5 AI Advances That Defined 2024

Looking Back

The road ahead with AI is coming into focus. The hype around Generative AI is giving way to the hard work of determining how to scale this new type of AI for business value. Meanwhile, deep learning and other AI types continue to fuel a range of applications in every industry and domain. The future is bright, and to understand where we’re headed and the opportunities on the horizon, let’s take a look back at some of the most important AI advances and insights in 2024.

1. The Rise of Agentic AI - The early chatbots and co-pilots enabled by large language models(LLMs) have shown some value in augmenting work, such as data summarization and content creation. Yet, value-driving GenAI at scale has proven challenging in part because there are limits on what chatbots can accomplish independently. Agentic AI is the next step toward more powerful systems that take on a greater degree of autonomy in planning and executing steps to accomplish a given goal – all with limited human intervention. This shows the concept of AI is evolving. It is decreasingly a matter of creating a single model purpose fit for a single task. Instead, AI applications are becoming systems of agents communicating and working in the background, accomplishing tasks based on human-set goals, not human-given instructions. This avenue of research and development could help enterprises push through GenAI pilot purgatory, from which 70% of pilots never escape, according to Deloitte research.
2. The EU AI Act – AI regulations have been anticipated for years, and the European Union’s AI Act is the first comprehensive AI regulatory regime. It establishes a legal framework for AI development, deployment, and use, and it applies to stakeholders across the AI lifecycle, including developers and end users. As with the EU’s General Data Protection Regulation, the AI Act is extraterritorial, applying to any organization worldwide whose AI products or contributions affect EU citizens. At its core, the Act is intended to mitigate risk to users. The regulation sets forth compliance requirements according to AI risk. Some AI applications are banned outright; others are treated with requirements for risk mitigation and transparency guidelines. As the heady period of innovation gives way to more mature technologies, governing bodies elsewhere in the world may release their own rules. In this, gaps or differences may emerge across geographies, requiring continued regulatory monitoring and preparation as a part of enterprise AI programs.
3. The Complexity of Scaling AI – It is one thing to identify an AI use case that could create business value and another thing to build and deploy at scale. Achieving scale is complex, with stakeholders and dependencies across the organization. Disjointed, random acts of AI for the sake of novelty are unlikely to drive business outcomes. Instead, AI use cases at scale answer a business need, whether to increase productivity, decrease costs, enhance quality, or differentiate the business in the marketplace. At the same time, creating value needs to come in tandem with managing risks. This requires attention to governance, change management, process and workflow transformation, and a focus on provisioning and securing the right data and infrastructure. Taking an intentional focus on scale, enterprises can identify where to surge investments to access the greatest potential value of an AI deployment.
4. The Emergence of Domain-specific LLMs – Since their initial releases in 2022 and 2023, general purpose, consumer-facing LLMs have shown remarkable capabilities, from generating poetry to mimicking a colloquial style of speech to providing summaries on most topics and in a variety of languages. For businesses, these general purpose LLMs may overgeneralize complex topics while creating privacy and security risks, which can limit their usefulness in certain domains. Domain-specific LLMs are the emerging solution to these challenges. They are trained deeply on specific subject matter, and the benefits include potentially greater accuracy and reliability in outputs. There are also benefits for protecting intellectual property and sensitive information. Inputting protected data into a general purpose LLM may create risk of inadvertent data leakage, whereas training a domain-specific LLM on proprietary data in a secure environment can help avoid some of these risks. By hosting the LLM on a cloud instance or even on-premise infrastructure, the enterprise has more control over data and security. There are also cost advantages, as smaller domain-specific LLMs may be bring lower compute costs and thus more savings in the aggregate at scale.
5. The Entanglement of Cyber and AI – Data is a strategic asset, requiring cybersecurity treatment across collection, transfer, storage, and usage. As AI consumes and computes enterprise data, it can potentially create vulnerabilities that bad actors could exploit to exfiltrate or compromise IP, financial data, or other sensitive information. Knowledge and data management are coalescing, and enterprises are challenged to secure the entire AI lifecycle and everything that touches it, including the assets, data flows, and development environments. At the same time, AI itself is creating new cyber threats. Synthetic content can be used as part of cybersecurity attacks, such as by mimicking a person’s words, voice, or likeness in a social engineering exploit to trick humans into divulging information or taking inappropriate action. These so-called “deepfake" threats complicate the cybersecurity landscape. Employees across all business domains will need to be trained and incentivized to watch for these kinds of exploits, and enterprises may need to develop new protocols or processes to validate communications.

Today, the only constant in the AI field is change. GenAI models are becoming more sophisticated. The union of models and enabling infrastructure is allowing enterprises to discover new ways of working and bring new products and services to market. Where we go next is being defined today, and it’s clear the AI era is well on its way.