The Setup: 2,000 Tickets, 19 Agents, and an AI Tool That Was Not Working
When we first connected with the support team at a growing B2B SaaS platform, the situation was familiar but urgent. They were handling roughly 2,000 support tickets per month -- a volume that was steadily rising as their customer base expanded. A 19-person team was managing everything over email through Zendesk, with no live chat support because the team did not have the bandwidth to staff it in real time.
The ticket mix broke down into two distinct streams. About 70% of volume came from end-user tickets -- employees at client companies hitting login issues, website access problems, password resets, and other common friction points. The remaining 30% were admin tickets -- raised by the platform administrators at each client company, covering product functionality questions, feature requests, bug reports, integration issues, and other items that required deeper technical expertise.
The team had already tried to automate. They had been paying for an AI automation tool for six months. It was not working. And the reasons it was not working turned out to be more instructive than the tool itself.
Why the Previous AI Tool Failed
The team had invested in an AI automation tool with a simple premise: upload a knowledge base, connect it to Zendesk, and let the bot handle the repetitive stuff. In theory, this should have been straightforward. Their end-user knowledge base was modest -- roughly 15 documented issue types on a single page. The use case was as simple as AI support gets.
Six months in, the tool was still only generating internal notes for human review, not sending responses to customers. The support lead was manually reviewing every AI-drafted response in a Google Sheet, flagging errors, and feeding corrections back to the vendor. It never got to production.
Three specific failures kept the tool from going live:
Hallucination on a Simple Knowledge Base
Even with only 15 documented scenarios, the AI was inventing answers. A customer would report a blocked website, and the bot would suggest steps that were not in the knowledge base. The team could not trust the output. They tried adding explicit instructions, refining the knowledge base document, and working with the vendor directly. The hallucinations persisted.
No Customer-Specific Routing
The platform serves enterprise clients, and not all clients are equal. Some enterprise customers have their own internal ticketing systems and specific escalation paths. The previous tool had no concept of this. It was a single-agent system with a single knowledge base. Every customer got the same response regardless of who they were.
No Image Understanding
End users frequently attach screenshots when reporting issues. About 20% of tickets included images as the primary context. The previous tool could not process images at all. It would ignore the attachment and attempt to answer based solely on the text.
Zero Implementation Support
The vendor support team was unresponsive. The support lead had been working solo for six months trying to make the tool functional. The tool that was supposed to save the team time was consuming more time than the manual process it was replacing.
The Solution: Multi-Agent Workflows With Conditional Routing
The core insight was that a single AI agent cannot handle the variety of scenarios a B2B support team faces. Instead, we built a multi-agent workflow with conditional routing -- three distinct AI agents, each specialized for its job, with a triage layer that decides which agent handles each ticket.
How the Triage Works
The workflow triggers whenever a new Zendesk ticket is created. A conditional agent examines the ticket and routes it to one of three branches based on the requester email domain and the nature of the issue. This is not keyword matching. The triage agent understands intent.
Branch 1: Generic Customer, Knowledge-Base Response
For the 70% of tickets that are standard end-user issues, the ticket routes to an AI agent trained exclusively on the team knowledge base. The agent refers only to the uploaded documentation. It cannot invent answers. If the knowledge base does not cover the issue, the agent escalates to a human rather than guessing. This agent also processes images. When a user attaches a screenshot, the agent analyzes the image and maps it to the relevant knowledge-base entry.
Branch 2: Enterprise Customer, Tailored Response
For tickets from specific enterprise domains, the workflow routes to a completely different AI agent. This agent has no knowledge base. Its only job is to send a standard acknowledgment without attempting to resolve the issue. No hallucination risk because there is nothing to hallucinate from. This branch is infinitely configurable with time-based rules and domain-specific behaviors.
Branch 3: Technical Issue, Slack Escalation
For API issues and integration failures, the workflow routes to a third agent that does not respond to the customer at all. Instead, it summarizes the ticket and posts it directly to a designated Slack channel. The Slack channel has its own AI agent with its own knowledge base and connected tools. A monitoring agent watches the thread and routes the resolution back to the customer via Zendesk.
The Zendesk Integration: Everything Stays Where It Is
Nothing changes in the team day-to-day Zendesk workflow. The AI agents operate on top of Zendesk. Tickets still live in Zendesk. Status changes, tag additions, assignee updates, priority settings -- all of it happens through the AI workflow directly in Zendesk fields.
Knowledge Insights: The Flywheel That Was Missing
The previous tool biggest structural failure was the absence of a feedback loop. The knowledge insights module continuously analyzes all tickets and surfaces new article recommendations. It recommends updates to existing articles. This is how a 15-article knowledge base becomes 20, then 30, then a comprehensive resource that covers 90% of incoming scenarios. Custom insights let the team create their own analytics rubrics running on 100% of tickets.
The Projected Impact
Of 2,000 monthly tickets, approximately 1,400 are end-user tickets with a small, well-defined set of issues. The support lead estimated that with working automation, 95% of these could be deflected. The total addressable automation reaches roughly 70-75% of all volume. Implementation timeline: one week for the core workflow, two to three weeks for the full deployment.
Key Takeaways for B2B SaaS Support Teams
- Single-agent AI does not work for multi-scenario support. The hallucination problem is often a routing problem in disguise.
- Customer-specific responses require multi-agent architecture. Enterprise clients expect tailored treatment.
- Image understanding is no longer optional. If 20% of your tickets include screenshots, an AI tool that ignores images is ignoring 20% of the problem.
- Implementation support is the actual product. The difference between a tool that sits unused for six months and one that goes live in a week is almost always the vendor willingness to own the implementation.
- The flywheel matters more than the launch. Going from 50% to 75% deflection is about a system that continuously identifies knowledge gaps.
What This Could Look Like for Your Team
If you are running support for a B2B SaaS product on Zendesk, the questions worth asking are: How much of your ticket volume is genuinely repetitive? Do you have enterprise customers who need different handling? Is your current AI tool actually sending responses, or is it stuck in internal-note purgatory?
Want to see this on your own tickets? We build a working demo on your actual knowledge base in under a week. Book a demo
