ElevenLabs says its internal agents answer more than 5,000 questions a month and opened 150 tickets last month, and the fix that mattered most was deleting fields rather than tuning prompts
AspireAI community · ElevenLabs
ElevenLabs describes an internal agent estate that started as a Slack question-answerer and grew into agents that act - building customer demos, opening Linear tickets, and answering data questions in domain-specific channels. The company reports more than 1,200 hours a month saved, more than 5,000 questions answered a month, dozens of customized demos built for customers, and more than 150 tickets created in the last month by a triage bot that searches Slack for context, dedupes against existing tickets, and only opens a ticket once there is enough detail to diagnose it.
The more useful part is what the team says it learned. Schema curation beat prompt engineering - the biggest accuracy gains came from removing fields rather than tuning the system prompt, because every extra field gave the model another place to guess wrong. Picklist typings were fetched at startup and injected into the MCP tool type definitions so the model can only choose real Salesforce values. Procedures get pruned as they overlap, and task verification after ticket creation checks the fields and the destination team to curb hallucinations and wrong-team routing.
This is a vendor's own account with no independent measurement and no baseline for the hours figure. One inconsistency inside the page itself is worth noting: the social meta description says '1,200 hrs/week' while the on-page headline and body say 1,200 hours a month, and only the on-page figure is quoted here. ElevenLabs also sells the product being described, and the Architect component it credits for restructuring procedures is not a public tool.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.