AI-Powered Pitch Training
for cult Centre Managers
Chaitanya Gupta
Engineering Lead, Growth Org
The question
Could we make pitch training feel like an actual conversation?
When my team and I looked at center managers pitch training across cult centres for prospective members, we realised there was a gap between knowing the pitch and applying it in real conversations. They were trained on what to say, but customer conversations rarely followed a fixed script. When prospective members raised unexpected fitness concerns or nuanced questions, newer managers sometimes struggled to adapt in the moment. It wasn’t because they lacked the information, but because they had limited opportunities to practise handling situations outside the standard pitch.
As cult expanded, we needed a way to give managers more opportunities to practise these situations without making training dependent on one-on-one coaching or repeated classroom sessions.
“The real challenge wasn’t teaching Centre Managers on what to say...it was helping them adapt to every kind of customer conversation. With AI-powered personas, we could give managers personalised practice across different consumer types and scenarios, helping them build confidence while ensuring every prospective member received the right guidance.”
Chaitanya Gupta · Engineering Lead, cult
We Turned Pitch Training Into a Conversation
We didn't want Centre Managers to learn the pitch by reading another document or answering a fixed set of questions. We wanted to see how they handled the conversation itself. So I worked with the team to build an AI voice-based training experience that could simulate different prospects.
Persona model
One pitch. Many kinds of prospect.
Instead of giving every manager the same conversation, we could simulate complete beginners, experienced athletes and more challenging negotiators. Each persona could be calibrated to a difficulty level and brought together into structured training modules.
Assessment criteria
Judge the scenario, not a fixed script.
Each conversation could be evaluated against parameters defined for that persona such as active listening, objection handling or pitch accuracy with those parameters passed to the LLM as part of the assessment prompt.
If every conversation can take a different path, how do we decide whether a manager handled it well?
The answer
Define assessment criteria for each persona. This meant the conversation could be judged against criteria relevant to the scenario rather than a single fixed checklist. The assessment then generated a score that could be compared against the passing score configured for that persona within the training module.
Live simulation
Riya · Beginner
“I’m new to fitness. What would you recommend?”
Pitch practice · Stage 02
Active listening
—
Objection handling
—
Pitch accuracy
—
Simplifying the Training Experience for Our Employees
The first version worked.
The setup didn’t.
The first version of the assessment ran through the Partner Portal as a web experience. It worked, but the setup created a few practical problems:
Centre Managers needed a laptop to complete the voice training.
Background noise could interfere with the conversation.
Carrying a laptop around the centre wasn't always practical.
The web-based setup gave managers less flexibility over when they could complete the training.
Meet managers where
they already work
We realised the problem wasn't the training itself. It was how managers were accessing it. We moved the training experience into Teams, an app that Centre Managers and Trainers already use as part of their everyday work, including marking attendance and checking member check-ins at the centre.
That made it possible to run shorter voice simulations within a tool managers were already using, without requiring a separate device or workflow.
“The AI pitch training is helping us improve our communication skills. It is good for practice and also helps build our confidence.”
• Sejal, Centre Manager
That feedback reinforced something we were learning as we built: a training experience isn't useful just because the underlying technology works. It has to fit into the environment where people actually work.
The conversation needed a system around it
BEYOND THE CALL
The voice conversation was only one part of the system. We also needed to know who was having it, which training module it belonged to and what happened during the interaction. We separated those responsibilities into a few layers.
01 / PERSONA
Simulated prospect
An agent persona defines the customer, their context, difficulty and the objections a manager must navigate.
02 / CURRICULUM
Training module
A module brings personas into a curriculum. Their mapping controls order, passing score and retry rules.
03 / ASSIGNMENT
Employee mapping
A second mapping connects the module to the employee taking it, so every attempt belongs to a learning journey.
One ID ties every attempt together.
When a manager started a conversation, the system created an assessment entry. A unique Conversation ID linked the audio, transcript and metrics to that attempt. Once processed, the interaction was evaluated for active listening, objection handling and pitch accuracy, to determine a clear passing score that showed managers where they needed to improve. The L&D team could create modules, assign them to employees and track performance, including access to interaction recordings.
TRAINING FLOW
The conversation became a trackable unit of training.
Simplified system / sequence diagram
Voice AI at scale and its impact
20×
growth in assessment attempts between December and May
>90%
completion rate, sustained through June and July
2×
improvement in first-sale conversion after training
30–45
days in a typical self-paced completion window
Fewer calls. More structured learning.
The aim wasn’t just to have fewer conversations. It was to make each one count.
We deliberately designed the tool around broad completion timelines, typically 30 to 45 days. Within that window, managers can train at their own pace across multiple stages, each featuring a distinct persona and difficulty level. Managers move through those stages in sequence. Pass, and the next stage opens. Fail, and they retry the same stage.
This timeline structure allows managers to complete the curriculum across multiple self-paced sessions. Admins can create modules, decide how they are structured and assign them to employees. The system then tracks progress across the module rather than treating each conversation as a separate attempt.
Voice interactions became part of a structured learning cycle rather than the training itself.
What the new model makes possible
The new model changed more than the training schedule. For our engineering and L&D teams, it provided a holistic view of skill development across a full learning journey rather than judging performance on isolated call snapshots.
The architecture also leaves room for multiple training modules and different learning journeys for different employee groups. Future iterations can build on the same foundation with more flexible programmes and deeper performance analytics.
Shoutout to
The amazing team who made this initiative happen!