Analyst Perspectives

It is now more than two years since the launch of ChatGPT introduced the world to generative AI (GenAI) and large language models (LLMs). GenAI-based assistants and co-pilots are now widely adopted, with individuals and enterprises adopting GenAI models to automate the generation of text, digital images, audio, video and code, amongst other things.

More than four-fifths (81%) of participants in ISG’s Market Lens study on GenAI are using chat-based services, such as ChatGPT, with 72% using GenAI for text generation and translation, 43% for code generation, and 42% for image generation. ISG’s recent Market Lens study on GenAI also indicated that the primary motivations for enterprise adoption of these GenAI models include improved efficiency, business growth and cost savings.

The rapid rate of adoption of GenAI models and assistants has been matched by the rapid pace of expansion of GenAI-related capabilities. While LLMs triggered the initial adoption of GenAI, attention quickly expanded to associated capabilities such as retrieval-augmented generation. This is being adopted by enterprises to make GenAI more suitable for addressing enterprise use-cases by grounding the output of GenAI models in enterprise data to improve trust.

The latest focus of attention in the AI space is agentic AI, with multiple software providers introducing agents that combine multiple models and assistants to take GenAI beyond the automation of discreet functions to address complete business tasks.

The rise of agentic AI is a natural outcome of enterprise requirements related to GenAI. While individual users can play around with various GenAI models and services to fulfill discreet requirements, enterprises have more complex processes and requirements that need to be served by a combination of models and applications.

The variety of GenAI models available presents enterprises with the upfront challenge of understanding the suitability of models to specific business tasks and processes and approving specific models for enterprise use. While many models can be used without modification for non-critical use-cases, in many cases enterprises will need to train, tune, prompt or ground the model with enterprise information to ensure that they can trust its output.

Grounding a model with enterprise data can only be successful if the enterprise has trust in the data being used to ground the model. As such, it is imperative for enterprises to ensure the validity, quality and reliability of enterprise data. As was explained in ISG’s State of Generative AI Market Report, AI-ready data is data that is clean, well-organized and compliant with regulatory standards. The need for good data management is by no means new. Enterprises have been concerned about data quality and reliability as it relates to business intelligence for many years. Despite that, many enterprises find themselves struggling with data that is fragmented, inconsistent and not easily accessible.

The importance of good data management is an even greater priority in relation to AI for three key reasons. Firstly, while BI uses data to support decisions that are taken by employees, enterprises adopting AI are relying on data to make and automate decisions in real-time.

Secondly, those decisions are increasingly critical—for example spotting fraudulent activity and delivering customer service—so the risks associated with failure are heightened. Thirdly, the use of AI is increasingly a board-level concern. If the CEO is driving increased adoption of AI, and has expectations about improving efficiency, innovation and growth, then AI initiatives had better deliver.

Increased enterprise focus on AI is therefore a forcing function for enterprises to take long-overdue steps to improve data management and data governance.  As my colleague David Menninger recently explained, enterprises also need to govern the security, privacy and explainability of the application, while also being conscious of the criticality of detecting and preventing bias as well as identifying and eliminating toxicity. I assert that through 2026, one-third of enterprises will realize that a lack of AI and ML governance has resulted in biased and ethically questionable decisions. 

Finally, once models have been selected, trained, grounded and governed, enterprises also need to be able to ensure that recommendations and automated decisions are conducted in the context of reasoning and driver-based planning models and applications, and that they can also identify and measure the value delivered by an AI project to explain and justify the investment.

Individual models deployed in isolation can be used to serve discreet, well-defined functions (like code conversion or drafting an email). Given the variety of requirements, however, it is clear that individual models will not be enough to serve more complex business use cases, such as responding to a customer service inquiry. In addition to training, tuning, prompting or grounding via RAG, the output of GenAI models also needs to be combined with other AI models to address predictive machine learning, reasoning, planning, budgeting and risk assessment. Similar to the way in which different combinations of atoms come together to form different molecules, we see software providers and enterprises alike bringing together different combinations of AI models to form agents to fulfil specific business tasks. I recommend that all enterprises explore the potential benefits of agentic AI. While they do so, however, they should also be cognizant of the advice of my colleague Robert Kugel and his assertion that enterprises need to invest in generative automation to address the even greater levels of complexity involved in collaborative business processes.

Regards,

Matt Aslett