Use a plain script when inputs are predictable and the logic fits in a flowchart; reserve an AI agent for tasks that need judgment on unstructured text, images, or messy data. The Microsoft news from this week sharpens that rule: if an agent gets the keys to your CRM or warehouse, you need separate monitoring around it, which is overhead a simple script does not demand.
Microsoft rolled out Guardian Agents and an agent detection feature inside Defender that can spot up to 60 types of agents employees bring in, including Claude Code and Codex. In a statement reported by TechCrunch, security chief Hayete Gallot said frontier models cannot be trusted - not because they are malicious, but because they are unpredictable and were trained to finish tasks, not follow company norms. CEO Satya Nadella's line is cleaner: separate the supply of intelligence from the authority over it.
What this means for a small business
For most small-business processes - moving a lead from a form to a CRM, generating an invoice, posting a stock movement - a scripted n8n flow or a few hundred lines of code is safer and cheaper than an agent. Bring in an LLM only at the step that actually needs reasoning: classifying an email, extracting fields from a scanned receipt, drafting a reply. Where I do use agents on sensitive data, like a legal document audit pipeline, the agent reads but a deterministic layer decides what gets written back and a human signs off on anything irreversible.
Why it matters: The cheapest security control is scope. Every action you let a script handle instead of an agent is one less thing you need Guardian Agents - or a human - to watch.
Source: TechCrunch