The over-engineering trap
There's a pattern we see repeatedly as AI tooling matures: teams reach for agents to solve problems that a simple rule-based automation could handle in an afternoon. The appeal is understandable — AI agents feel like the future, they're technically impressive, and they generate enthusiasm in leadership presentations. But an AI agent that reasons, calls tools, and maintains state costs 10–50x more to build and operate than a deterministic automation. If you don't actually need the intelligence, you're paying a tax you can't recover. The question isn't 'can an AI agent do this?' Almost certainly yes. The question is 'does this problem actually require intelligence to solve?' That's a much more useful frame.
The three signals that a simple automation is sufficient
First: the inputs are structured and predictable. If you're moving data from a form submission into a CRM, matching invoice lines to purchase orders, or generating a weekly report from a database query — the data is structured, the logic is deterministic, and a workflow tool (n8n, Make, custom code) will handle it reliably at much lower cost. Second: every decision in the process has an explicit rule. If you can write the logic as 'if X then Y, else if A then B' with finite branches, you don't need an AI to reason — you need a decision tree. Third: the output format is fixed. If the result of the process is always the same type of object (a row in a database, an email from a template, a PDF from a report), deterministic code will be faster, cheaper, and more reliable than an LLM generating that output each time.
When an AI agent genuinely earns its cost
Agents justify their overhead when the input is unstructured and variable. Processing incoming emails from customers, extracting intent and entities from free-text support tickets, or reviewing contract documents for non-standard clauses — these are tasks where the variability of input makes rule-based approaches brittle and unmaintainable. Agents also justify their cost when the decision requires reasoning across context: not just 'does this invoice match a PO' but 'does this contract clause represent unusual risk given the counterparty's industry and the clause type?' That requires judgment, not pattern matching. Finally, multi-step research and synthesis tasks — produce a briefing document on this company, research competitive pricing, summarise these ten customer interviews — are well-suited to agents because the task structure varies and results need to be synthesised rather than extracted.
The decision matrix in practice
When a client asks whether a process needs an agent, we run through four questions: (1) Is the input structured or unstructured? Structured → automation first. (2) Are the decisions deterministic or judgment-based? Deterministic → rules, not AI. (3) Is the output a fixed template or generated text? Fixed → deterministic generation. (4) Does the process require multi-step reasoning or tool use? If no to all four, use a simple automation and save the agent budget for a problem that actually needs it.
Found this useful?
We write about automation, software strategy, and engineering once a month. No spam.