A team is using Claude for two different tasks. Task A: Extracting structured data from invoices (vendor, date, line items, total). Task B: Determining whether a refund request is eligible based on a 4-step policy (check time window, check item condition, check customer history, check refund cap). They're debating whether to add chain-of-thought reasoning ("walk through your reasoning step by step before answering") to one or both tasks.
For which task is chain-of-thought most beneficial?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Chain-of-thought helps when there are reasoning steps to lay out. Multi-step policy application (the refund task) is the textbook fit: each step depends on the previous, and externalizing the reasoning makes errors visible. Structured extraction (the invoice task) doesn't have this shape — there's no chain to reason through, just fields to find. Adding "think step by step" to extraction adds tokens without adding reliability.
CoT adds latency and cost on every call. For tasks without reasoning steps, the cost buys nothing — and can occasionally hurt by encouraging the model to overthink simple transformations.
Field-by-field extraction isn't reasoning — it's mapping. Asking the model to "think through" each field wastes tokens and can introduce errors when the model talks itself out of correct extractions.
CoT is heavily used in production for exactly the cases where it helps (multi-step reasoning, policy application, complex decisions). "Overkill" is a non-criterion — match the technique to the task shape.