Eliminating Manual Email Triage with an AI-Driven Workflow

Time-Intensive Manual Triage

Each inbound email required a team member to read, interpret, classify, and manually create a Jira ticket — a process that took 5 to 10 minutes per email. With a continuous stream of scheduling requests, corrections, and information queries arriving daily, this created significant operational overhead and delayed response times.

Inconsistent Priority Assignment

Priority decisions were made manually based on individual judgment, leading to inconsistency across client tiers and SLA requirements. High-priority clients such as L1 strategic networks including NBC, CBS, FOX, and ESPN did not always receive the urgency they required, increasing the risk of SLA violations.

Duplicate Ticket Management

Without an automated duplicate detection mechanism, the same scheduling request or update was frequently logged as multiple separate tickets. This created confusion for the operations team, cluttered the Jira backlog, and made it difficult to track the true status of any given request.

Limited Audit Visibility

There was no structured audit trail linking inbound emails to the tickets they generated. This made it difficult to investigate issues, verify classifications, or demonstrate SLA compliance — creating operational blind spots across the triage workflow.

01.

Email Ingestion via Microsoft Graph API

The system connects to the scheduling support inbox using an OAuth2 client credential flow through the Microsoft Graph API, reading unread emails every 5 minutes and processing them oldest first. Pre-filters were applied before any AI processing to skip system-generated messages, prevent self-loop emails, and exclude irrelevant recipients — ensuring only actionable emails entered the pipeline.

02.

Content Extraction and Attachment Processing

Email content was extracted from HTML into clean plain text, and attachments across formats including Excel, CSV, PDF, Word, and TXT were downloaded, stored in Amazon S3, and processed for text content up to a 100,000 character limit. Schedule dates were extracted from attachments using AI in parallel, ensuring all relevant context was available for accurate classification.

03.

AI Classification Using Amazon Bedrock

Amazon Bedrock powered the core classification engine, with Claude Sonnet as the primary model and Claude Haiku as a fallback. Each email was analyzed and returned a structured output including the recommended action, a confidence score between 0 and 100, the identified client name, target date, description, and action points. A low temperature setting of 0.2 ensured deterministic, consistent outputs across all classifications.

04.

Deterministic Business Rule Overrides

AI classification was supplemented with hard business rules to enforce SLA compliance. Comscore emails were automatically forced to P1 priority with a critical label, regardless of AI output. Emails referencing past target dates were always classified as past data modifications, a rule the AI could not override. These hard overrides ensured that critical clients and time-sensitive requests were never misclassified.

05.

Priority Matrix and Client Tier System

A structured priority matrix was applied based on client tier and target date proximity. L1 strategic networks including NBC, CBS, FOX, ESPN, and BBC received the highest priority handling with automatic owner assignment. T2 major networks and standard clients were assigned priorities based on SLA windows, ensuring consistent and defensible priority decisions across over 6,100 affiliate station mappings.

06.

Duplicate Detection and Multi-Client Handling

A two-layer duplicate detection system using JQL fuzzy search and exact subject normalization identified existing tickets before creating new ones. Active tickets received update subtasks rather than duplicate entries, and priority could only escalate — never downgrade. Emails referencing multiple clients generated one parent ticket with independent subtasks per client, each with its own priority, assignee, and classification.

07.

Audit Logging and Observability

Every processed email generated a complete audit trail stored in Amazon S3, including the raw email, AI extraction output, Jira payload, execution log, and all attachments. Metadata was stored in Parquet format for analytics, providing full traceability from inbox to ticket for every classification decision made by the system.

100% Automated Triage

Every inbound scheduling support email was classified, prioritized, and actioned without manual intervention — eliminating 5 to 10 minutes of triage time per email and freeing the operations team to focus on higher-value work.

Deterministic Priority Assignment

The combination of AI classification and hard business rule overrides ensured that every ticket received a consistent, defensible priority based on client tier and SLA requirements — replacing subjective, inconsistent manual decisions.

Controlled Duplicate Management

The two-layer duplicate detection system significantly reduced redundant ticket creation, keeping the Jira backlog clean and ensuring that updates to existing requests were appended rather than duplicated.

Full SLA Compliance

Automated enforcement of client tier priorities, Comscore overrides, and past-date classification rules ensured that high-priority requests were never delayed or misprioritized — maintaining SLA compliance at scale.

Complete Audit Visibility

Every email and its corresponding classification, payload, and outcome were logged in Amazon S3, providing the operations team with a full, searchable audit trail for investigation, reporting, and compliance purposes.