Repeat questions consume most of the headcount
When does it ship, how do I return it, which size should I pick — these dominate the queue. None of them is hard, but each one still has to be read and answered by a person.
Multilingual customer service: from slow, inconsistent replies to one line with a standard
The real cost in support is not writing a reply, it is **consistency**: when the same refund question gets three different answers, the refund policy stops existing. Meanwhile the handful of tickets that should have been escalated sit under hundreds of repeats.
01
Where it used to get stuck
Languages multiply the problem
The same standard replies are maintained separately per language, so one policy change means several edits. Miss one language and that market is quietly running on the old policy.
What should be escalated gets buried
Complaints, platform interventions and anything involving compensation sit in the same list as routine questions. Finding them means scrolling, and the busier it gets the more get missed.
02
How it runs once handed over
Classify first, then decide how to handle it
Intent is sorted into pre-sale, shipping, after-sales, refund and review, and each category takes a different path. Not every message deserves the same treatment.
One source of policy, rendered per language
Replies and policy are maintained in one place and produced in each language. Change the policy once and every market follows — no quietly stale market.
Drafted, not sent
Replies are drafted and wait for approval. High-volume low-risk categories can be loosened, but that is a setting you choose, not a decision it makes.
Risky items escalate immediately
Anything involving compensation, platform intervention or an angry customer is pulled out and pushed to a person instead of waiting in the ordinary queue.
03
Where it stops for a person
- Sending anything outward waits for a person, unless you explicitly loosen a category
- Anything involving money — refunds, compensation — is never executed automatically
- When intent is unclear it escalates instead of guessing a category and replying
How it is accepted
- Replay a batch of historical conversations and compare the classification against human judgement
- Change one refund policy and confirm every language's output changed with it
- Construct a few risky conversations and confirm they are pulled out rather than queued
04
Where the definitions come from
06
Questions
Which languages are supported?
The common cross-border market languages are in scope. The number of languages is not the point — **keeping policy in one place** is, because otherwise every added language is another place for the policy to drift.
Could it approve a refund on its own?
No. Actions involving money are never executed automatically; that is a hard constraint, not a setting. What it can do is assemble the reasoning and evidence and wait for you.
Can it work with our current ticketing system?
Through its API where one exists, and through a real browser carrying your signed-in session where it does not. Both go through the same approval gate.