Aderant's hourly Nova Lite job routed 109 support tickets with about 96% accuracy

Aderant's hours estimate fits its own baseline. At 34 to 40 tickets a week and 15 to 25 minutes of investigation per ticket ^[10], manual triage took between 8.5 and about 16.7 engineering hours a week ^[1]. The claimed 8 to 14 recovered hours ^[9] sits inside that band. At both ends the recovered figure is below the manual one ^[1].

The accuracy figure rests on a smaller base. About 96% of 109 tickets is roughly 105 routed correctly and four routed wrong ^[2]. In a sample that size, each misroute moves the rate by almost a full percentage point ^[3]. Whether the number carries to another queue depends on three values the post does not state: the confidence threshold, how many teams a ticket can be sent to, and whether tickets a human checked before routing count toward the 96%.

The model choice is a result on Aderant's tickets. The team compared several foundation models on real ticket data ^[12]. Each model got the same ticket content, environment data, runbooks and resolution history. According to the post, Nova Lite extracted more specific resolution steps and correlated past issues better than the others ^[12]. For that ranking to hold on another team's queue, that team needs context of similar depth to feed in. With no runbooks or searchable ticket history, any model has less to correlate.

The build is small and well made. One Lambda function, triggered by Amazon EventBridge every weekday hour, runs all five stages ^[4]. Its Act stage posts analysis and acknowledgments, reassigns misrouted tickets, notifies the right Microsoft Teams channel and applies predefined actions for support-resolvable cases ^[6]. Metrics go to Amazon CloudWatch ^[5]. There is no always-on service to operate ^[4]. At under $30 a month in total operating cost ^[9], the weekly correction review is probably the most expensive part of the system.

The analyzer came after earlier AI work at Aderant, which used Amazon Quick for unified search and on-demand investigation ^[13]. The account is a guest post on the AWS machine learning blog, co-written by Aderant's Angela Mapes and Adam Walker ^[1]. One stated reason for choosing Nova Lite was that it runs inside Aderant's existing AWS environment, reducing the need for an additional external model provider or vendor contract ^[14].