Why One Engineering Team Stopped Using AI for Automated Workflows

A platform engineering team supporting hundreds of developers initially built an AI-powered self-service system using Amazon Bedrock to generate Terraform code dynamically, but later removed it after recognizing the approach introduced unnecessary uncertainty. When the team later considered automating DNS record creation in Cloudflare, a similar AI-first proposal prompted deeper reflection on when AI is actually appropriate. The core argument is that LLMs excel at handling ambiguous, unstructured inputs — such as interpreting natural language requests — but are poorly suited for deterministic, step-by-step automation pipelines where predictable outcomes are required. Traditional workflow automation is reliable and testable precisely because it eliminates probability, whereas AI reasoning by definition introduces it. The team concluded that engineers should first ask whether a problem genuinely requires reasoning before defaulting to AI tooling, rather than choosing the technology before defining the architecture.