ElevenLabs hands routine work to internal agents and saves 1,200 hours a month

ElevenLabs has published a breakdown of how its own agents on the ElevenAgents platform work inside the company. The numbers: more than 1 200 hours saved per month, over 5 000 answers to employee questions, dozens of custom customer demos and more than 150 tickets filed by agents in the past month. 1 200 hours is roughly seven to eight full-time positions. The author, ElevenAgents head of product and growth Jake Tennant (previously a PM at Coinbase for four years), openly describes both the results and the engineering techniques behind them.

It started small. Agents in Slack helped people get internal information faster: user retention data, the right documentation, account history. Then they learned to act. An employee asks a question, reports a bug or requests a demo, and the agent either answers or carries out the request.

Who does what in Slack

Agents are assigned to specific teams:

• GTM bot (sales) answers questions about customers. Example queries: "Who are our top customers with logo rights in the UK, France and Germany?" or "What are our main API use cases?"
• FDE bot (Forward Deployed Engineer, sales and presales) answers product questions, for example which integrations a customer needs if they want an outbound AI SDR to book meetings on behalf of account managers. From a single message, it also builds a working agent for a specific prospect's scenario.
• Triage bot (support and engineering) turns feedback and bug reports into tickets. It created 150 tickets in a month; this used to be manual work for two teams.
• Product and product marketing agents answer data questions in topic channels. One example: how the free-to-paid conversion rate among US self-serve ElevenCreative subscribers changed over six months, and what drives it. The analysis arrives in a thread, where people can ask follow-up questions.

A ticket you're not embarrassed to open

The triage bot has separate Procedures for feature requests and for bugs. Before creating a ticket, it looks for context in Slack, checks existing tickets to avoid duplicates and verifies that there are enough details to diagnose the problem. Only then does it open a ticket in Linear with the appropriate priority. The bot determines the owning team from a document on areas of responsibility uploaded to the agent's knowledge base.

After the ticket is created, one more step kicks in: the agent checks that the fields are filled in correctly and that the ticket went to the right team. According to the author, this reduced cases of hallucinations and wrong team assignment.

Trim the schema, don't polish the prompt

The core engineering section of the article describes four techniques the team took away from running the agents in production:

• Schema curation matters more than prompt engineering. For retrieval agents, the biggest quality gain came from removing fields, not from tuning the system prompt. Every extra column or object field is one more point where the model can guess wrong. Cutting the schema down to a limited set of critical fields improved accuracy more than any prompt edit. The side effect is obvious: a shorter schema means less context per request.
• Picklist typing. Salesforce fields such as deal stage or account type have a fixed set of values. Without constraints, the model can produce a plausible but nonexistent value. Now the allowed values are pulled in at startup and injected into the type definitions of the MCP tool, so the model chooses only from real options.
• Procedure pruning. As an agent accumulates Procedures, they start to overlap. The Architect tool in ElevenAgents can restructure them so that trigger conditions are unambiguous and the logic doesn't conflict.
• Task verification. Checking the fields and the team after a ticket is created in Linear, as described above.

One separate observation on adoption: internal agents need to be accurate from day one, otherwise the team stops trusting and using them. In a shared channel, mistakes are more visible, but so are successes, which, according to the author, helped adoption.

What to take away

If you are building internal agents on top of a CRM, a tracker and a corporate chat, this case suggests the following order of work. First, cut the data schema down to the minimum fields without which an answer is impossible. Then restrict value types to the real enumerations from the source, for example through MCP tool definitions. Next, separate procedures by task type and periodically clean out overlaps. Finally, add a self-check step after every write action. It is best to launch agents in shared channels: errors are visible immediately there, and the agent's usefulness is visible to everyone.

This piece came out two weeks after ElevenLabs' earlier breakdown, in which the company described how it handles 22 000 support requests a month with ElevenAgents. Now the same product is shown as an internal tool for the company itself.