Customer support workflows

Distinguish answers from actions

Prepare an approved knowledge source, a list of permitted business actions and the conditions under which a person must take over. Answering a policy question and changing an order require different checks.

The desired output is either an answer grounded in current information, a clarification question, or an explicitly confirmed action request. Your business system remains authoritative for order state, prices and customer permissions.

Build the workflow

1. Identify the customer and the information they are authorized to access in your application.
2. Retrieve relevant policy passages or account data using that context.
3. Let the model draft an answer or propose a structured tool call for an allowed action.
4. Validate arguments, permissions and business conditions in application code. Obtain confirmation when the action requires it.
5. Execute the action through your business system, then report the actual result. Escalate unresolved cases to a person.

Test failure cases

Test missing orders, stale policies, ambiguous requests and attempts to act on someone else's account. Check whether the assistant asks for clarification and whether duplicate submissions can cause repeated business actions. A model proposing valid-looking JSON does not establish that an action is permitted or that it succeeded.

For sourced answers, the knowledge assistant workflow explains document handling. Use the selected model's API documentation to check tool-call support and limits.

Implementation boundary

SeedRouter provides the model calls. Retrieval, user identity, tool execution, audit records and human support remain in your application or agent environment. Choose an integration according to where those responsibilities already live.

Evaluate answer correctness and completed actions separately. Use current model prices and actual conversation records to measure costs; conversation length and repeated retrieval can change the total.