Internal Ops Agents

ElevenLabs, the voice-AI company, says the biggest accuracy gains for its internal Slack agents came from narrowing what the model could get wrong, not from rewriting prompts. In a post by Jake Tennant, who works on its ElevenAgents agent-building product, the company describes internal agents that answer staff questions and act on requests: a sales bot that answers account questions, a sales-team "forward deployed engineer" bot that answers product questions and assembles working customer demos, a triage bot that turns bug reports and feature requests into Linear tickets, and a data-analysis bot for product teams. ElevenLabs says they save more than 1,200 hours a month and answer more than 5,000 questions a month, and that the triage bot opened 150 tickets in the past month. The post does not say how the hours were measured.

The lessons all remove room for guessing. Cutting the data schema its retrieval agents see down to a few critical fields improved accuracy more than any prompt change the team tried, because each extra column was "one more place to guess wrong." For Salesforce fields that accept only a fixed list of values, such as deal stage, the agents fetch the valid values at startup and write them into the tool's type definitions, the description of what inputs a tool accepts, which the agents reach through MCP, the protocol they use to call tools. The model can then pick only a real option instead of inventing a plausible one. After the triage bot opens a ticket, it checks that the fields are set and that the ticket reached the right team, which the company says has curbed occasional hallucinations and wrong team assignments.

ElevenLabs' reason for the effort is trust: internal agents have to be accurate from the start, because a team that sees mistakes, especially in a shared channel, stops using them. The post reports no error rates or before-and-after comparisons, so the three techniques are design choices supported by the company's account rather than measured results.