Why Most AI Coding Feels Like Driving a Ferrari to Buy Milk

Neel Sundaresan, the founding engineer of GitHub Copilot and now general manager of IBM Software Automation and AI, is developing IBM Bob, an AI‑powered coding assistant already used by about 80,000 IBM developers.

He notes that roughly 30% of a developer’s code consists of API calls. Traditional IDEs present long lists of possible functions, forcing developers to manually select the right one—a major friction point.

“If you use a class to call a function, you get a long list of callable functions and you have to pick one. That itself is a pain point.”

Sundaresan’s first system was not a code‑generator but an API‑call recommendation engine that ranks the most appropriate function at the right moment, treating the problem as a search‑ranking task rather than a generative one.

He emphasizes that user experience matters more than the underlying model. Even a better model can produce a worse product if the UI design is flawed.

Bob runs models locally on the client because many enterprise customers refuse to send data to the cloud. To make this possible, IBM invested heavily in engineering to ensure the models run efficiently on laptops.

Bob does not expose the underlying model; instead it automatically routes each request to the most suitable model—Anthropic Claude, open‑source Mistral, IBM Granite, or proprietary fine‑tuned models—based on task requirements and cost considerations.

“Even our customers don’t want to send data to our cloud. We actually run the model on the laptop, and we did a lot of engineering to make sure it works.”

Within IBM, Sundaresan conducts A/B tests across model variants, monitoring usage patterns to identify tasks where a cheaper model performs as well as a more expensive one, thereby controlling operational costs.

He chose IBM over other tech giants because IBM’s large internal developer base (about 20,000 engineers) and diverse workload—from Python and Rust to PL/I, COBOL, and JCL—provide a “zero‑customer” environment for rapid iteration and feedback.

Sundaresan warns that many developers misuse AI coding tools, treating them like a Ferrari for a simple milk run—overkill that incurs unnecessary token costs (e.g., a million tokens costing $40).

“It’s like driving a Ferrari to buy milk—completely unnecessary.”

Regarding AI agents, he argues that agent‑based development is not new; what is new is the probabilistic, conversational interface that introduces both new capabilities and new risks.

“We can either do nothing out of fear, or move forward bravely and methodically.”

He stresses discipline in AI projects: merely signing contracts with frontier model providers is insufficient; teams must enforce rigorous integration practices to avoid the high failure rate he estimates at 91%.