Building an AI Conversation Coach at Rightmove To Help Practice Tough Conversations

What they built

• An AI conversation coach built inside Microsoft Copilot to help people leaders at Rightmove practice difficult conversations before having them in real life
• The agent operates as an interactive, back-and-forth tool (not a fully autonomous agent) — it asks clarifying questions, provides frameworks, and offers to role-play the conversation
• Pre-loaded prompt starters cover the most common scenarios (e.g. "I have a tricky conversation coming up", "I want to recognise someone positively")
• Built with internal guardrails so it stays within approved content and does not browse the internet or hallucinate from external sources
• Includes hard escalation rules: if a sensitive topic arises (e.g. sexual harassment), the agent immediately directs the user to a human business partner

Why they built it

• The people team's service ticket system was flooded with requests for informal pre-conversation chats with business partners - leaders wanted reassurance and a sounding board before having tough conversations
• Two external vendors quoted significant sums to build an equivalent tool; Mark decided to build it internally instead
• Goal: free up business partner capacity for higher-value work while giving people leaders a judgment-free, always-available space to practice

Tools used

• Microsoft Copilot (agent-building capability within Microsoft 365) - chosen because Rightmove had already rolled it out enterprise-wide, keeping data inside the corporate environment
• Internal frameworks loaded into the agent:
  • SBI (Situation–Behaviour–Impact) feedback model
  • Radical Candor
  • SCARF motivational model (Status, Certainty, Autonomy, Relatedness, Fairness)
  • Erin Meyer's Culture Map (added after pilot surfaced cross-cultural conversation needs)
  • Rightmove's own internal performance approach
• Copilot's dictation/voice input feature - useful for external processors who prefer to talk through their thoughts

How you could build this yourself

1. Audit your ticket/request data - identify the most common scenarios where managers seek pre-conversation support; these become your agent's core use cases
2. Choose a platform already in your tech stack that keeps data inside your enterprise boundary (e.g. Microsoft Copilot, if you have M365 licences)
3. Identify the frameworks your organisation already uses for feedback, performance, and inclusion - these become the agent's knowledge base and replace generic internet content
4. Draft your system prompt / agent instructions - specify tone, scope, which frameworks to apply, and what the agent should not handle
5. Build hard escalation guardrails - define the categories (e.g. harassment, safeguarding) where the agent must immediately redirect to a human; write these into the instructions explicitly
6. Run a focused pilot with one high-volume team or division - recruit willing managers, observe where the agent gets stuck (e.g. neurodiversity, cultural nuance), and use that data to iterate
7. Add missing context from the pilot - in Rightmove's case this meant adding the Culture Map and inclusion guidance after the first round of testing
8. Pair the rollout with a complementary people initiative - Rightmove launched the coach at the same time as a strengthened business partnering model, giving managers a clear choice between AI and human support
9. Use pilot participants as internal advocates - their honest, lived-experience stories (positive and "I still needed a human") are more credible than any top-down communication
10. Track leading indicators - monitor manager effectiveness scores in engagement surveys and watch for a reduction in informal pre-conversation tickets to evidence impact
11. Listen for adjacent use cases - Rightmove's IC community asked for something similar, which led to a separate performance and growth agent

Source

Built by Mark Abraham, Group Head of Organisational Development at Rightmove, and presented as part of an Open Org "Shipped by HR" session.