AI Last Mile Problem: Bridging Model Capability and Cost Sustainability with Imperfect Data

Joe Rose, President at strategic technology provider JBS Dev, is on a mission to debunk one of the most persistent myths surrounding generative and agentic AI adoption in the enterprise.

“It's a common misconception that your data has to be perfect before you do any of these types of workloads.”

— Joe Rose, President, JBS Dev

As highlighted in a recent article in AI Fieldbook, vendors and consultants often suggest that organizations need massive data lakes and multi-year data transformation programs before they can begin leveraging AI — a narrative that leaves many executives overwhelmed. But the reality, according to Rose, tells a different story.

💡 "The tooling has never been better than it is now to deal with poor quality data," says Rose. "It's almost remarkable what an LLM can understand on a half-written prompt."

The inherent unpredictability of AI models does require careful handling — particularly when it comes to bad output. This is where the human-in-the-loop model becomes essential. Rose notes that for textual or categorical data, there is already a natural resilience built in, but organizations must move away from the "set it and forget it" mindset.

“People are used to — 'we build it, it works, we forget about it.' That's just not how these systems work.”

— Joe Rose, JBS Dev

🏥 Real-World Example: Medical Billing Reconciliation

To illustrate the power of AI with imperfect data, Rose points to a client in the medical sector that needed to migrate to a new billing reconciliation system. The records were a chaotic mix:

• Some records were in PDF format, others were image files
• Procedure details were sometimes filed under the doctor's name
• Doctor names appeared in patient name fields
• Data was inconsistent across the entire system

Despite this, generative AI was able to scope and clean the data using a simple prompt — leveraging OCR for images and text extraction for PDFs. More agentic approaches were then applied, such as comparing customer records against insurance contracts to verify billing accuracy.

📈 The Incremental Automation Model

“You start to layer different use cases on top of one another. We started at 20% automated, then 40%, then 60, 80% — and grow that over time.”

— Joe Rose, JBS Dev

Rose emphasizes that this is not about achieving perfection overnight. A phased, incremental approach — with human oversight at every stage — is both realistic and effective. The goal is progressive automation, not a one-time transformation.

🔗 The Future of AI: Cost, Portability & Sustainability

Looking ahead, Rose believes the next major shift in AI discourse will move away from model capability breakthroughs and focus instead on cost sustainability and portability.

“How do we make the cost more sustainable so we don't have to build data centres at the rate we're building data centres?”

— Joe Rose, JBS Dev

📱 The "last mile" challenge, as Rose describes it, is getting these models to run on a laptop or a smartphone rather than requiring large-scale data centre infrastructure. He also challenges the assumption that new training data will drive the next major AI breakthrough — arguing that models have already been trained on an enormous corpus of existing information.

💬 Stop Buying From SaaS Vendors — Build It Yourself

At AI & Big Data Expo, where JBS Dev is an active participant, Rose is preparing to share what he calls a more controversial take: organizations should stop purchasing from SaaS vendors when they have the capability to build solutions themselves.

“It's not as hard as it sounds. Almost everybody's got some kind of cloud presence — and that's where I would start. The cloud tooling, especially for the big three, has everything you need to start implementing agentic workloads tomorrow, without new software licenses and new training.”

— Joe Rose, JBS Dev

✅ The message is clear: organizations don't need perfect data, massive budgets, or lengthy vendor contracts to get started with AI. The tools are already available. The cloud is already there. The time to act is now.