AI cut energy use in some factories—but the review could not predict the next site
A new peer-reviewed review found promising reductions in individual industrial studies, but incompatible baselines, single-site tests and a single-reviewer process ruled out a dependable pooled effect.
The short version
- What happened
- The PRISMA-guided review searched four databases, screened 310 unique records, assessed 95 full texts and retained 42 studies published from 2023 to March 2026.
- Who it affects
- Factory teams may gain from better monitoring or scheduling, but an efficiency project can also add sensors, computing, maintenance and oversight work. Those costs and burdens need measuring alongside energy use.
- What remains unproven
- Screening, eligibility, extraction and quality scoring were conducted by one reviewer, without an inter-rater statistic.
Keep in mind
Screening, eligibility, extraction and quality scoring were conducted by one reviewer, without an inter-rater statistic.
The protocol was not prospectively registered, increasing the risk that decisions changed after results were seen.
The 42 studies used incompatible baselines, system boundaries and outcomes, so the review reports no pooled estimate.
English-language journal articles dominated the evidence; small firms and developing economies were underrepresented.
Source analysis · 1 linked source. Primary analysed source date: 25 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Promising examples did not add up to a transferable number
Researchers from Lebanon, Kuwait and France reviewed how artificial intelligence, the Internet of Things, digital twins and data analytics were linked to resource efficiency and emissions in operations. Searches of Scopus, Web of Science, ScienceDirect and IEEE Xplore on 15 March 2026 found 420 records. After 110 duplicates, the team screened 310 records, assessed 95 full texts and retained 42 peer-reviewed studies.
Some results were meaningful inside their own settings: a physics-guided neural network was associated with 4.21% lower energy use across two industrial cases, while digital-twin scheduling cut automated-guided-vehicle energy use by 7.62%. Another machine-level monitoring study reported 51.5%, and a pelleting-press model predicted 6.0%. The review correctly treats these as named observations, not a common effect. Different baselines, system boundaries and definitions make a pooled percentage misleading.[1]
The review maps options; it does not price a retrofit
For plant managers and workers, the practical message is to measure the whole workflow before buying on headline percentages. A local test should define the machine or process boundary, compare against the existing operation, record production quality and downtime, and include the electricity, water and hardware burden of the digital system itself. Otherwise an apparent efficiency gain can shift cost or emissions elsewhere.
Confidence is limited by the review process. One reviewer handled screening, eligibility, extraction and a bespoke quality score, with no independent second assessment or inter-rater statistic. The protocol was not prospectively registered, English-language journal articles only were eligible, and evidence for smaller firms and developing economies was sparse. The authors report no funding or competing interests. Stronger evidence would require preregistered dual-reviewer synthesis, common sustainability measures and prospective multi-site comparisons.[1]
The distinction that matters
Why one factory's saving is not the next factory's forecast
The review moves from a broad search to context-specific results, then stops before an unjustified average.
Find
Search four databases and remove duplicates
310 unique records screened
Assess
Read eligible reports in full
95 full texts assessed; 42 studies retained
Interpret
Keep results tied to each site's boundary
No pooled effect or cross-study percentage range