Beyond the hype: AI, not for AI’s sake

Business leaders and governments are engaged in a fierce competition for AI dominance, driven by cautionary tales about the potential dangers of superintelligence. However, enterprises must maintain a grounded perspective and avoid the temptation to embark on a race to deploy increasingly complex models simply for the sake of it.

In the realm of enterprise AI, there is no reward for launching the most powerful model, generating the highest number of agents, or forcing AI into every workflow. True success is achieved by solving real business problems and delivering outcomes that justify the costs and complexities involved. While the idea of a singular, all-powerful large language model (LLM) may be enticing, today's business landscape necessitates a more measured approach: the right tool for each problem, often a simpler one.

Generative AI is not synonymous with LLMs. Businesses have been utilizing machine learning for tasks like recommendations, forecasting, fraud detection, and customer lifetime value estimation long before generative AI took the spotlight. These technologies have undergone years of development, including deployment, monitoring, and enabling rapid, cost-effective predictions.

An LLM may be capable of addressing some of the same issues, but this does not automatically make it the optimal choice. More advanced models can substantially increase operational costs and introduce unnecessary variability for straightforward prediction tasks. In certain cases, AI may not be required at all. Processes governed by known rules, such as data transfer between systems at scheduled intervals or triggering actions based on predefined conditions, can be efficiently handled through conventional automation techniques, including pipelines, schedulers, and rules.

If these established methods prove to be reliable, swift, and cost-effective, incorporating an AI agent does not enhance the process's intelligence; it could merely introduce additional complexity. Generative AI finds its purpose when creativity or interpretation is necessary, such as understanding language or images, or managing processes where the subsequent steps vary based on context.

Agents can further augment this by leveraging the understanding to interact with systems and execute actions, unlocking processes that were previously challenging to automate. The future of AI is expected to involve a harmonious blend of conventional automation, traditional machine learning, generative AI, and agents, rather than one technology supplanting all others.

The AI infatuation is waning. As enterprises transition from experimentation to production, the aforementioned distinctions become increasingly significant. Deploying a model or agent in a production setting entails running costs, security considerations, monitoring requirements, and the ongoing effort to maintain increasingly autonomous systems.

An analogy comparing the process to nurturing a tree, where branches growing in the wrong direction must be pruned, aptly illustrates this point. Identifying and addressing all such branches poses a significant challenge. Developing an agent can be similar. While obvious behaviors can be controlled relatively easily, real-world use often reveals edge cases that were not apparent during development.

The greater the autonomy afforded to a system in interpreting information and making decisions, the more crucial evaluation, monitoring, and guardrails become. Control must extend beyond just model behavior to encompass data access, authorized actions, retained information, and the point at which human intervention is necessary.

Cost considerations also come into play. The most sophisticated models can be prohibitively expensive at scale, so enhanced capabilities must yield sufficient additional value to justify the investment. As businesses shift from experimentation to scrutinizing return on investment (ROI), they are less concerned with whether something can incorporate AI and more focused on whether AI leads to demonstrable improvements in commercial or operational outcomes sufficient to warrant the cost.

Data should guide decision-making. Avoiding over-engineering begins with a clear understanding of the business need and the available data, rather than selecting technology first. Data engineers and analysts play a crucial role in comprehending the origins and reliability of information, as well as its predictive potential. While AI and machine learning engineers can apply mathematical models to data, the sophistication of these models cannot compensate for data that fails to support the task at hand.

The nature of the data and the desired outcome often point to the most suitable approach. Predictions derived from structured numerical data can be effectively addressed using traditional machine learning techniques. On the other hand, language, images, and other unstructured information can benefit from generative AI, while processes necessitating human-level understanding to trigger actions across multiple systems may warrant the deployment of agents.

In some instances, traditional software solutions remain the optimal answer. The misguided notion that "we need an agent" should be replaced with a more informed approach. The right individuals hold equal importance to the technology being employed. As AI becomes increasingly capable, there is a prevailing assumption that businesses will inevitably require fewer engineers.

While AI does enhance productivity, allowing experienced engineers to complete tasks more swiftly and junior engineers to become productive more quickly, businesses still necessitate individuals who possess a deep understanding of the underlying systems. Drawing an analogy with smartphones, it is evident that most users effectively utilize their devices without comprehending the intricacies of the hardware, operating system, and networks involved.

This does not imply that the foundational engineering skills have vanished. Just as there are still engineers who understand these systems, AI is poised to develop in a similar fashion. Many individuals will become proficient users, while businesses will continue to require engineers with the necessary technical expertise to build, control, and maintain the systems that underpin AI capabilities. This underscores the importance of continuing to invest in junior talent.