The 90% Rule: Why Real AI Agent Work Begins After Launch

Salesforce's experience with 20,000 Agentforce deployments reveals a critical inversion: AI agent development, unlike traditional software, shifts the real work to production rather than pre-launch planning. Success depends on avoiding three key anti-patterns—mixing deterministic logic into LLM reasoning, relying on prompts instead of code-enforced rules, and overwhelming models with excessive context—insights that apply regardless of scale.