Executive Overview
The race to achieve corporate net-zero targets has run headlong into one of the most notoriously arduous bottlenecks in modern supply chain management: calculating Product Carbon Footprints (PCFs). For decades, sustainability professionals have bogged down in the grueling, manual labor of Life Cycle Assessments (LCAs)—hunting down elusive data from tier-one through tier-four suppliers, cross-referencing industry databases, and reconciling conflicting material specifications. It is a process so resource-intensive that recent data from PwC reveals a startling reality: nearly 70 percent of companies have managed to calculate PCFs for less than a quarter of their actual product lineups.
Into this breach has stepped artificial intelligence. A burgeoning market of automated platforms—ranging from specialized startups like Makersite and Terrascope to established enterprise carbon accounting suites like Watershed—now promises a technological silver bullet. By feeding a simple product description or a bill of materials into a large language model (LLM) or specialized predictive algorithm, businesses are told they can generate comprehensive carbon metrics in seconds rather than months.
Yet, beneath the glossy marketing promises and impressive top-line numbers lies a troubling disconnect. A landmark new study led by Krishna Rao and researchers at Watershed exposes a critical vulnerability in the current generation of generative AI and machine learning tools applied to sustainability. While these models can occasionally hit close to the mark on a final aggregate emissions figure, their internal reasoning is riddled with compounding errors. When researchers pulled back the hood to examine the intermediate steps—how the AI broke down a product’s components and assigned emissions values to individual sub-assemblies—accuracy plummeted to as low as 37 percent.
This investigation explores the mechanics of AI-driven PCF automation, the stark divergence between impressive macro-level outputs and deeply flawed micro-level analyses, and the profound implications this has for corporate decarbonization strategies. As regulators tighten greenwashing rules and stakeholders demand audit-grade supply chain data, the message from the research is unequivocal: in the realm of carbon accounting, trusting an AI’s final number without inspecting its math is a dangerous gamble.
Detailed Chronology: The Evolution and Testing of AI Carbon Accounting
The Manual Bottleneck (Pre-2023)
For decades, Life Cycle Assessments remained the gold standard of environmental accounting, yet their operational friction limited their widespread adoption. Conducting an LCA meant tracing every raw material extraction, manufacturing process, transportation leg, and disposal phase associated with a product. For complex goods like smartphones, automobiles, or industrial chemicals, this meant managing spreadsheets containing thousands of line items sourced from dozens of global vendors. Because suppliers often guarded proprietary data or lacked emissions tracking infrastructure themselves, analysts relied heavily on generalized industry averages. The immense time and capital required made comprehensive portfolio-wide carbon accounting an exclusive luxury for only the largest multinational corporations with dedicated ESG divisions.
The Generative AI Boom and the Promise of Automation (2023–2025)
As generative AI capabilities exploded across industries following the mainstreaming of large language models, enterprise software developers quickly recognized a prime use case in sustainability. Tech startups and established carbon platforms began integrating LLMs and proprietary predictive models designed to bypass the supplier-data bottleneck. Companies like Makersite marketed tools capable of automating accurate LCAs across entire product portfolios in mere seconds. Competitors like Terrascope claimed they could achieve up to 70 percent accuracy without needing to badger suppliers for primary data. The narrative was seductive: artificial intelligence could synthesize chemical formulas, material weights, and global supply chain pathways instantaneously, democratizing carbon accounting for small and medium-sized enterprises (SMEs) while saving Fortune 500 sustainability teams thousands of man-hours.
The Watershed Benchmark Study (2025–2026)
Skeptical of the black-box nature of these rapid outputs, a research team at Watershed—led by Krishna Rao—decided to rigorously test the capabilities of leading foundation models when applied to carbon accounting. The researchers assembled a diverse test bed of 175 products spanning multiple industrial sectors, including advanced electronics, heavy chemicals, specialized textiles, and raw materials.
They submitted product descriptions and material lists to frontier AI models developed by Anthropic, DeepSeek, Google, and OpenAI. The goal was twofold: evaluate how closely the models could replicate expert-derived carbon footprints, and benchmark the internal logic used to arrive at those conclusions. Published in early 2026, the study’s findings sent a ripple through the corporate sustainability community. While the best-performing models landed within two multiples of the human expert benchmark 77 percent of the time on final figures, the granular breakdown of how they reached those figures revealed profound, hidden inaccuracies. The illusion of automated perfection was officially broken.
Supporting Context & Metrics: The Anatomy of a Hidden Failure
To understand why the Watershed study’s findings are so consequential, one must examine the operational realities of corporate decarbonization. A product carbon footprint is rarely calculated solely for corporate ESG reporting or marketing brochures; it is an active operational tool meant to guide procurement, engineering, and supply chain decisions.
Macro Precision vs. Micro Chaos
When a consumer electronics manufacturer sets out to reduce the carbon footprint of a new laptop, for instance, a macro-level emissions score is largely useless on its own. Knowing that the device generates 250 kilograms of $textCO_2$ equivalent ($textCO_2texte$) tells the design team nothing about where those emissions originate. To drive actual reductions, engineers must interrogate the supply chain: Is the aluminum chassis the primary culprit, or is it the rare-earth metals in the battery? Should the company drop its current vendor in Shenzhen for a lower-carbon supplier in Vietnam?
This is where the Watershed benchmarking exposed a fatal flaw. While an AI model might correctly guess a total product footprint by leveraging aggregate sectoral training data, the intermediate steps—decomposing the product into specific hardware components, assigning material weights, and calculating localized manufacturing emissions—suffered from catastrophic degradation. When Rao and his colleagues evaluated the step-by-step logic chains of the models, accuracy plummeted down to 37 percent.
[Product Description / Bill of Materials]
│
▼
[AI Model Analysis]
│
┌─────────┴─────────┐
▼ ▼
[Final Macro Number] [Intermediate Steps]
(Looks Impressive: (Catastrophic Drop:
~77% within range) Accuracy sinks to 37%)
│ │
▼ ▼
(Good for high-level (Flawed component data
general reporting) leads to failed procurement
decisions & wasted capital)The Cost of Compound Errors
This disparity creates a dangerous operational trap. If an enterprise relies on an AI tool to identify emission hot spots, a 37 percent component accuracy rate means that nearly two-thirds of the sub-assembly breakdowns are unreliable. A manufacturer could spend months auditing a specific supplier, renegotiating contracts, or redesigning a component based on ghost data generated by an LLM hallucinating emissions factors for obscure chemical precursors or specialized alloy grades.
Furthermore, the study highlights the risks of relying on models that rely on pattern matching rather than thermodynamic or chemical accounting. Large language models are trained to predict the next most likely token based on vast corpuses of internet text. When asked to estimate the emissions of a novel material or a proprietary manufacturing process, the AI may generate a plausible-sounding figure simply because it looks statistically similar to standard manufacturing data found in its training set, completely masking a severe error in its underlying logic.
Official Statements and Industry Expert Perspectives
The release of the Watershed study has sparked vital conversations across the tech and sustainability sectors regarding the limits of automation in environmental science. Industry leaders and researchers are increasingly urging caution against uncritical tech adoption.
In an interview detailing the study, lead researcher Krishna Rao expressed astonishment at the sheer magnitude of the performance gap between top-line outputs and granular analysis.
"What surprised me most was the size of the gap," said Krishna Rao, researcher at Watershed. "You can look at a final score and think the model has mastered the product’s life cycle. But when you examine the steps it took to get there, the reasoning often falls apart completely."
Rao emphasized that while his team designed the benchmarking process explicitly to help evaluate specialist AI-powered PCF tools—including those offered by Watershed itself—the immediate takeaway for the broader market is clear. Enterprises must shift away from treating carbon accounting software as a black-box oracle.
"For now, our message to users is clear: do not rely on final results alone," Rao advised. "To make actionable business decisions, users have to understand the intermediate steps. If you cannot audit the math, you cannot trust the reduction strategy."
Other industry voices echo these sentiments, warning that as regulatory scrutiny intensifies under frameworks like the European Union’s Corporate Sustainability Reporting Directive (CSRD) and strict anti-greenwashing laws enforced by bodies like the FTC and ESMA, companies using unverified AI-generated carbon metrics could face severe legal and financial liabilities. Auditors will not accept "the AI told us so" as a valid defense for erroneous Scope 3 supply chain reporting.
Future Outlook: The Road Ahead for AI and Sustainability
Despite the sobering revelations of the Watershed study, artificial intelligence is not destined to remain a blunt or deceptive instrument in carbon accounting. The trajectory of enterprise software suggests that the industry is entering a more mature, rigorous phase of AI integration.
1. The Shift Toward Hybrid Architectures
Pure generative AI—relying solely on unconstrained large language models—is proving ill-suited for the strict mathematical and physical realities of Life Cycle Assessments. The future belongs to hybrid architectures that couple the natural language processing capabilities of LLMs with deterministic, rules-based calculation engines and verified database backends (such as Ecoinvent or US LCI). In these systems, the AI acts strictly as an intuitive user interface—parsing unstructured bills of materials or supplier emails—while verified databases and thermodynamic equations handle the actual emissions calculations.
2. Audit-Ready Traceability and Explainable AI (XAI)
As enterprise clients demand transparency, software vendors are being forced to build "explainable AI" features directly into their sustainability suites. Future PCF platforms will not simply spit out a final tonnage of carbon; they will provide a fully traceable audit trail, highlighting exactly which database was queried, what assumptions were made for missing supplier data, and where confidence intervals are high or low. This level of traceability will be essential for passing third-party sustainability audits.
3. Closing the Primary Data Gap
Ultimately, artificial intelligence in carbon accounting must be viewed as a bridge, not a permanent destination. While tools like Makersite, Terrascope, and Watershed help companies triage massive product portfolios and prioritize high-impact categories, they cannot permanently replace primary data collection. Over the coming years, regulatory pressures and digital product passports (DPPs) will compel suppliers throughout global value chains to measure and share verified emissions data directly. As primary data becomes standardized, AI’s role will shift from guessing missing values to managing, validating, and optimizing mountains of real-world supply chain telemetry.
Conclusion
The integration of artificial intelligence into product carbon footprinting represents one of the most tantalizing frontiers in corporate sustainability. The promise of automating a notoriously tedious, resource-prohibitive process holds immense appeal for businesses struggling to meet aggressive climate goals. Yet, as the Watershed study powerfully demonstrates, convenience must not eclipse rigor.
Impressive top-line numbers can easily camouflage a cascade of internal errors, turning strategic decarbonization initiatives into exercises in chasing phantom data. For sustainability professionals, procurement officers, and corporate executives, the path forward requires a balanced approach: embracing AI to unlock operational scale and speed, but demanding absolute transparency and auditability every step of the way. In the high-stakes arena of global climate action, precision and truth will always outweigh speed and illusion.
