ShipStation reports its Deckhand agent resolved 97.5% of help desk channel queries

Both figures are self-reported. The post runs on AWS's blog and was co-written with five ShipStation Global staff ^[16]. The text does not define "resolved", give a query count, or say what share of answers drew a rating.

Deckhand hands off what it cannot answer. Chris Collins, senior director of IT at ShipStation Global, said that in those cases it "converts the conversation into a fully contextualized ticket with one click" ^[15]. Whether that handoff counts as resolved changes the rate. So does whether the messages the Slack service chooses not to answer count as queries at all ^[7].

Volume sets how fragile the figure is. The channel carried more than 1,500 threads in the 12 months before launch ^[4], which is more than 125 a month ^[18]. If post-launch volume were similar and the unit were threads, one thread would move the rate by 0.8 percentage points ^[20], and 2.5 percent unresolved would be about three threads a month ^[19]. Both assumptions are ours.

The 96 percent has its own limit. A satisfaction rate over rated answers describes only the employees who rated. The post says feedback is stored in DynamoDB ^[12].

For the numbers to transfer, a help desk would need a similar workload. That means repeated questions, a fifth or more of them arriving outside local business hours ^[5], and the same definition of resolved. In our view, a channel dominated by one-off questions would score lower on the same design.

The engineering holds up where it can be checked. AgentCore runtime hosts the agent in isolated sessions, and the authors write that they run no servers or containers ^[13]. X-Ray traces request paths using OpenTelemetry and the services log to CloudWatch ^[12]. A tracing stack is dull and exactly right for a post-mortem on a wrong answer.

Collins described the goal in terms of tickets. "We believe that tickets are an indication of operational friction not productivity," he said ^[14].