Aderant's Ticket Triage Bot: A Small, Boring AI Deployment Worth Copying

Aderant, which sells business management software to law firms, says it built a scheduled AI workflow that reads unassigned support tickets, pulls context from four internal systems, recommends a team and a resolution path, and then takes routing actions on its own when confidence is high enough. In a guest post published on the AWS Machine Learning blog and co-written by Aderant's Angela Mapes and Adam Walker, the company reports that during its first 2.5 weeks in production the system reviewed 109 tickets with roughly 96% routing accuracy, at a total operating cost of under $30 per month. The numbers are small and self-reported, but the architecture is the more interesting part: this is a narrow, governed automation built on a cheap model, not an ambitious autonomous support agent.

What the system actually does

According to the post, the Intelligent Ticket Analyzer supports a 38-person team called SierraOps, which runs Aderant's Expert Sierra product across 268 client environments. Amazon EventBridge triggers a single AWS Lambda function hourly on weekdays. For each unassigned ticket in the CloudOps queue, the workflow runs five stages: pull unassigned tickets from Jira; enrich them with client metadata from Amazon Athena, knowledge articles from Confluence and Microsoft SharePoint, and comparable resolved tickets from Jira; send the assembled context to Amazon Nova Lite through the Bedrock Converse API for structured classification; act on the result by posting comments, reassigning misrouted tickets, notifying a Microsoft Teams channel and applying predefined actions; and finally publish metrics to CloudWatch while routing low-confidence classifications to human review.

Amazon DynamoDB handles cross-ticket pattern tracking, and credentials come from AWS Secrets Manager. The whole thing deploys from a single AWS SAM and CloudFormation template. Aderant also specifies a data boundary: the analyzer touches internal operational ticket data, operational metadata and internal knowledge sources, and does not access, process or store client matter data or client application business data. For a vendor serving law firms, that constraint is not incidental — it is probably what made the project approvable in the first place.

One detail worth flagging for anyone building something similar: the confidence threshold that separates autonomous action from human review is described as configurable, but the post does not disclose what that threshold is, how confidence is computed, or whether it comes from the model's own output or a separate scoring step. That is the single most important knob in the design, and it is left unspecified.

The baseline, and what the accuracy number means

Aderant says the SierraOps team handles an average of 34–40 tickets per week, and that before automation each ticket required 15–25 minutes of manual investigation before resolution work could start. That baseline came from a four-week pre-deployment measurement. From it, the company estimates the analyzer recovers 8–14 engineering hours per week, or 32–56 hours per month.

The accuracy figure has a clear definition, which is more than many deployment write-ups offer: the analyzer's assignment was compared against the team that ultimately resolved the ticket, yielding four misroutes out of 109. That is a reasonable ground truth. It is also a small denominator. At n=109, a 96% rate carries a wide confidence interval, and 2.5 weeks of summer traffic may not represent seasonal peaks, major release windows or incident surges. Aderant says as much, describing the findings as an early production period rather than a long-term benchmark — a caveat that deserves more prominence than it usually gets in posts like this.

The time-savings figure is an estimate derived by multiplying the manual baseline by ticket volume, not a measured post-deployment time study. It assumes engineers no longer spend the 15–25 minutes, which in practice is rarely a clean subtraction: reviewers still read the analyzer's output, low-confidence tickets still get manual attention, and someone spends time on the weekly routing-correction review the post describes.

Why a small model, and what the cost figure hides

Aderant says it evaluated multiple foundation models through Bedrock using real ticket data and chose Nova Lite for three reasons: it extracted more specific resolution steps and correlated past issues better than the other models tested on identical inputs; it integrates natively with the Converse API, IAM and the AWS SDK, avoiding another vendor contract; and its cost profile made it viable to analyze routine tickets rather than only high-priority ones. Which competing models were evaluated, and how the comparison was scored, are not disclosed — a meaningful gap in a guest post published on the model provider's own blog.

The headline economics are striking: under $30 per month total, with Bedrock inference under $1 per month. Read that second number carefully. Inference is roughly 3% of the bill. The rest is Lambda, EventBridge, Athena, DynamoDB, Secrets Manager and CloudWatch. At this volume — roughly 150 tickets a month — the model is close to free, and the real cost is plumbing. That inverts the usual assumption that model spend dominates. It also means the cost figure says little about how the system scales; Athena queries and API calls to Jira, Confluence and Microsoft Graph grow with volume, and none of the per-unit costs are broken out.

What the numbers do support is the argument for using a cheap model on high-frequency, low-stakes decisions. Routing a ticket to the wrong queue is recoverable. That is exactly the risk profile where sub-frontier models earn their keep.

The rollout pattern is the transferable lesson

Aderant describes going from concept to production in five weeks through progressive validation: a rules-based Jira automation first, which the company says could not incorporate AWS context, ticket history or model reasoning; then validation using an existing Amazon Quick agent to confirm operational data could support useful recommendations; then a production build with Lambda and Nova Lite, with Kiro used for integration scaffolding, architecture validation and test generation; then roughly two weeks of monitoring-only operation where classifications were compared against manual decisions and prompts were refined; then launch on June 30, 2026.

That shadow-mode period is the part most teams skip. It produces the ground truth needed to set a confidence threshold responsibly, and it converts a leap of faith into a measurable decision.

The post also describes a knowledge loop: when recurring issues appear across clients on the same software version and hit a configured threshold, the workflow records the pattern, updates a Confluence tracking page, watches resolved tickets for confirmed fixes and links future matching tickets to the documented resolution. Aderant does not disclose how pattern matching works or what the threshold is, and this capability is presented without any usage metrics — no count of patterns detected or documented. Treat it as a stated design goal rather than a proven result.

Aderant lists three next steps: deeper client-history analysis across related tickets, evaluating Bedrock Knowledge Bases for semantic matching when ticket and documentation wording diverge, and keeping the CloudOps Helper bot for on-demand investigation alongside scheduled triage. The second item is a quiet admission that keyword-style knowledge retrieval is currently a weak link.

What to watch

• Whether the ~96% routing accuracy holds across a larger sample and through release cycles, incident spikes and staffing changes — 109 tickets over 2.5 weeks is a pilot, not a benchmark.
• What share of tickets fall below the confidence threshold and go to human review. Aderant did not disclose it, and it determines how much work is actually removed versus reshaped.
• Whether a measured post-deployment time study confirms the estimated 8–14 hours per week, which was extrapolated from a pre-automation baseline rather than observed.
• How the under-$30 monthly cost behaves at higher volume, given that inference is under $1 and the bulk of spend sits in orchestration and data services.
• Whether the knowledge loop produces documented, reusable resolutions at a measurable rate, or remains an architectural intent without published metrics.