Tier-1 ticket triage with HITL agents: what to expect

Tier-1 ticket triage is the workflow most ops leaders ask us about first, and for good reason. It is high volume, mostly structured, has clean reversibility, and the cost savings show up in the quarter. Here is what running it on a HITL agent actually looks like, week by week, with the numbers we tend to see.

The intent taxonomy is the project

Most teams underestimate this part. Before the agent does anything useful, you need a clean intent taxonomy: refunds, address changes, order status, subscription pause, password reset, return label, bug report, billing question, and so on. Twelve to twenty intents covers 80%+ of tier-1 volume on a typical SaaS or e-commerce queue.

You will not get this right on the first try. Plan for one taxonomy revision in week 2.

What the agent does on each ticket

For each incoming ticket, the agent:

1. Pulls the customer record, their last ten orders or interactions, and any churn-risk flags.
2. Classifies the intent against the taxonomy with a confidence score.
3. For routine intents above threshold, drafts and applies the resolution: sends the email, applies the refund, updates the address, pauses the subscription.
4. For ambiguous, high-dollar, or churn-flagged tickets, builds a context pack and routes to a supervisor in <5 seconds.

The supervisor sees a one-screen decision pack: customer history, the agent's draft, the model's confidence, the policy clauses that fired, and a one-click approve / edit / reject.

The numbers we tend to see

These are post-week-4 numbers from real deployments, with names removed:

• Auto-resolution rate: 70–76%.
• Supervisor agreement on auto-resolved: 94–97%.
• Median supervisor review time: 30–50 seconds.
• p99 escalation latency: <10 seconds.
• Accuracy gain per week from supervisor corrections: +1.0 to +1.5 points.

The accuracy gain is what compounds. Six months in, auto-resolution rates above 85% are not unusual on the most repetitive intents.

The supervisor side of the math

Tier-1 ticket triage is the workflow that most clearly proves "humans involved cheaply." Before the agent, a supervisor handled around 60–80 tickets a day. With the agent, the same supervisor handles 200–300 reviews a day at higher quality, because they only see the cases the agent flagged. Total team headcount usually goes down 30–40% within a quarter, and the remaining team handles 4–5x more total volume.

We wrote up the structured product side of this on the use case page. The page covers the integration pattern, the timeline, and what the supervisor's day looks like.

When it does not work

If your tickets are >40% free-form support questions that need product knowledge to answer, this is not a tier-1 problem, it is a documentation problem. Get the docs in shape first; the agent will pick up the speed gain afterwards.

If the team picks tier-1 because executive leadership wanted the savings, but no one on the ops team owns the rollout, the project will stall regardless of the technology. Pick a workflow with a named owner.