An owner faces a crowded queue
In this hypothetical retailer, an owner must choose how an AI assistant handles 30 refund requests daily within 90 staff minutes. All timings are assumed, not reported results. Customers want quick answers. The owner assumes automating more decisions will free more staff time. Yet a refund promise commits the business to an action; polished wording cannot justify it.
The record behind the promise
For one request, the assistant sees order_status = delivered but no final_sale value; the sales system holds final_sale = true. A default treats the missing restriction as permission, producing “Your refund is approved.” The retailer's assumed policy excludes final-sale items. Sending that reply could require a correction or a refund from the retailer's cash. A reply preview is a shop window: it displays the promise, while eligibility depends on the record behind it.
Two routes through the same queue
Option A lets the assistant draft while staff verify eligibility and approve every reply. At two minutes per request, 30 × 2 = 60 staff minutes daily, leaving 30 minutes free.
Option B automates complete, policy-matching requests and sends incomplete records to staff. Assume daily monitoring takes 30 minutes and ten exceptions take four minutes each, including triage and handoff. That is 30 + (10 × 4) = 70 staff minutes, leaving 20. Both fit; choose A under these assumptions.
Check the decision before it travels
Before enabling automatic approvals, compare every request's final_sale value with the sales system. Missing or conflicting values require staff review; approval despite that gap blocks automatic release. Confirm that a test request with final_sale = true receives no refund approval. Time verification, exceptions and monitoring across the whole queue.
When the balance changes
Now suppose this same retailer fixes the data connection and only five daily requests need staff review. Option B becomes 30 + (5 × 4) = 50 minutes. Choose it if eligibility checks pass. With monitoring unchanged, automation beats A at seven or fewer exceptions daily; missing policy information still requires review.