AI's shift from promise to practice

Edge AI, which moves AI computation to end users' machines. Though not a new concept, we expect that the increasing use of small language models (SLMs) by companies will support the scalability of edge AI. That shift offers numerous advantages, including, notably, a reduced risk of mass outages associated with failures at major infrastructure or technology providers--known as centralization risk. The combination of SLMs with crypto technologies should further mitigate centralization risk, paving the way for more secure and decentralized computing and data storage (see AI & DeFi: Can Crypto Innovations Offset Artificial Intelligence Concentration Risks, Dec. 4, 2025). And because SLMs at the edge require less computational power, they will also offer energy efficiency advantages compared to data and resource-hungry LLMs (see Can AI become net positive for net-zero? Nov. 14, 2024). However, meaningful application of scalable edge AI will have to overcome the sizable challenge of ensuring that nascent technologies make their way from the lab to the field and are deployed in sufficient numbers to be impactful.

Causal AI, which can reason and make choices like humans do, uses causal inference to fundamentally understand why processes and decisions are made. This goes beyond the pattern recognition and correlation that underpins much of the current ability of AI (particularly discriminative AI and generative AI). We expect that capability will become increasingly important with increased AI adoption and the corresponding demand for data (both real and synthetic), as a means to identify true causality from spurious correlations. It should also enable AI models to deliver more reliable query responses, make more informed decisions, and potentially aid in the discovery of underlying causes--which could prove a powerful tool across sectors including healthcare and finance. Causal AI is generally regarded as one of the missing links to create artificial general intelligence (which matches or outperforms human cognitive abilities across any task). By nature, casual inferencing reduces complexity in decision-making, meaning it requires less data and power to deliver typically better performance. Active use cases are still relatively limited and tend to center on manufacturing, life sciences, decision sciences, and the study of human cognition and behavior.

Agentic AI, which describes the development of increasingly autonomous "AI agents" that make decisions and execute tasks based on a set of provided instructions (see Podcast: Artificial Intelligence Insights, Episode 4: Agentic AI, Oct. 30, 2024). This advancement opens the door for AI to shift from meeting a problem by providing relevant data to formulating and enacting solutions. As Microsoft put it, when it launched its generalist agentic system Magentic-One: the difference is "between generative AI recommending dinner options… (and an) agentic assistant that can autonomously place your order and arrange delivery." Agentic AI promises to reduce the complexity inherent in many systems, enabling better human decision-making and a focus on higher-value activities like creativity, leadership, and communication and clarification of decisions.

Cognitive AI, which refers to AI being able to mimic human cognitive functions such as reasoning, learning, and problem-solving--including by using cognitive computing structures known as neuro-symbolic architecture that simulate human patterns of learning, reasoning, and understanding. We expect cognitive AI will follow the widespread adoption of agentic AI to become the next step in the technology's evolution. Recent multimodal virtual chatbots, such as ChatGPT-4o and its latest model o1 (designed to think more deeply before responding), have some features that qualify as cognitive AI (i.e., the use of multimodal AI to provide reasoned responses and understand complex queries, and showcasing of adaptive learning in human interactions). Yet the models are not fully autonomous, meaning they require human inputs to achieve goals (unlike agentic AI), and do not yet integrate sensory information (from Internet-of-Things (IoT) devices, for example), telematics, or other sources of augmented real-time data needed for decision-making.