The AI Runtime Field Lab

Field Briefs

B2B SaaS Starter Agentic Workflow Open Draft, pending editorial review

Triage First: A Support Ticket Routing Agent

Build an agent that classifies inbound support tickets, drafts a suggested reply, escalates uncertain cases, and explains why it routed each one.

Why this matters

Support teams drown in repetitive tickets while urgent ones wait in the same queue. An agent that triages and drafts replies clears the easy volume, but the failure that matters is a confident wrong route on a ticket that needed a human. Calibrated escalation, knowing when to defer, is the real target.

Persona

Support lead at a small B2B SaaS company

Current manual workflow

A support representative reads every inbound ticket, tags it by topic and urgency, writes a first reply or forwards it to the right person, and judges case by case whether the ticket is routine or needs escalation.

The AI workflow to build

The agent classifies each ticket by topic and urgency, drafts a reply grounded in help-center content, and assigns a confidence score. High-confidence routine tickets get a suggested reply. Low-confidence or high-urgency tickets are escalated to a human with a short note on why. Every decision carries a one-line routing explanation.

Inputs

  • inbound tickets
  • help-center articles
  • routing rules
  • past resolved tickets

Outputs

  • topic and urgency labels
  • suggested replies
  • escalation decisions with reasons
  • routing explanations

Definition of done

On a synthetic ticket set seeded with ambiguous and high-urgency cases, the agent labels topic and urgency, drafts grounded replies for routine tickets, escalates uncertain ones with a reason, and never auto-sends a reply to a flagged-urgent ticket. A routing explanation is present on every ticket.

Example input

A ticket: exports have been failing all morning with a 500 error, and this is blocking a month-end close.

Example output

Topic: export failure. Urgency: high. Decision: escalate to engineering on-call, do not auto-reply. Reason: a production outage signal plus business-critical timing exceeds the auto-reply threshold.

Data plan

synthetic data

Boundaries and non-goals

  • auto-sending replies without review
  • real customer data
  • live helpdesk integration

Evaluation ideas

  • classification accuracy
  • escalation calibration (precision and recall on the should-escalate set)
  • groundedness of drafted replies
  • explanation quality

Run Level target

R2 Visible Plain translation: handles real cases.

Scope envelope

Buildable by one solo builder in 20 to 30 focused hours, on public, synthetic, or sanitized data, with a demo path that requires no production access.

Suggested tools

Suggested options, never requirements; briefs are tool-agnostic.