Building With AI: An Honest Field Note

We set out to build a healthcare navigation product using AI. Not as an experiment, but as a real tool for real people. This is what actually happened—the good, the hard, and the honest limitations.

The Problem We Set Out to Solve

Healthcare navigation is broken. Patients face fragmented information, confusing insurance terminology, and overwhelming choices. We wanted to build something that could guide people through the maze—explaining terms, clarifying options, and providing clear next steps.

The vision was simple: a conversational interface that could understand a person's situation (symptoms, insurance, location) and provide personalized guidance on what to do next and why.

How AI Genuinely Accelerated the Build

Research phase compression was the first win. Instead of spending weeks reading medical guidelines and insurance documents, we trained a specialized agent to digest the information and extract actionable patterns. What would have taken a human researcher months took days.

Prototyping speed was dramatic. We could describe an interface concept and have working code in minutes. The AI didn't just write boilerplate—it understood the healthcare context and built appropriate validation, error handling, and user flows.

Content generation at scale was transformative. The product needed explanations of hundreds of medical terms, insurance concepts, and procedural steps. AI generated clear, consistent explanations that we could then refine and validate.

Testing automation saved countless hours. We built AI-powered test agents that could simulate user journeys, identify edge cases, and even suggest improvements to the flow.

The Honest Limitations

Accuracy is non-negotiable, and AI isn't there yet. Every medical recommendation had to be validated by human experts. The AI could suggest plausible guidance, but we couldn't trust it with literal life-and-death decisions.

Context understanding has hard limits. The AI could understand individual pieces of information but struggled with complex, multi-factor decision trees involving insurance coverage, symptom severity, geographic availability, and personal preferences simultaneously.

Regulatory compliance is a minefield. Healthcare products face strict regulations (HIPAA, FDA, etc.). AI couldn't navigate these automatically—every decision needed human legal review.

The "uncanny valley" of empathy. The AI could provide technically correct information, but it often missed the emotional context of healthcare decisions. We had to carefully design the tone and add human oversight for sensitive topics.

What We'd Do Differently

Start with validation, not generation. We spent too much time generating content and features before validating what users actually needed. Next time, we'd use AI to build rapid prototypes for user testing first.

Human-in-the-loop from day one. We tried to automate too much too soon. The sweet spot is AI handling the 80% of routine work with humans overseeing the critical 20%.

Better error boundary design. AI makes different kinds of mistakes than humans. We needed more robust systems to catch and correct AI-specific failure modes.

Embrace the hybrid model sooner. The most effective approach turned out to be AI handling information gathering and initial recommendations, with human experts providing final validation and emotional intelligence.

The Reality Check

AI didn't build the product for us. It accelerated specific parts of the process dramatically while requiring careful oversight in others. The overall timeline was still measured in months, not days—but those months were more productive than they would have been otherwise.

The product that emerged was better for the AI assistance but fundamentally human-designed and human-validated. The AI handled the heavy lifting of information processing and content creation, freeing human experts to focus on validation, empathy, and strategic decisions.

This experience changed how I think about AI in product development. It's not a magic bullet that eliminates the need for expertise. It's a force multiplier that lets expertise work faster and at greater scale—but only when paired with human judgment at critical points.

The healthcare product is still in development, evolving based on real user feedback. The AI continues to accelerate the work, but humans remain firmly in control of the decisions that matter. That's the balance that actually works.