Building AI-Enabled Engineering Teams: Lessons from Semiconductor Manufacturing

Leading an engineering team in a domain as demanding as semiconductor inspection and metrology means there is no room for slow feedback cycles. Every defect in the software directly impacts manufacturing yield. Over the past two years at Applied Materials, I have guided our team through a deliberate transformation—one that puts AI-assisted development at the centre of how we build, review, and ship software.

The Productivity Challenge

Legacy semiconductor software stacks are notoriously difficult to change quickly. Codebases are large, the domain knowledge required is deep, and the cost of a regression in a production tool can cascade into millions of dollars of yield loss. These constraints meant that moving fast traditionally came with very high risk.

The core question I asked myself was: can we get the speed benefits of AI assistance without sacrificing the rigour our domain demands?

Introducing Agentic Workflows

We started with GitHub Copilot for code completion, then layered Microsoft Copilot for documentation and test case drafting. The real step-change came when we integrated Anthropic Claude Code into our engineering workflow.

Claude Code’s ability to understand large, multi-file codebases and reason about architecture—not just complete the next line—was exactly what semiconductor software development needed. Instead of AI suggestions that required heavy correction, we got suggestions that understood invariants, thread safety constraints, and real-time performance budgets.

What Changed

• Code reviews went from 2-day cycles to same-day, with AI pre-screening catching ~60% of issues before human review
• Root-cause analysis of customer-reported defects accelerated significantly—AI could trace log patterns across thousands of lines in seconds
• Algorithm documentation, historically a bottleneck, became a by-product of development rather than a post-hoc effort
• Team throughput roughly doubled on feature delivery within six months of full adoption

The Human Side

The biggest risk I had to manage was not technical—it was cultural. Engineers who had spent years mastering complex C++ and image processing pipelines were understandably cautious about tools that could “do their job.”

The framing that worked: AI is a force multiplier for expertise, not a replacement. The engineers who adopted it most effectively were also the most experienced ones, because they knew exactly where to apply scepticism and where to trust the output.

What I Would Do Differently

Start with a specific, high-value problem rather than broad adoption. For us, using AI to accelerate root-cause analysis of customer issues was an immediate win that built credibility. Once the team saw concrete results in a high-stakes context, enthusiasm for broader adoption followed naturally.

Building AI-enabled teams in regulated or precision-critical industries requires more governance than consumer software contexts. We invested in prompt standards, output validation checklists, and clear escalation criteria. That infrastructure made the speed gains sustainable.