← All postsBlog

Meet Priya - the customer support triage agent that runs on your laptop

· Avery NXR

Customer support has a structural problem that's gotten worse as AI has been added to the workflow. The cloud-LLM tools that handle triage work well — they classify, route, and draft responses faster than a human can — but every ticket flows through a third-party AI provider. For most support orgs, that's tolerable. For support orgs handling healthcare data, financial data, or anything covered by an enterprise contract with stricter handling requirements, it's not.

Priya is one of the 7 production-ready agent templates that ship with Avery NXR. She pulls new help-desk tickets every 15 minutes, classifies them, detects sentiment and urgency, auto-replies to the simple how-tos, and routes the rest to the right human with a drafted response ready to send. She runs entirely on your laptop. The ticket content never leaves.

What Priya actually does

Every 15 minutes (or on a webhook trigger, your choice), Priya:

  • Pulls new tickets from Zendesk, Intercom, or whichever help-desk you use
  • Classifies each by topic, intent, and product area
  • Detects sentiment (happy / neutral / frustrated / angry) and urgency (low / medium / high / critical)
  • For simple how-to questions with KB-matchable answers: drafts and posts the reply directly
  • For everything else: routes to the right queue or agent with a drafted starting response
  • Tags the ticket with structured metadata so analytics tools can use the classification
  • Optionally posts urgent items to Slack for immediate attention

A typical loop takes 5-15 seconds per ticket.

Why the auto-reply tier matters

The most underrated efficiency in support is not the agent who handles the hard tickets faster. It's the agent who never has to see the easy ones.

Across most support orgs, 20-40% of tickets are how-to questions with answers already in the knowledge base. Priya can answer those directly — drawing from your KB via the vector knowledge base she has per-agent — without involving a human. The remaining 60-80% get to the human with the classification already done and a drafted response already started.

The result is that the human queue gets shorter, and the time-to-first-response gets faster for the tickets that actually need human judgment.

What's running under the hood

Priya's graph in Avery NXR:

Scheduled trigger (every 15 min) OR Webhook
  → Fetch new tickets (Zendesk/Intercom API node)
  → Loop over tickets:
       → Classify topic + intent (local LLM)
       → Detect sentiment (classification node)
       → Detect urgency (classification node)
       → Vector KB retrieval (search per-agent KB)
       → IF simple how-to with KB match:
            → Draft response (local LLM)
            → Post reply via Zendesk/Intercom API
            → Tag ticket as auto-resolved
       → ELSE:
            → Draft starting response
            → Route to appropriate queue/agent
            → IF urgency = critical:
                 → Slack notification to on-call

What it costs

A cloud-LLM equivalent processing 10,000 tickets per month at $0.05 per ticket runs about $500/month or $6,000/year. For larger support orgs handling 100K+ tickets per month, the bill is in the high-five to low-six figures per year.

Priya runs on your local model. Cost per ticket is electricity.

The privacy story

Support tickets are surprisingly sensitive. They contain customer account details, often PII, sometimes billing information, sometimes information about the customer's business that the customer expects you to handle carefully. The cumulative content of a year of support tickets is a comprehensive view of your customer base — exactly the kind of data third-party AI providers shouldn't have arbitrary access to.

Priya keeps it all in-house. The tickets stay on your infrastructure. The classification stays local. The KB retrieval is from your KB. The drafted responses are produced on your laptop.

For support orgs in regulated industries — healthcare, financial services, government services — this is often the only architecture that passes legal review.

Try Priya in 5 minutes

If you've already got Avery NXR:

  1. Open the Agents tab
  2. Find Priya (PRIYA · SUPPORT)
  3. Click "Use this template"
  4. Plug in your help-desk (Zendesk or Intercom)
  5. Connect your knowledge base (upload docs, point at a URL, or sync from Notion/Confluence)
  6. Configure the routing rules
  7. Hit Run

She'll process the most recent batch of tickets immediately. Adjust the routing and auto-reply thresholds based on what you see.

If you don't have Avery NXR yet, request access at avery.software. Free Desktop tier, no card required.