Zendesk Integration
Agent works inside existing queues, macros and SLAs.

An AI agent integrated into Zendesk and the product's admin API that resolves routine support tickets end to end, escalates with full context and is measured by an evaluation set of real conversations.

Overview
A B2B SaaS company with 4,000 customer accounts received 9,000 tickets a month. Two thirds were routine: password resets, billing questions, configuration changes and how-to questions already answered in the documentation.
We integrated an agent into their Zendesk and product admin API. It reads the ticket, retrieves the right documentation, performs permitted actions such as resending invoices or adjusting seat counts, and replies in the company's tone. Anything outside its scope is escalated with a summary and the evidence it gathered.
The Challenge
The client had tried a chatbot before. Customers hated it because it could not do anything, only link to articles. The new system had to take real actions, but a wrong action on a customer account was unacceptable.
Support leads also needed proof: which tickets the agent handled, how customers rated the answers and where it failed, before giving it more autonomy.

Our Solution
We defined three permission tiers for tools: read-only lookups run freely, reversible actions run with notification, and irreversible actions require a support engineer's approval in Zendesk. Guardrails check every outgoing message for policy, tone and data leakage.
Following eval-driven development, we built an evaluation set from 1,200 historical conversations. Every prompt or model change is scored against it before deployment, and production conversations are sampled weekly to extend the set.

Key Features
Agent works inside existing queues, macros and SLAs.
Grounded answers with links to the exact help-centre section.
Invoices, seats and settings changed through the admin API.
Read, reversible and approval-gated actions.
Humans receive context, evidence and a suggested reply.
Accuracy, tone and resolution rate tracked per release.
The Impact
The support team now handles complex cases while the agent clears the routine queue.
62%
Tickets resolved without a human
4.6 / 5
Customer rating of AI replies
-71%
First-response time
0
Unapproved irreversible actions