Table of Contents
Industrial AI projects usually fail for organisational and data reasons, not because the algorithm is wrong. The common causes are a vague problem with the wrong success metric, data that is not ready, pilots never designed for production, weak integration with plant systems, operators who do not use the output, and models that drift after go-live.
Key takeaways
- By estimates cited by RAND, more than 80% of AI projects fail, about twice the rate of IT projects without AI. In RAND's interviews, misunderstanding the problem was the most common cause.
- In 2019, the World Economic Forum and McKinsey reported that more than 70% of industrial companies were stuck in "pilot purgatory" with Fourth Industrial Revolution technologies.
- BCG attributes about 70% of AI implementation challenges to people and processes, 20% to technology and only 10% to algorithms.
- Plant data is "often abundant but poor in information" (Lawrence et al., 2024), and labelled examples such as lab results or recorded failures are scarce.
- Models lose accuracy when the process changes. Joint guidance led by CISA calls for monitoring, failsafe fallback and operator training.
How often do industrial AI projects fail?
There is no reliable failure rate for industrial AI, and we found none for Algeria or the MENA region. The available figures come from surveys and interviews, each defining failure differently.
- RAND (2024) cites estimates that more than 80% of AI projects fail. Its own research is qualitative: interviews with 65 data scientists and engineers with at least five years of experience (Ryseff et al., RAND).
- WEF and McKinsey (2019) found that only 29% of industrial companies were deploying Fourth Industrial Revolution technology at scale and 30% had yet to pilot. This covers digital manufacturing technologies broadly, not only AI (WEF, 2019).
- BCG (2024) surveyed 1,000 executives in 59 countries and found that 74% of companies "have yet to show tangible value from their use of AI" (BCG).
Read these as a warning, not a forecast. What they share is that most projects stall between a working demonstration and daily use.
What are the most common causes?
| Cause | What it looks like on site | Early warning sign |
|---|---|---|
| Unclear problem or wrong metric | An accurate model that changes no decision | Nobody can name the KPI the project will move |
| Data not ready | Months spent finding, cleaning and aligning tags | Lab results or failure records are on paper or in spreadsheets |
| Technology chosen before the problem | "We need an LLM" or "we need a digital twin" as the starting point | No simpler method was compared |
| Pilot not built for production | A demo on one engineer's laptop | No owner, budget or infrastructure for after the pilot |
| No integration with plant systems | Operators copy values between screens | No named interface to the historian, DCS or ERP |
| Operators do not use it | Recommendations ignored or switched off | Operators were not involved in defining the use case |
| Drift after go-live | Accuracy falls months later, after a process or feed change | Nobody is assigned to monitor the model |
RAND's five root causes cover much of this table: a misunderstood problem, missing data, technology over users, inadequate infrastructure, and problems too difficult for AI (RAND). Industry adds OT integration, operator trust and process drift.
Why does a vague problem sink the project?
RAND found that misunderstanding or miscommunicating the problem is the most common reason AI projects fail. Models are deployed that have been "optimized for the wrong metrics" or do not fit the business workflow (RAND).
Model accuracy is not business value. A survey of machine learning deployments describes how Booking.com, after deploying 150 models, found that better model performance did not always translate into business value, partly because proxy metrics did not convert into the business metric (Paleyes et al., 2022). In a plant, the equivalent is a model that predicts a quality variable well but reduces no off-spec product, because nobody changes a setpoint based on it.
Fix the metric before the model. Rule #2 of Google's Rules of Machine Learning is "First, design and implement metrics." Name the decision the model will change, who makes it, and today's baseline. Our guide to building an AI business case turns that into figures.
Why is plant data rarely ready?
A survey of machine learning in the process industries describes process data as "often abundant but poor in information" because of disturbances, outliers and missing values (Lawrence et al., 2024). It names the lack of labelled data as the main challenge in soft sensor development, since quality variables are often sampled once per 8-hour shift or once a day.
The CISA-led guidance adds data silos caused by OT/IT network segregation, proprietary protocols and many suppliers, and the difficulty of capturing safety-related edge cases, for which "OT operator expertise is necessary" (CISA et al., 2025). Failure records are a common gap: a failure prediction model needs examples of failures, and a well-maintained plant has few.
Check this before signing anything. An AI readiness assessment scores the historian, tagging and records.
Is AI the right tool at all?
RAND's third and fifth causes are chasing the latest technology instead of the user's problem, and applying AI to problems too difficult for it (RAND). The CISA-led guidance asks operators to assess whether AI is the most appropriate solution and to consider "whether an established capability meets their needs" before pursuing more complex AI (CISA et al., 2025).
In a process plant, the established capability might be a better-tuned control loop, an alarm review, a first-principles calculation or a fixed rule. Rule #1 in Google's guide is "Don't be afraid to launch a product without machine learning" (Google). If a simple method gets you most of the way, it also becomes the baseline the AI must beat.
Why do pilots never reach production?
The WEF found that pilots were "often too long and/or complex": 56% of respondents reported pilots of one to two years and 28% of more than two years. Companies struggled to scale because of difficulty aligning on value and return on investment, uncertainty about the value, and the cost to implement and scale (WEF, 2019).
Part of the reason is that the model is the small part. In a widely cited paper on technical debt in machine learning systems, Google engineers note that "only a small fraction of real-world ML systems is composed of the ML code"; the surrounding infrastructure "is vast and complex" (Sculley et al., 2015). Google Cloud's MLOps guidance describes the least mature stage as a manual, script-driven process with infrequent releases and no active performance monitoring (Google Cloud).
RAND recommends up-front investment in infrastructure for data governance and model deployment, and committing a team to one problem for at least a year (RAND). Put the production phase in the pilot plan: who hosts the model, who maintains it, and what it costs after year one. Our post on industrial AI project cost breaks those costs down.
Why does integration with OT systems stall projects?
A model that cannot read live data, or deliver its output where operators work, stays a report. Connecting it to a historian, DCS or SCADA also crosses the boundary that protects the control network. The CISA-led guidance recommends push-based architectures, where data is pushed out of the OT network without persistent access into it, new connection paths through the existing DMZ or jump host, and testing AI systems for safety impacts such as latency, like any new OT system (CISA et al., 2025).
Involve OT engineers and cybersecurity from the first week, not after the pilot. See integrating AI with SCADA and OT systems.
Why do operators ignore the output?
The survey of deployment case studies notes that machine learning "is often met cautiously by the end users", that time must be invested to build trust, and that involving the audience early is an efficient way to do it (Paleyes et al., 2022).
The CISA-led guidance also warns of the opposite risk: overreliance that erodes manual skills, and skill gaps where personnel misinterpret AI outputs. It recommends training OT teams to interpret and validate AI outputs, procedures that cover AI-related operations, and failsafe mechanisms that revert to traditional automation or manual operation. For safety reasons, it advises limiting active control of OT by AI without a human in the loop (CISA et al., 2025).
In practice, operators are more likely to use a tool they helped define, that shows why it recommends something, and that they can switch off.
Why do models degrade after go-live?
A model learns the plant as it was. The CISA-led guidance lists model drift as a risk: models "may become less accurate over time" when they see data not represented in their training data, and alterations to production processes can affect model performance (CISA et al., 2025). Catalyst ageing, a new feed or ore blend, a replaced instrument or a debottlenecking project can all cause it.
Google Cloud's guidance is to actively monitor model quality in production and retrain on recent data (Google Cloud). For an industrial model, that means a named owner, a comparison with lab or inspection results on a fixed schedule, and a defined fallback. Our soft sensor guide shows this for one type of model.
How do you check a project before it starts?
- Write the problem as a decision. Which decision changes, who makes it, which KPI moves, and what the baseline is today.
- Compare with a simpler method. A rule, a calculation or better control tuning. If it is good enough, stop there.
- Audit the data. Historian coverage, tag consistency, the number of labelled examples (lab results, failures) and their timestamps.
- Involve the users. Operators, maintenance staff and process engineers help define the output and how it is displayed.
- Agree the pilot protocol in writing. Baseline, metric, test period, pass mark and duration, before the pilot starts. Our vendor selection guide lists the questions to put to suppliers.
- Design the production path. Hosting, the integration route through the DMZ, the security review, and the budget after year one.
- Assign an owner for after go-live. Monitoring, retraining triggers, the fallback, and who can switch it off.
The NIST AI Risk Management Framework, released in January 2023 for voluntary use, gives a broader structure for managing these risks.
What we have seen at KABAS AI
Our two published projects are R&D projects, not control-room deployments, but they follow the order of work described above. At ENAD Shymeca, the project began with a discovery phase alongside the R&D team, then configured Virtual Lab for detergent-specific parameters, migrated historical formulation data and trained R&D staff. The case study reports a 40% reduction in physical testing cycles. At ADFC, comparisons between simulated and actual results guide refinements, and the deployment sits within Sonatrach's IT environment.
Two lessons apply beyond R&D. Keep the physical measurement as the reference, whether it is a lab test, an inspection or an analyser. And decide early where the system will run: when data must stay on site, hosting shapes integration, hardware and cost. DALILE and Virtual Lab can both be deployed on the customer's own infrastructure.
Where should you start?
If a project has stalled, or is about to start, review it against the seven steps above before funding another pilot. Our consulting team can run that review with your operations, IT and management leads, and you can contact us to discuss a specific case.
Sources
- Ryseff, De Bruhl and Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND, August 2024
- World Economic Forum with McKinsey & Company, Fourth Industrial Revolution: Beacons of Technology and Innovation in Manufacturing, January 2019
- BCG, AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, October 2024
- Paleyes, Urma and Lawrence, Challenges in Deploying Machine Learning: a Survey of Case Studies, ACM Computing Surveys, 2022 (arXiv 2011.09926)
- Lawrence et al., Machine learning for industrial sensing and control: A survey and practical perspective, Control Engineering Practice, 2024 (arXiv 2401.13836)
- CISA, ASD's ACSC et al., Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- Sculley et al., Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015
- Google Cloud, MLOps: Continuous delivery and automation pipelines in machine learning
- Google for Developers, Rules of Machine Learning
- NIST, AI Risk Management Framework
Frequently asked questions
What percentage of AI projects fail?
There is no single reliable figure. RAND reports that, by some estimates, more than 80% of AI projects fail, about twice the rate of IT projects without AI. BCG found in 2024 that 74% of companies had yet to show tangible value from AI. Each source defines failure differently, so treat these numbers as a warning rather than a forecast.
What is pilot purgatory?
Pilot purgatory is the situation where a company runs pilots of new digital technologies but never deploys them at scale. A 2019 World Economic Forum white paper written with McKinsey reported that more than 70% of industrial companies were in this situation with Fourth Industrial Revolution technologies, and that many pilots lasted one to two years or longer.
Is poor data the main reason industrial AI projects fail?
It is one of the main reasons, but not the only one. RAND's interviews found that misunderstanding the problem to be solved was the most common cause, and lack of the necessary data was another of its five root causes. In plants, process data is often plentiful but noisy, and labelled examples such as lab results or recorded failures are scarce.
Can an AI project fail even when the model is accurate?
Yes. A model can predict well and still change nothing if no decision depends on it, if operators do not use it, or if it is never connected to the systems where people work. Booking.com found that better model performance did not always translate into business value, partly because proxy metrics did not convert into the business metric.
How do we keep an industrial AI model accurate after go-live?
Assign a named owner, compare the model's output with lab, inspection or analyser results on a fixed schedule, and retrain when the process, feed or equipment changes. Keep a fallback to the previous method or manual operation. CISA-led guidance lists model drift as a specific risk when AI is used with operational technology.



