Insurance Sales Lab
Insurance agents practice calls with an AI prospect, scored on the company's script
Insurance Sales Lab trains insurance agents on its own learning platform. We added AI voice roleplay inside it: agents call an AI prospect that follows the company's scripts, get scored on its 15 point checklist, and admins stay in charge of what the AI says.
EdTechAI voice roleplay trainerAI added to a live product
The challenge
Agents learned the company's structured sales script through human led roleplay and coaching, which does not scale and gives uneven feedback from one coach to the next.
What we did
We built voice roleplay into the client's platform extension: an AI prospect that follows admin owned scripts over live voice, and a LangGraph workflow that scores each call on the company's 15 point checklist.
The outcome
The trainer is in production with the client's agents inside the platform they already use, practice now sits inside the lessons, and the client's admins own the scripts, scenarios and scoring standard.
- Status
- In production with the client's agents
- Lives inside
- The client's learning platform, no separate sign in
- Scored on
- The company's own script and 15 point checklist
- Controlled by
- Admins: scripts, scenarios, AI instructions
The starting point
Insurance Sales Lab trains insurance sales agents and sells that training through its own membership and learning platform, with courses, lessons, organizations, sub organizations and subscription tags. Agents learn a structured sales script and practice it before real calls.
What had to change
Practice relied on human led roleplay and coaching, which does not scale and gives uneven feedback. The company wanted agents to rehearse realistic voice calls with an AI prospect at any hour and be scored the same way every time, against its own script rather than a generic rubric.
What could not break
- No separate sign up or sign in, since the trainer runs inside the client's platform extension
- Agents are scored on the company's own script and 15 point checklist, never a generic rubric
- Training data stays model neutral, and voice, logic and scoring can each be swapped
- Only an admin, under manual review, can change a scorecard
- Recording needs the agent's consent at first use
The calls we made
We framed it as putting the company's own coaching standard into the AI, inside the platform agents already use, with its admins in charge of that standard.
- 01
A short proof of concept, built to move in
We started with a standalone proof of concept with hard coded scenarios and a live voice demo, scoped from day one to move into the client's platform extension on its existing sign in, users and documents.
Ruled outA separate product beside the platform, with its own sign up, sign in and user list.
- 02
Scripts in the database, not in the code
After the first demo, scripts, scenarios, checklists and global AI instructions moved into the database and prompts are built from them, so the client's admins steer the AI without a code change.
Ruled outThe hard coded scenarios of the proof of concept.
- 03
Small scoring nodes, not one big prompt
Feedback runs as a LangGraph workflow of separate nodes for the checklist, script adherence, objection handling, coaching summary and analysis, each with structured output, so every step stays narrow and easier to debug.
Ruled outOne prompt that grades the whole call at once.
- 04
Model neutral by design
Training data is stored in a model neutral format, and voice, logic and scoring are separate modules, so each can be upgraded without touching the others.
Ruled outTying the training data and scoring to one model vendor.
- 05
Scores change only through admin review
Agents see a read only scorecard, an admin can edit it only under manual review, and coach mode instructions stay hidden from agents.
Ruled outLetting agents edit their own scorecards.
How the system fits together
What we built
Live voice roleplay
Agents pick a scenario, a prospect personality and a difficulty, then talk over WebRTC to an AI prospect that stays on the selected script, in practice or coach mode. Asking to change the script mid call updates the prompt through a backend tool call.
Admin control
Script and scenario management with categories, global AI instructions, prompt visibility per user, an admin controlled beta feature list and session logs for every call.
Feedback after every call
An overview with AI summary, recording, transcript, scorecard and key recommendations, an AI coaching tab against the 15 point checklist, and objection handling with timestamped questions and objections.
Analytics for training managers
Dashboards for sessions, completion rate, average score, top performers and the most challenging objections, plus engagement reports for each sub organization.
Practice inside the lessons
Roleplay scenarios embedded in course lessons and linked to the platform's documents, with related lessons suggested after each call.
How it works
Pick
The agent picks a scenario, a prospect personality and a difficulty, from the roleplay screen or inside a lesson.
Call
The AI prospect talks back in real time, stays on the selected script and persona, and switches script when the agent asks.
Score
The feedback workflow checks the call against the 15 point checklist, the script and the objections raised.
Coach
The agent reads the scorecard, coaching notes and timestamped objections, with related lessons to study next.
Review
Admins see trends across agents and sub organizations and can review and adjust any score.
A closer look
What changed
The roleplay trainer is in production with the client's agents, inside the platform they already use, with no separate sign in. Admins own the scripts and the scoring standard, and practice now sits inside the course lessons members already pay for.
Under the hood
- Datics role
- AI added to a live product
- Voice
- OpenAI Realtime over WebRTC, with voice activity detection tuned for the client and mid call script changes through a backend tool call.
- Scoring
- A LangGraph workflow with structured output per node, AssemblyAI transcription with speaker detection and an audio analysis node for talk and listen time.
- Platform
- The client's GraphQL APIs for sign in, organizations, subscription tags, documents, courses and lessons, with the trainer's own data in PostgreSQL behind a Fastify backend.
- Controls
- Coach mode instructions encrypted and hidden from agents, prompts split into static and dynamic parts, scorecards editable only through admin review and recording consent at first use.
AI and models
Data and retrieval
Apps and APIs
Also in the stack
How it shipped
Proof of concept
WebRTC voice, hard coded scenarios and a first live demo, built to move into the client's platform extension.
Admin owned AI and feedback
Scripts, scenarios and global AI instructions moved into the database, the LangGraph feedback workflow went in, and a production ready build went to the client for a feedback round.
Dashboards and tuning
Admin and user dashboards, session logs and voice notes, plus model and voice activity detection tuning the client asked for.
Coaching modes
Practice and coach modes with separate prompts, difficulty levels, scripts linked to scenarios, mid call script changes, AssemblyAI transcription and audio analysis of talk and listen time.
Inside the lessons
Prompt visibility per user, beta feature control, platform documents linked to scenarios, related lessons, scenarios embedded in course lessons and sub organization reports.
Built by a team that knew the platform
Our engineers already worked in the client's own repository on its sales agency platform, directed through daily syncs, so the trainer was designed around a platform we knew.
Inside the client's platform
The trainer lives in the client's platform extension and uses its sign in, organizations, subscription tags, documents and lessons through the platform's GraphQL APIs.
A demo after each step
We shared a demo recording with the client after each step and folded its feedback, such as model and voice detection tuning, into the next sprint.
What this unlocked
The client then asked us to score its agents' real sales calls the same way, which led to a desktop recorder that works with any meeting tool. Its roadmap adds real time AI feedback during live calls.
Related work
Have a product with the same kind of constraints?
Tell us what already exists and what you want it to do. We will tell you what it takes.