Webinar Recap: Build an AI GTM Coach with Lovable & ElevenLabs

Most enablement teams have run a role play program. But it’s hard to run one that reps actually finish. And it’s even harder to build one around the unique conversations your team is actually having. 

In Live Workshop: Build an AI GTM Coach with Lovable & ElevenLabs, Morgan Jacobson (GTM Enablement at Lovable) and Remy Khoung (GTM Enablement at ElevenLabs) walked through how they solved for this by building an AI GTM coach.

They shared the behind the scenes of how they rolled them out to their teams at Lovable and ElevenLabs - and some of the tactical learnings behind the build. 

Limitations of sales role play 

Remy and Morgan were consistently running into these limitations with their sales role plays: 

1. Practice partners vary in quality, so sessions could sometimes become a walkthrough rather than a real buyer conversation
2. There was no objective measurement until the rep was already in front of a customer
3. Live sessions were hard to coordinate across time zones
4. Peer role play asks reps to risk looking unprepared in front of the people who evaluate them

A sales rep managing how they come across to their manager is spending attention on the room rather than on the skill.

Taking the audience out of the exercise is what makes real repetition possible.

To help solve for some of these, they decided to build an always-on AI GTM Coach that was trained on the best buyer conversations. 

How they built an AI GTM Coach 

ElevenAgents acted as the underlying agent with Lovable acting as the front end. 

The agent
The system prompt was the key piece of the agent. The scenario, the rules about what the buyer will and will not volunteer, and the scoring rubric all were part of the instructions. The model choice, language, and voice acted as the configuration on top of that.

The front end
Lovable handled everything the rep touches - the role play studio, transcript upload, the live scorecard, certification paths, and drill assignment with completion tracking.

Three decisions behind the buyer persona

Three choices shaped the build.

1. Tiered disclosure. The AI buyer does not volunteer information, so the rep has to earn it with a point of view or a sharp question. It is also what keeps a rep from clearing the exercise with three generic questions.
2. Structured scoring. The agent breaks character at the end and scores against defined criteria. In the build shown, that was technical accuracy, buyer credibility, and whether the rep pitched at the buyer's altitude.
3. Teaching mode. A rep who is stuck says so, the agent pauses, teaches the skill with a concrete example question or customer story, and the exercise resumes. This is what makes repeated use feel safe rather than exposing.

Demo 1: The Lovable role play studio

Scenario: A rep runs a discovery call against an AI buyer named Sarah, whose team struggles with consistent messaging and takes too long to train new reps. 

What was shown:

• The rep opening with an agenda, leading with a point of view on ramp difficulty, then pushing past the stated problem to business impact, at which point the buyer volunteers a lost deal
• A live scorecard filling in during the call and a full scorecard at the end, both configured in natural language against the company's own methodology
• Scenarios generated by prompting the Lovable agent directly, or by pointing Claude via MCP at a library of historical calls and having it build replicas of the most common ones

Why it matters: Role play quality and interface quality are two separate problems, and most programs only get to solve the first.

Demo 2: Configuring the ElevenAgents agent 

Scenario: The agent layer underneath the interface, shown through the configuration of an ElevenAgents build.

What was shown:

• The system prompt: scenario, disclosure rules, and scoring rubric, all as plain English instructions
• A first message that sets the tone and opens the scenario the moment the agent picks up
• Language selection across 70 plus languages, either one agent handling several or one agent per language
• Voice selection from over 10,000+ voices, with some performing better in particular languages
• An optional knowledge base for ICP and persona documents

Why it matters: A role play only works when the agent performs well. That comes from voice quality, low latency, and turn-taking - which sit in one stack. 

Best practices for building an AI role play coach

Write the rubric before the prompt. Define the criteria reps have to hit, then build the agent to test for them. Working the other way round produces a scorecard reps read as generic, which is the fastest way to lose them. In the rollout shown, five criteria came first and the agent was built against them.

Mine real calls for the scenarios. Rather than inventing scenarios, point Claude via MCP at a library of historical sales calls, have it summarise the ones that come up most, and build replicas of those. 

Keep responses short in the prompt. Two to four sentences. More than that, the agent tends to waiver more often. 

Match the model to the complexity of the conversation. A light, fast model for straightforward role plays, something more capable where the conversation genuinely needs it, balanced against latency and cost. 

Tune prompts for regional pushback. Deployment across regions was the easy part. The hiccup was prompt tuning, because a US enterprise buyer pushes back differently from one in EMEA, India, or Japan. Those conventions have to be worked out with local reps and written in.

Treat adoption as change management. Leadership alignment, VP led distribution, a credible scorecard, and competitive drills. The build takes an afternoon, but bringing tenured reps along takes longer. 

Watch the full session

Watch the full webinar here.