 terawatt-hours (1.5% of global electricity), projected to reach ~945 TWh by 2030 in base scenarios. Electricity-related emissions would rise from 180 to 300 million tonnes by 2035 (up to 500 Mt in accelerated cases), remaining under 1.5% of energy sector totals but standing among the fastest-growing categories. Conversely, widespread deployment of existing AI solutions could avoid emissions equivalent to ~5% of energy sector totals by 2035—more than data center consumption, but far below total needs. The IEA's dual conclusion: catastrophic climate alarms regarding AI appear overstated, but so are expectations that AI will solve the climate crisis on its own.
A critical counterpoint emerges elsewhere. Analyst Ketan Joshi, in a report commissioned by an environmental coalition, examined 154 public corporate claims regarding AI climate benefits and found 74% unproven; only 26% cited academic research and 36% offered no evidence whatsoever. His most compelling objection is that most claimed gains stem from traditional, far less energy-intensive AI, whereas generative AI impacts are blurred under generic labels that lump everything together. Among sources examined was the IEA, and it should be noted that the commissioning coalition holds an activist stance. The full report is public.
MIT's report does not single out generative AI: it speaks of industrial and agentic AI—decision-making in production, procurement, and warehousing. The comparison is therefore asymmetrical: Joshi targets consumer-facing AI, whereas warehouse data describe industrial operations. This neither refutes nor confirms his thesis, but shows that "AI" as a label is too broad for meaningful accounting. The IEA adds a vital real-world constraint: power grids are already strained, and ~20% of data center projects face potential delays, making energy impacts local long before they become global.
Measure Before Automating
What should an enterprise do? The report prescribes no step-by-step recipe, but its findings suggest a path. Start with structure: a cross-functional team empowered to decide rather than merely advise, and board-level oversight tied to targets, budgets, and risk management. Expand focus beyond transportation and procurement into capital investments and network design, where footprint is locked in for years. Set rules for AI: define metrics in advance (energy, direct/indirect emissions, waste, suppliers), pair automation with continuous monitoring—especially in warehouses—and treat estimates as hypotheses to be verified against primary data. Finally, track net emissions alongside efficiency metrics, because that is where survey data diverge. All of this costs money, and large networks appear better equipped: one may wonder whether AI sustainability will become a privilege of scale.
Counterarguments remain. AI is young—true, but adoption is already widespread. Data are self-reported—correct, which is why we treat them as directional signals rather than proof. Is a lack of data the real problem rather than AI? Partly true, but unmonitored AI risks generating unverified estimates. Is data center energy a separate issue? Not if you evaluate holistic sustainability. Is growing supplier collaboration a good sign? Yes, but it should not be confused with verified decarbonization.
The core thesis can be stated in a few lines. AI is an amplifier: it amplifies efficiency where processes are well designed, and it amplifies opacity, energy consumption, and emissions where governance, data, and accountability are lacking. MIT's report does not claim AI is inherently incompatible with sustainability. It says something far more interesting: a supply chain can become more efficient without becoming more sustainable. The difference is made not by the model, but by who decides what to measure, who answers for results, and which trade-offs they accept to make.
Technical note: Unless otherwise indicated, figures originate from the MIT report (September 2026, version 1) and the MIT press release dated October 15, 2026. These reflect self-reported perceptions from a non-probability sample, and sample sizes for subgroups (mature AI users, warehouses, fleets) are not always disclosed. Causal interpretations and commentary regarding confidence in estimates represent our analysis.