When should a workflow actually use an AI agent?

OpenAI's practical guide to building agents argues that agents — which it defines as "systems that independently accomplish tasks on your behalf" by using an LLM "to manage workflow execution and make decisions" — should be reserved for workflows that have "previously resisted automation," not deployed by default. The guide gives three criteria for fit: "complex decision-making" involving "nuanced judgment, exceptions, or context-sensitive decisions" (its example is refund approval); "difficult-to-maintain rules," meaning systems that "have become unwieldy due to extensive and intricate rulesets, making updates costly or error-prone" (its example is vendor security review); and "heavy reliance on unstructured data," meaning work that requires "interpreting natural language, extracting meaning from documents, or interacting with users conversationally" (its example is processing a home insurance claim). OpenAI states the gate plainly: "Before committing to building an agent, validate that your use case can meet these criteria clearly. Otherwise, a deterministic solution may suffice."

The mechanism OpenAI gives for why agents suit these cases is a contrast in reasoning style, not just capability: "a traditional rules engine works like a checklist, flagging transactions based on preset criteria," while "an LLM agent functions more like a seasoned investigator, evaluating context, considering subtle patterns, and identifying suspicious activity even when clear-cut rules aren't violated." Having picked a workflow, OpenAI's guide describes governing it through three components — model, tools, and "instructions," which it defines as "explicit guidelines and guardrails defining how the agent behaves" — and the guide's conclusion stresses that "guardrails are critical at every stage, from input filtering and tool use to human-in-the-loop intervention."

OpenAI's three criteria sit alongside When should you actually deploy an AI agent instead of simpler automation? as a second vendor framework answering the same underlying question — when should a workflow get an agent instead of simpler automation — but in different vocabulary: McKinsey sorts workflows by variance and standardization, OpenAI by the nature of the judgment, the brittleness of existing rules, and the structuredness of the data. Both converge on the same caution against defaulting to agents, which is also consistent with Why do production AI agents stay deliberately simple?: practitioners observed in production already constrain agents rather than deploying them broadly, matching both vendors' advice to gate agent use to workflows that need it.

The excerpt gives no measured comparison of agent versus rules-engine performance on any of its three example workflows (refund approval, vendor security review, insurance claims) — the "seasoned investigator" contrast is OpenAI's own illustrative framing from "numerous customer deployments," not a reported benchmark, and no deployment counts, accuracy figures, or costs are given for the claim. The guide is also OpenAI's own commercial guidance for teams building on its models, not an independent evaluation, so the three criteria should be read as OpenAI's recommended decision heuristic rather than a validated boundary between where agents help and where they do not.