An overseas customer reports a machine fault and asks for a replacement-part quotation in the same email. Sending the message to either sales or support loses part of the request. A classifier that always returns one neat label can make an inbox look organized while creating extra work for the people receiving it.
AWS published a decision-model deployment guide on October 12 that describes returning probabilities over declared options rather than unrestricted prose. Structured choices are convenient for routing, but the guide cautions against treating model probabilities as operational confidence without validation. For sales operations, the practical design question comes first: which messages should be routed automatically, and when should the system decline to decide?
Three reasons not to make a single choice
Missing information, an out-of-scope message and mixed intent are different conditions. They need different responses. Treating them all as an undifferentiated unknown queue merely moves classification work from the model to an operator.
A short request for information may belong to the product business but lack the product family. One targeted clarification can help. A clearly unrelated service pitch does not need a procurement questionnaire. A message that includes an existing order reference and a new quotation request may need an order association plus a separate sales task.
These are illustrative message types, not claims about a particular company's customers. Their purpose is to expose a basic limitation of a closed option list. If the available labels cover only product categories, a message outside those categories can still receive a highest-scoring product label. The model has compared the choices it was given; it has not established that the list covers reality.
Define the abstention paths before measuring automation coverage. Otherwise, adding an unknown option late in development can appear to reduce performance simply because the previous system hid uncertainty inside confident-looking assignments. A lower proportion of automatic labels may be the more useful result if it removes costly misroutes.
A routing threshold should reflect the cost of being wrong
A general brochure request sent to the wrong sales desk may need a routine transfer. An urgent operational problem placed in a marketing nurture queue can be more consequential. The categories should not automatically share one threshold just because the integration expects one number.
Build a labeled test set with the people who receive the work. Include brief messages, references to attachments, follow-up threads, local-language expressions and requests containing more than one action. Clean example inquiries are useful for a demonstration, but they are insufficient to represent an international inbox.
Review each category in both directions. Which messages assigned to it were wrong? Which relevant messages went somewhere else? A single overall accuracy score can hide a category that rarely occurs but matters greatly when it does. This precedes the reward-design question in lead qualification: first establish whether the allowed choices fit the messages. Preserve ambiguous examples where reviewers disagree; those may need a business rule or clarification, not a more forceful prompt.
When a clarification is appropriate, phrase it in the buyer's terms. Ask whether they need documentation for a particular model or help with an existing order. Do not expose an internal label such as queue B and expect the buyer to understand the organizational structure. Check the thread and attachments before asking for information already supplied.
Routing is not permission to send a response
Placing a message in a work queue and contacting the customer are separate actions. An initial deployment can propose the category, retain the supporting message excerpt and let an operator accept or correct it. Those corrections provide evidence for later decisions about where automatic routing is justified.
For Chinese exporters working with overseas distributors, role descriptions also need maintenance. A product may move to another regional team, a distributor may gain responsibility for support, or a new product family may be added. The classifier can continue producing valid old labels after the organization has stopped using them. Review the option descriptions whenever responsibilities change.
Messages concerning payment-detail changes, contractual commitments or urgent technical incidents need their own controlled human paths. Do not let a routing probability silently become approval for an external action. The model's task can remain modest: identify the likely owner, explain what is missing, and preserve every part of the request.
An abstained message should arrive with a usable reason: the missing information, the out-of-scope purpose or the separate actions it contains. Sending only an other label makes the next person repeat the classification from scratch.
Sources
AWS, Deploy decision models with llama.cpp on Amazon SageMaker AI, October 12, 2026.