ServiceNow Ticket Auto-Filling: A Practical Agent

Ten fields. Five tools. One agent. The integration that fills tickets correctly more often than the humans did, with the gold-set used to prove it.

The 10 fields

The ticket has 10 standard fields. Title, description, urgency, impact, category, sub-category, affected service, assignee, environment, links to evidence; each field has a constrained set of valid values (urgency: 1-4; category: defined list). The agent’s output respects the constraints because structured output forces the model to pick from the enum, with no free-form leakage into structured fields.

• 10 standard fields. Title, description, urgency, impact, category, sub-category, service, assignee, environment, links.
• Constrained value sets. Urgency 1-4; category defined list; the agent picks from enums.
• Structured output enforces constraint. No free-form leakage into structured fields.
• Per-field validation. Each field validated; supports correct ticket shape.

The 5 tools the agent uses

The agent has five tools at its disposal. Pull alert payload (the trigger event); pull service registry (identify the affected service); pull on-call schedule (identify the assignee); pull recent metrics (for the description); pull recent deploys (for context). Each tool is bounded and read-only.

• Alert payload. The trigger event; the starting context.
• Service registry. Identify the affected service from the alert.
• On-call schedule. Identify the right assignee.
• Metrics and deploys. Description content plus context for the ticket.

The gold-set proves accuracy

A 100-ticket gold set proves field-by-field accuracy. 100 historical tickets hand-validated by the team; the agent fills each based on inputs available at original filing time; compare agent output to original ticket field-by-field with target 90%+ accuracy per field. Disagreements get reviewed because sometimes the agent is more accurate than the human (typo) and sometimes the agent missed context.

• 100 historical tickets. Hand-validated by the team; the gold set.
• Same-input filling. Agent uses inputs available at original filing time.
• 90% per-field target. Field-by-field accuracy; the bar.
• Disagreement review. Sometimes agent more accurate; sometimes missed context; both improve.

Human review before file

Human review starts as required and graduates per-category. Agent fills the ticket as a draft, human reviews, edits, submits; after 90 days of reliable performance on a category, auto-file is enabled for that category while other categories remain draft-only; auto-file with notification gives the team a 5-minute window to override.

• Draft-by-default. Human reviews, edits, submits; the discipline starts manual.
• Per-category graduation. 90 days reliable performance enables auto-file.
• Other categories draft-only. Graduation is per-category, not platform-wide.
• Auto-file with notification. 5-minute override window; safety net.

How this compounds

The compounding effect is what makes the investment pay back. Each ticket the agent fills gets reviewed (or auto-filed) and the reviews train the next iteration; common issues identified in review become prompt updates so the agent improves field by field; after a year the agent fills tickets more accurately than the average team member and the team uses the agent as their default.

• Reviews train next iteration. Each review feeds prompt update; the loop closes.
• Common issues become prompts. Field-by-field improvement; structural learning.
• Year-over-year accuracy. Eventually exceeds average team member.
• Per-team default. Team uses agent as their default ticket-filing tool.