The emergence of generative AI, particularly marked by the launch of ChatGPT-3.5 in November 2022, has transformed how corporations approach new technologies. As organizations scramble to harness this advancement, it is essential to recognize the unique challenges that come with building and maintaining a high-quality generative AI assistant, which can be substantially different from traditional tech initiatives. As the two-year anniversary approaches, evidence suggests a substantial risk of failure for many companies trying to build generative AI assistants. Such initiatives not only require a clear understanding of current technology but also demand a dynamic, adaptable approach to managing rapidly evolving AI tools, or risk making decisions that lead to redundant or inefficient work.
To illustrate these challenges, consider two hypothetical projects undertaken by ImaginAiry Airlines: the development of a customer-focused mobile app versus a generative AI assistant for customer service. The standard process for creating a mobile app generally follows a structured path—initial understanding of requirements, stakeholder approval, and execution. While such a method can yield positive results for non-AI software, the approach falls short in the context of generative AI. The rapidly changing landscape of generative AI means that decisions made today—regarding budget and tools—may well become obsolete as new technological advancements reshape the field in the next few years, necessitating a re-evaluation of earlier decisions and potentially a complete overhaul.
There are primarily three risk factors that could thwart an organization’s efforts in developing a generative AI assistant. First, selecting an unsuitable large language model (LLM) vendor poses a significant risk, as the performance landscape of LLMs is in constant flux. Predictions regarding which models will lead the industry in the near future are uncertain, and a selection made now may soon be outclassed, leading to potential high costs in re-engineering earlier efforts. Second, organizations face the challenge of deciding between open-source versus closed LLMs. While closed options typically offer more straightforward implementation and robust vendor support, they incur higher fees and less customization. Open-source models can be cost-effective and adaptable but often require in-depth engineering expertise, creating a difficult choice for firms with varying resource levels.
Beyond provider selection, technological breakthroughs introduce additional uncertainties. Currently, a common best practice in AI assistant development is retrieval augmented generation (RAG), which enables adaptive use of various LLMs. However, advancements in AI research could soon render RAG obsolete, as newer methodologies emerge that enhance the capabilities of generative AI assistants. Innovations like multiple AI models working collaboratively or neuro-symbolic AI could revolutionize the way these systems function, resulting in fundamental shifts in development strategies and necessitating significant changes to existing codebases, potentially nullifying previous investments.
Given these myriad uncertainties, organizations must adapt their strategies for building generative AI assistants. Traditional methods—like completing a structured business case followed by implementation—are inadequate. A lack of agility can hinder progress. Instead, firms should establish cross-functional teams comprising senior stakeholders who can make timely decisions in response to ongoing technology developments. This approach mirrors practices in agile environments and ensures that organizations remain attuned to the fast-paced evolutions within the generative AI landscape.
On the budgeting front, organizations should avoid perceiving their AI investments as fixed, one-time escalations. Flexibility should be embedded into financial planning to accommodate changing needs and unexpected technological developments. Firms should consider building a continuous operational budget that allows for iterative investments in generative AI development—enabling quick pivots if initial choices do not yield anticipated results. This funding model should support a dedicated team year-round, emphasizing not only the development of the assistant but also the modernization of existing infrastructure to optimize performance.
In conclusion, while the risks associated with generative AI initiatives seem daunting, they can effectively serve as catalysts for reformation and refinement within organizations. Accepting the fluidity of technology necessitates an agile and proactive approach to development and investment. As generative AI continues to mature, firms must commit to long-term strategic investments, fostering a culture of innovation and integration that positions them favorably in the competitive landscape. With the right operational processes and foresight, organizations can ultimately leverage generative AI to improve customer experiences and drive overall efficiency, establishing robust systems that adapt to the shifting dynamics of this transformative technology.