The convergence of Artificial Intelligence and the Internet of Things (AIoT) promises transformative efficiency and predictive capabilities. We expect to see a major uptick in IoT projects in the next 2-3 years as AI-driven operational support and analysis takes hold.
Indeed, we expect there to be 30 billion or more connected IoT devices installed by 2028-2029, as instrumentations and autonomous objects proliferate. With enterprises in manufacturing, healthcare, transportation, utilities and more increasingly focused on operational improvements, incrementing a wide range of operational components with the promise of AI being able to interpret the data created and then operate those devices in an autonomous fashion for corporate benefit is a major opportunity.
But despite this promise, and sometimes unproven hype, moving from a proof of concept experiment to factory-scale deployment remains the industry's greatest hurdle. According to major industry benchmarks, nearly three-quarters of organizations fail to transition their AIoT projects into production, and approximately 75% of IoT pilots never scale out to full production. This is caused by several factors, including:
• Pilots being implemented without total organizational Buy-In which creates a major stumbling block to success
• Choosing the most effective AI model takes time and effort and often requires using customized and/or specialized smaller models
• Proof of Concepts (PoC) often don’t show a realistic Return on Investment (ROI), which is required for full scale implementation
• And often a Proof of Concept doesn’t prove the concept at all!
Getting Beyond AI-IoT Failures
More broadly, AI IoT failures are often created by a lack of synchronization between data science, hardware engineering, and business strategy. Such roadblocks are common in many enterprise application areas, but are particularly critical when implementing any AI directed solutions, like IoT. Further, governance, usability and security are often undermined by lack of a full product and process specification/plan. This makes it difficult to achieve a true ROI for the organization, and often makes any proof of concept a “shot in the dark” at achieving expected goals. It also creates a potential security concern that can result in not only failure of operation, but potentially physical damage to both people and things.
Primary reasons for failures
There are a series of failure mechanisms at play, with each organization having some of their own unique mechanisms. Below we highlight the major, common challenges that affect most enterprises when trying to deploy any IoT solution.
Data Quality and "Data Silos"
AI is only as good as the data it consumes. In IoT, data often originates from heterogeneous sensors with different formats, sampling rates, and calibrations. This can result in Dirty Data from Sensors in the field which are subject to environmental wear, leading to noise, drifts, or complete outages that "poison" the AI models. Further, supervised AI learning requires accurately labeled data. In many IoT environments, it is difficult or expensive to obtain accurate labels for events (e.g., exactly when a machine began to fail). Finally, data is often scattered into many data silos, making a consolidated view difficult to achieve.
The "Cloud-Only" Fallacy
Many projects fail because they assume all data must be processed in the cloud. A cloud-only approach causes several issues, including:
• Latency Issues: since real-time IoT applications (like autonomous robotics or safety shut-offs) cannot wait for a round-trip to a cloud server.
• Excessive Bandwidth Costs: Streaming raw data from thousands of sensors 24/7 is prohibitively expensive. Projects often fail when they realize the operational expenditure (OPEX) of data transmission outweighs the ROI.
• OPEX Explosion: Projects often stall when the cost of cloud transmission exceeds the operational ROI. And Cloud processing costs often exceed budgeted costs.
In reality, many, if not most, AI IoT solutions are better implemented by a hybrid approach by deploying Edge AI for processing data locally, and only sending insights to the cloud as necessary for secondary analysis.
Power and Resource Constraints
Running complex AI requires significant computational power, and power is one of the major restrictions associated with IoT. The result is a Power Paradox where complex AI inference can reduce IoT battery life without implementing hardware-specific optimization. Running a continuous AI inference engine can reduce a device's lifespan from years to days. Further, standard AI models are too large for the low power microcontrollers (MCUs) typically used in IoT and newer more capable processors often exceed cost and/or size restrictions. Finally, Parameter Creep often comes into play as initial projected operations increase as more features get added, resulting in reduced performance and stretched processing resources as well as increased AI model size.
Scalability
What works for five devices in a lab rarely works for 50,000 in the field. This is further complicated by Deployment Logistics, including a significant technical challenge of Updating AI models over-the-air (OTA) across a geographically dispersed fleet. It may also be affected by Environment Variance, since a model trained on a sensor in a climate-controlled lab may fail when deployed in a humid factory or a freezing outdoor environment.
Lack of Domain Expertise
Many IoT projects start with a technological view rather than a problem-solving view of what’s required. Adding technology “because we can” is a recipe for failure. And data scientists may build a sophisticated deep-learning model where a simple physics-based equation or a basic heuristic would have been more robust and explainable. Technology-first projects often ignore the actual problem that needs solving in a cost-effective manner and often create "solutions in search of a problem." Success requires merging Data Science with deep Industrial Knowledge and full partnership with the actual user community. Indeed, if the AI provides "insights" that teams don't find actionable or trustworthy, the system will be ignored and the project will face ultimate rejection at the end user level.
Security and Privacy
Adding AI to IoT projects increases the attack surface. This can result in Adversarial Attacks where malicious actors manipulate sensor inputs to trick AI models into making wrong decisions. Further, any solution must take into account Privacy Compliance, especially in consumer IoT, where moving personal data (audio, video, health metrics) to the cloud for AI processing often runs afoul of GDPR or CCPA regulations.
Conclusion: A Path to Success
Focus on measurable industrial outcomes (MTBF, Energy efficiency, operational excellence) and secure stakeholder buy-in early. Key success factors include:
• Getting to the right AI model for IoT is not trivial. Focus on “Right Sized AI for IoT"
• Move AI processing to the sensor level if possible or at a close edge processor. Only use cloud for higher level AI processing functions if needed.
• Optimized models reduce latency, lower cloud costs, and preserve battery life by processing insights locally
• Concentrate designs on Problem-First Logic, not technology-driven expansion
Bottom Line: AI IoT solutions can provide a valuable addition to running many operations. But many fail due to overreach and/or a lack of understanding of what the problem really is. Companies must lead with a Problem-First, ROI approach and not a Tech-First approach, while also bringing the end user community into the project early for maximum possibility of success. Finally, concentrate on the ROI aspects of any AI driven IoT solution in order to provide a real return to the organization.
Jack Gold is founder and principal analyst at J.Gold Associates, LLC. With more than 45 years of experience in the computer and electronics industries, and as an industry analyst for more than 25 years, he covers the many aspects of business and consumer computing and emerging technologies. Follow him on Twitter @jckgld or LinkedIn at https://www.linkedin.com/in/jckgld.