AI in Mechanical Integrity Programs: Real Applications vs. Vendor Hype
According to Inspectioneering Journal, the promise of AI transforming mechanical integrity programs is loud at industry conferences right now, but the reality is considerably more complicated. Writing in the May/June 2026 issue, Matthew K. Caserta, PE, Manager of the Corrosion, Materials, and Integrity Division at Becht, opens with a sobering data point: a recent MIT study (as cited by Caserta) found that 95% of AI pilots fail to achieve their promised results. For inspection and maintenance contractors working in refineries, petrochemical plants, and other fixed-equipment facilities, that number deserves serious attention before the next vendor pitch arrives.
Background
Caserta’s article frames the central question clearly: the issue isn’t whether AI can contribute to mechanical integrity work, it’s where, when, and how. The piece lays out working definitions of the key technologies in play, including artificial intelligence broadly, machine learning (ML), and large language models (LLMs) such as ChatGPT, Google’s Gemini, and Meta’s LLaMA. Machine learning, as Caserta describes it, allows models to identify patterns from historical data rather than relying on fixed rules, making predictions that improve over time. LLMs, by contrast, process and generate human-like text responses from massive training datasets.
The article notes that AI is already embedded in everyday industries, from fraud detection in financial services to customer service chatbots in retail to navigation systems in autonomous vehicles. The natural follow-on question for the MI community is how that same computing power applies to fixed equipment inspection, corrosion management, and integrity program decision-making.
Where AI Actually Works, and Where It Doesn’t
The framing Caserta uses is valuable precisely because it resists the binary debate that dominates most AI coverage. The conversation in field operations too often splits into two camps: true believers who treat AI as a near-term replacement for experienced inspectors, and skeptics who dismiss it outright. Neither position serves the people actually running inspection programs or managing corrosion circuits at a processing facility.
The 95% pilot failure rate from MIT is the most important number in the publicly available portion of this article, and it points to a structural problem that goes beyond technology. AI pilots in industrial settings typically fail not because the algorithms are bad, but because the underlying data is inconsistent, incomplete, or poorly organized. Mechanical integrity programs at aging facilities often carry decades of inspection records in formats that range from paper logs to incompatible software systems. Feeding that kind of data into a machine learning model and expecting reliable output is optimistic at best.
This is where the distinction between ML and LLMs matters practically. Machine learning tools that flag anomalies in thickness measurement trends or predict corrosion rates under specific operating conditions require clean, structured, historical datasets. That’s a real data management challenge for most facilities, but it’s a solvable engineering problem. LLMs, on the other hand, can potentially help inspectors navigate complex code requirements, summarize inspection histories, or draft fitness-for-service justifications. The risk there is different: LLMs can produce confident-sounding answers that are technically wrong, which in an MI context isn’t just an inconvenience, it’s a safety issue.
The broader implication for the inspection and maintenance contracting community is that AI doesn’t eliminate the need for qualified human judgment, it shifts where that judgment gets applied. Experienced inspectors and corrosion engineers may spend less time manually sorting through data and more time reviewing and challenging AI-generated outputs. That’s a meaningful change in workflow, but it’s not the “hands-off MI decisions” that vendor marketing decks promise.
For owners and operators pushing contractors to adopt AI tools, the Caserta article implicitly raises a governance question: who is accountable when an AI-assisted inspection recommendation turns out to be wrong? Under OSHA’s Process Safety Management standard (29 CFR 1910.119), mechanical integrity is a named program element with specific owner-operator obligations, but the standard doesn’t address AI-generated outputs. That regulatory gap doesn’t have a clean answer yet, and it’s one that subcontractors should be asking about before they sign on to any AI-integrated scope of work.
What It Means for Subcontractors
- The 95% AI pilot failure rate cited in Inspectioneering Journal is a useful number to have in your back pocket when a client or vendor is pushing AI adoption on an accelerated timeline. Ask what the data quality looks like before the pilot starts.
- Subcontractors who own or manage inspection data, thickness readings, CML records, inspection histories, are in a better position to benefit from ML tools. If your data is a mess, AI won’t fix it, and will likely amplify it.
- LLM tools may offer near-term value for document-heavy tasks like report drafting, code lookups, or inspection history summarization, but any output touching safety-critical decisions needs qualified engineer review. Don’t let a client treat AI-generated recommendations as a substitute for a PE stamp.
