The conversation about AI in operations tends to mix two different things: classic automation — deterministic flows, n8n or Make style — and agents — systems that reason about a goal and decide steps. Confusing them is expensive in both directions.
When classic automation
When the process is stable and the rules can be written down: if A happens, do B. Invoices being entered, data being synced, alerts being fired. Here the deterministic flow always wins: cheaper, faster, auditable line by line, and it never improvises.
When agents
When the input varies and interpretation is required: a complaint written seven different ways, a document that needs reading and classifying, a decision that depends on context. There the agent brings what the flow cannot: understanding. In exchange it demands what the flow doesn't: designed human oversight, clear boundaries and quality metrics.
The architecture that works
The best systems we build don't choose: they combine. The agent interprets and decides at the ambiguous points; the deterministic flow executes everything else. AI at the hinges, engineering in the skeleton — and people watching the board.