ElevenLabs Hands Its Support Desk to an AI Agent: 22,000 Tickets a Month, 65% Without Humans

ElevenLabs resolves 22,000+ requests a month with its own product, ElevenAgents. This is not a customer demo but the company's actual support line. According to the team, the agent resolved 65% of tickets from start to finish over the past month. Average first-response time fell to 60 seconds; it used to take hours. Users rated 70% of the AI-led conversations positively, which matches the rating of human operators.

The agent works with text, screenshots and PDFs. It is connected to the ticketing platform, holds long email threads, looks up account data, takes actions and escalates cases to staff. If both figures refer to the same month, the total flow is about 34 thousand tickets (22,000 at a 65% share). This is our own calculation, not a company statement.

SOPs as the agent's source code

The team's key decision is a boring one, and that is why it works. Support already had detailed standard operating procedures for every major category: refunds, billing changes, questions about professional voice cloning. The engineering task came down to translating that knowledge into a format the agent executes predictably. That took two ElevenAgents mechanisms:

• Procedures turn SOPs into repeatable workflows. The team ported its support playbooks for every major ticket category.
• Structured Procedures are used for critical operations, such as refunding money. Each step runs deterministically, with no free improvisation by the model.
• Webhook tools give the agent live user data: subscription plan, credit usage, the voices the user owns. Without them, the agent could only offer advice; with them, it solves the problem outright.

The line between a flexible model and a rigid procedure sits where a mistake costs money. The model conducts the conversation, while a script controls the issuing of a refund.

How quality was tested before production

Before going live, and after every configuration change, the team used three tools:

• Channel behavior settings. An email user wants an answer in as few turns as possible, in a business-correspondence format. A voice customer needs a lively real-time dialogue with many more turns. Each channel gets its own bar.
• Simulation tests. The agent was run through roughly 50 of the most common support scenarios. Architect, a natural-language agent builder, then edited the system prompt and procedures until the tests passed consistently. The suite is rerun after any edit to avoid regressions.
• Shadow mode. The agent worked on real traffic inside the ticketing system, but its replies did not go out to customers. Once the team was confident in the quality of the answers, it was switched on in production.

Architect as a root-cause tool

Support staff use Architect every day to find out where the agent gets stuck. A single natural-language query returns a breakdown of tickets by main category. Staff can also ask separately why a specific case went to a human. The cause may be a gap in the knowledge base, a missing tool or an unclear internal policy. Once it finds the cause, Architect proposes fixes.

Changes also go through Architect: it drafts and validates the updates, then creates new simulation tests and adds them to the shared suite. That way new procedures are checked immediately, and old ones do not break unnoticed. Humans stay in the loop: they clarify customer needs, update policies and make sure the agent does not overfit to a single type of ticket. Architect handles the rest, which lets support keep pace with an expanding product portfolio.

What you can take from this case

The overall approach carries over to any team that runs support:

1. Start with existing SOPs for your top request categories. If there are no procedures, there is nothing to automate yet. 2. Separate free-form dialogue from irreversible actions. Anything involving money or access rights should be built as deterministic steps. 3. Give the agent access to account data through tools, or the share of fully resolved tickets will hit a ceiling. 4. Build a set of 30–50 typical scenarios and rerun it after every change to a prompt or procedure. 5. Launch in shadow mode on the live stream before showing replies to customers. 6. Use separate settings for email, chat and voice: expectations for the number of turns and for tone differ across them.

The 70% human rating shows that speed was not bought at the expense of quality. But a 65% end-to-end resolution rate means a third of tickets still go to humans, and those escalations supply the material for the agent's next iteration.

The case comes amid a series of ElevenLabs releases around its agent platform. In September, ElevenAgents arrived on the OpenAI Marketplace and Eleven v4 Turbo launched inside ElevenAgents; in August, ElevenLabs MCP for Claude came out. The company's own support serves as a showcase for the platform.