AfterShip Intelligence: the agent asks before it refunds

AfterShip launched AfterShip Intelligence, an AI agent for post-purchase operations, ITBrief reported on September 4. It watches shipments, flags exceptions, and reviews return authorizations. The feature worth copying is not the model — it is the rule that any financial or customer-facing decision still requires a human to approve it.

What actually happened

AfterShip Intelligence runs on tracking data the company has accumulated since 2012: more than 11 billion shipments, roughly 110 billion delivery checkpoints, and over 1,400 carriers, across a base of 20,000-plus brands. That corpus feeds delivery prediction, exception forecasting and return intelligence. AfterShip Agent is the piece that acts on it.

The launch-partner numbers are specific enough to be checkable. Dr. Squatch reports exception resolution time down more than 58%, "where is my order" tickets down 25%, positive sentiment on agent-initiated conversations up 42%, and 94% on-time estimated-delivery accuracy across 99.98% of orders. Naked Wardrobe reports return reviews roughly 40% faster — they previously took five to ten minutes each — with 15% to 20% of return authorizations resolved with no manual review at all.

Read that last figure carefully. Eighty percent of returns still go to a person. The agent is not clearing the queue; it is clearing the obvious end of the queue and handing over the rest with context attached.

Why the approval gate matters for your business

Most agent deployments we get called into fail in the same place. Somebody wires an LLM to a system that can move money — issue the refund, waive the restock fee, comp the shipping — and then discovers that "mostly right" is a very different standard for a support macro than for a ledger entry. A 3% error rate on draft text is a rounding error. A 3% error rate on refunds is a line item your bookkeeper will find in March.

The pattern AfterShip is shipping is the one to steal, and you do not need their product to build it:

Split the action from the authorization. The agent gathers evidence, writes the recommendation, and stages the transaction. A human clicks. That single boundary converts an autonomy problem into a throughput problem, and throughput problems are the good kind.

Make every action explainable before you make it automatic. If the agent cannot show which carrier scan and which policy line produced the recommendation, you cannot audit the 15% it handled alone — and you will not be able to defend it during a chargeback dispute.

Instrument the approval step. Track how often the human overrides the agent, by reason. That override rate is your real accuracy metric, and it tells you exactly which decision types are ready to automate next. Vendor benchmarks will not.

The right first automation in post-purchase is almost never the refund. It is the "where is my order" reply, drafted from live carrier data, sent without anyone touching it. Cheap to get wrong, constant, and it is the ticket volume actually drowning your inbox.

Sources: ITBrief UK, SalesTechStar.