JBS Dev: AI Doesn't Need Perfect Data to Excel

JBS Dev and the Myth of Perfect Data

Joe Rose, president of JBS Dev, a company specializing in strategic technologies, is tackling a persistent misconception in the field of artificial intelligence. According to him, it is misguided to believe that data must be flawless before it can be effectively utilized by generative and agentic AI systems. Rose insists that this belief is unfounded.

A recent article published in AI Fieldbook highlights this issue. Technology vendors and consultants often advocate for the necessity of vast data lakes and conducting multi-year data transformation programs. This approach leaves many executives perplexed. However, the reality is more nuanced. Rose asserts that current tools are capable of processing low-quality data with surprising efficiency. He emphasizes the ability of language models to understand instructions even when they are only partially formulated.

The Importance of Human Safeguards

The use of these powerful tools, however, requires the establishment of appropriate safeguards. AI models, by nature, can produce unpredictable results, making human intervention essential to manage potential errors. In the realm of textual or categorical data, a certain resilience is already present. Rose explains that users must become accustomed to an iterative process, where AI does not operate autonomously but requires constant oversight.

A Concrete Example in the Medical Sector

To illustrate his points, Rose cites a concrete case in the medical sector. A client wanted to migrate to a new billing reconciliation system. The records were in a state of disarray, with some in PDF format and others as images. Sometimes, information was misattributed, such as the doctor's name appearing instead of the patient's. Thanks to generative AI, it was possible to extract clean data from simple instructions, using technologies like optical character recognition for images and text extraction for PDFs. More sophisticated approaches then allowed for the correct billing verification by comparing records with insurance contracts.

Towards Gradual Automation

Rose describes a process of gradual automation. The idea is to start with a low percentage of automation, for example, 20%, and gradually increase it to 40%, then 60%, and so on. The goal is to grow over time while maintaining human supervision to ensure the accuracy and efficiency of the systems.

The Future of AI: Costs and Portability

Looking ahead, Rose anticipates that discussions around AI models will focus on cost sustainability and portability. He predicts a paradigm shift, moving from spectacular improvements in model capabilities to a consideration of how to make costs more manageable. The idea is to reduce reliance on massive data centers and enable models to operate on more accessible devices like laptops or phones. The models have been trained on a dataset that essentially includes every page of the Internet and other sources, meaning there is not an abundance of new data to integrate for a significant breakthrough.

Encouragement for Technological Autonomy

At the AI & Big Data Expo, Rose plans to share an opinion that may surprise: he encourages companies to consider developing their own solutions rather than relying on SaaS vendors. According to him, with a cloud presence, most companies already have the necessary tools to begin implementing agentic workloads without needing new software or training. JBS Dev positions itself to support these companies in the next stages of their technological journey.