# Where Does AI Belong in a Workflow? Suggested ticket labels and automatic refunds have different failure costs\. Start the pilot where a reviewer can still correct the decision\. ![A reviewer examining an ambiguous card beside an automated sorting device\.]() The action after a model response determines much of its risk\. Two support teams want to use the same language model\. One wants suggested ticket labels\. The other wants automatic refunds\. The first decision can usually be reversed in seconds; the second moves money and may be difficult to unwind\. The model is only one part of that difference\. ## Define the decision before the prompt Write down the input, the exact output and who acts on it\. For a ticket classifier, the output might be one of six queue names plus “needs review\.” Specify what a bad routing decision costs\. A delayed routine request is different from a security incident sent to an unattended queue\. Compare against a simple baseline\. If a form field or a few rules already route most tickets correctly, measure whether the model improves the ambiguous cases enough to justify its latency and maintenance\. A fluent explanation is not evidence of better routing\. ## Pilot where errors remain visible Run the model on historical or appropriately authorized examples without letting it change records\. Keep examples from quieter queues and uncommon languages, not just the most frequent label\. Have reviewers record the intended action independently before seeing the model's answer\. Look at a disagreement table by class\. Ninety correct routine labels can hide ten missed security tickets\. Define which errors require abstention, what evidence the reviewer sees and where the original input remains available\. ## Set an exit condition A limited pilot needs a stop button, an owner and a way to continue manually if the model becomes unavailable\. Log the model version and decision without collecting unnecessary customer data\. Expand automation only when measured errors, reversibility and the handling of exceptions support the next step\. Suggesting a label and executing a refund deserve different permission boundaries\. ## Further reading [NIST: AI Risk Management Framework]() Related guide: [Compare Language Models on the Work You Actually Do]()\.