Intel / article

Building An Agent For Our Own Docs

Source: elevenlabs.io
Author: ElevenLabs editorial sitemap
Date: Published Sep 22, 2026

Why it matters

Gives a replicable method for validating a support agent: an LLM judge with custom success criteria, checked against human review, plus a check against the knowledge base.

Key takeaways · AI-distilled

ElevenLabs reports its docs voice resolves or correctly redirects about 80% of inquiries across roughly 200 calls a day, as scored by an LLM evaluator. A human review of 150 calls found 89% of relevant support questions answered or redirected correctly.
The LLM evaluator agreed with human reviewers on 81% of calls for whether the inquiry was solved and 83% for whether answers stayed within the knowledge base. The team ran the human check because an evaluation LLM does not judge every call correctly.
The agent struggled with account issues, pricing and discount requests, and vague problems. Despite prompts to ask clarifying questions, the model tended to answer with loosely relevant documentation, so the team says support agents need aggressive prompting to ask first.
Voice changes the prompt design: the bans bullet points, long lists and code samples, spells email addresses out as spoken words, and sends coding questions to documentation pages or Discord instead of reading code aloud.
Rather than retrieval, the agent carries an 80,000-character summarized version of all ElevenLabs documentation plus FAQs in its , and uses three redirect tools: to docs pages, to email support, and to external links such as sales or Discord.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source elevenlabs.io