How We Hit 83% Autonomous Support (With Better Documentation)

Quick Overview

The presentation concludes that achieving 83% autonomous support for self-serve tickets is possible by implementing a Human-In-The-Loop (HITL) process that relies on composable AI agents, where documentation is consistently updated by an AI agent (Documentation Drafter) that drafts changes based on research from another agent (Support Researcher), with all changes requiring human review before merging.

Key Points: The primary outcome achieved was 83% of self-serve support tickets being solved end-to-end by AI autopilot as of January 2026, according to Relevance AI. The success is attributed to a Human-In-The-Loop (HITL) process ensuring that no change goes live without human approval, involving steps: Trigger -> Research -> Draft -> Review -> Merge. The documentation workflow relies on composable agents, specifically using Claude Sonnet 4.5 for complex reasoning and writing, and GPT-4.1 for structured API calls (like those to Linear). The 'Doctor' agent orchestrates the workflow, coordinating other specialized agents like 'Support Researcher' (using 14 tools including Google Search, Firecraw, Pylon, Notion, Linear) and 'Documentation Drafter'. The Documentation Drafter agent automates writing by creating a branch, reading existing content, committing changes to MDX files, updating docs.json navigation, and naming the branch after the Linear issue. A key pattern identified is that the bottleneck for documentation work is context-switching and research, not the writing itself, which AI agents excel at assisting with. Future planned additions include automated release detection to trigger documentation updates based on code changes in new releases.

Context: This presentation, titled 'Scaling Documentation with AI Agents' by Alex Waite from Relevance AI, details the methodology and results behind achieving high levels of autonomous support resolution through advanced AI agent workflows. The talk builds upon a previous session comparing 'Auto-Pilot vs Co-Pilot' agents, focusing specifically on the mechanics that make these agents effective, particularly in the context of technical documentation updates.