Building an AI GTM Coach with Lovable and ElevenLabs: Webinar Review

The GTM Enablement teams at Lovable and ElevenLabs built an always-available AI GTM coach to address the limitations of sales role-playing. Traditional role-playing made repetitive training difficult due to variability among practice partners, a lack of objective evaluation, time zone coordination issues, and the pressure of performing in front of colleagues.

The architecture uses ElevenAgents for the conversational agent, while Lovable provides the frontend used by representatives. The agent's system prompt includes scenarios, the scope of information the buyer voluntarily reveals, and evaluation criteria. Lovable offers a role-play studio, transcript uploads, real-time scorecards, certification paths, and task completion tracking.

Buyer personas were designed around three principles:

• Staged information disclosure: The buyer reveals additional information only when the representative presents a clear perspective or sharp questions.
• Structured evaluation: After the conversation ends, scores are assigned based on technical accuracy, credibility from the buyer's perspective, and appropriate pitching levels.
• Teaching mode: If the representative indicates they are stuck, the agent pauses the conversation, explains techniques using example questions or customer cases, and then resumes the role-play.

In Lovable's role-play studio, we conducted discovery calls with virtual buyers who struggle with message consistency and onboarding new representatives. The real-time scorecard and final evaluation rubric allow the company's sales methodology to be configured in natural language. You can also directly prompt the Lovable agent or use Claude and MCP to replicate frequent scenarios from past call records.

In the ElevenAgents setup, scenarios, information disclosure rules, and evaluation criteria are written in natural language, and the first message sets the conversation's tone and starting point. It also supports over 70 languages and more than 10,000 voices, and allows optional linking of ICP and persona documents to the knowledge base.

When building, you must first define evaluation criteria before writing prompts. The quality of the role-play depends not only on the evaluation rubric but also on agent performance aspects such as voice quality, low latency, and natural turn-taking, as well as an easy-to-use interface.