AI in Industry: What Works and What Needs Work

Progress is evident in prominent AI techniques, but more work is needed

Traditional AI excels at making predictions based on existing data and predetermined rules to automate well-defined, routine or tedious tasks. It analyzes historical and real-time data reactively for patterns, correlations, anomalies and trends to improve decision-making and enable predictive actions.

• What works: Among the plethora of Industrial AI use cases today, predictive maintenance stands out as one of the most commonly implemented, says Vivek Murugesan, senior research associate at LNS Research, a provider of research and advisory services. “Successful implementation of predictive maintenance solutions often lays the foundation for more complex use cases like quality, batch optimization, energy management, etc.”
• What needs work: Traditional AI (or any AI for that matter) is only as effective as the data used to train its algorithms. Incomplete, inaccurate, and noisy source data, along with a lack of contextualization, impact the quality and reliability of predictions. Traditional AI also struggles with unstructured and complex data sources and can inherit and perpetuate historical data biases. Active human oversight and data quality management are paramount.

Generative AI (GenAI) swiftly and proactively creates new content — such as text, images, audio and video on demand — based on provided prompts, by learning patterns from data. It enhances human-machine collaboration and powers tools like AI assistants and copilots. Output is improved with high-quality initial and follow-up prompts, high-quality data and human review.

• What works: GenAI-produced output ranges from compliance reports to personalized training materials and medical treatment plans. Retrieval-augmented generation (RAG) helps augment and enhance the accuracy and reliability of GenAI models while reducing hallucinations.
• What needs work: According to Gartner, AI leaders continue to face challenges when it comes to proving GenAI’s value to the business: “Low-maturity organizations have trouble identifying suitable use cases and exhibit unrealistic expectations for initiatives. Mature organizations struggle to find skilled professionals and instill GenAI literacy,” notes Khandabattu.

Edge AI moves AI processing to local devices such as Internet of Things (IoT) sensors, robots and smartphones, bringing it closer to the data source. Deploying AI models on edge devices enables faster data processing and analysis, enhances privacy and security by processing sensitive data locally, and reduces bandwidth by minimizing data sent to distant data centers or the cloud.

• What works: Applications for worker safety (e.g., hazard detection and restricted zone monitoring), smart cities (e.g., traffic optimization and air and water quality control) and healthcare (e.g., wearable health monitors and faster image analysis) are examples proving fruitful.
• What needs work: Edge devices present unique cybersecurity and maintenance challenges and are limited in processing power, memory, storage and battery capacity. Understanding and mitigating these issues while also maintaining data quality and human oversight is vital. 

Agentic AI systems are goal-oriented and able to reason, plan and act independently. Their AI agents have the capacity and agency to autonomously accomplish complex, multi-step tasks and make real-time decisions to solve problems, with minimal human intervention. AI agents are among the two fastest-advancing technologies, the other being AI-ready data, according to Gartner.

• What works: One use case is enhancing predictive maintenance systems by analyzing sensor data to predict asset failure, identify the cause, generate work orders and schedule repairs. Another is using internal system data, supplier inventory levels, demand signals and weather forecasts to continuously forecast demand.
• What needs work: Khandabattu says, “The complexity of AI agents makes them vulnerable to access security, data security and governance issues. Organizations also exhibit a lack of true trust in AI agents’ ability to operate without human oversight and concern about the significant impact of potential errors.”

Causal AI, still at a nascent stage, models cause-and-effect relationships to understand why things happen and apply that reasoning in its decision-making. Murugesan notes its emergence in manufacturing, saying: “Causal AI, which has its roots in the fintech world, is increasingly used to find underlying causes (as opposed to correlations and patterns) to uncover how changes in one variable will influence others on the factory floor.”

• What’s ahead: Two likely use cases for Causal AI include improving root cause analysis and offering prescriptive guidance to prevent future incidents, and powering “what if” scenarios to optimize supply chains. 

A simple, strategic approach eases the journey to AI success

Due to their novelty, AI technologies remain rife with unknowns regarding safety, security, reliability and economy. Conscious effort is necessary to overcome data quality deficiencies, distrust, ethical concerns and security vulnerabilities. AI aging, self-preservation tendencies and workforce skills erosion deserve attention, and human oversight must be maintained. Additionally, the technical and organizational costs associated with AI technical debt will be elevated unless the system is well implemented.

An AI strategy that recognizes and addresses these challenges and leverages the KISS principle while the technology matures is more likely to achieve sustainable business value, long-term growth and a competitive edge.