How we built a self-serve AI HR assistant - Coaching Academy | Episode 4

Let’s be honest: every People team knows the feeling. The answers managers need are scattered across many notion pages, old Slack threads, and the institutional memory of a few key people. Consistency is a constant struggle, and your team becomes a bottleneck for even the most common questions.

Time to step behind the scenes of our Coaching Academy! Through our experience, we reveal the coaching practices that prove effective daily, putting employee fulfillment at the heart of everything we do.

At Alan, we didn't want to just scale our People team; we wanted to amplify it. Our goal was to give every coach and lead a thinking partner—one that speaks our culture, understands context, and, most importantly, knows when to say, “You should talk to a human.”

This is the story of our journey building Lumo, our AI HR assistant. It’s not a perfect playbook. It’s a candid look at our guiding principles, our early stumbles, and the ongoing work of building a tool that’s genuinely helpful.

Step behind the scenes of our Coaching Academy! Through our experience, we reveal the coaching practices that prove effective daily, putting employee fulfillment at the heart of everything we do.

Before Lumo, we faced a familiar set of growing pains:

• Scattered knowledge: our main problem wasn’t too many tools; it was too much content. For any given HR topic, a dozen different Notion pages could exist: the official policy, an old project brief, a new template, and a “best practices” guide from a 2023 training. It was impossible for a coach to know which one was the current source of truth, creating inconsistency and uncertainty.
• Inconsistent application: without a single source of truth, the handling of similar situations could vary from one team to another.
• The People team squeeze: our team was spending too much time on recurring questions, delaying their ability to tackle more complex, strategic work.

We knew technology could help, but only if it was built on a foundation of trust and aligned with our unique "two-leg" model, where both a coach (for personal growth) and a lead (for team performance) support every Alaner.

You can't automate what you haven't documented. An AI is only as good as the knowledge it has access to. The most valuable work isn't prompt engineering; it's the unglamorous, manual effort of creating a clear, comprehensive, and accessible knowledge base.

That is the theory, at least.

Our approach was simple:

1. Consolidate: we gathered and gave Lumo access to our People knowledge, Alan’s toolkits, processes, leadership principles, and anonymized learnings from past situations.
2. Clarify: we defined bright lines for where self-service is appropriate and where a conversation with the People team is non-negotiable. In simple terms, we treat Lumo as a first-step thinking partner, not a final decision-maker.

   • Lumo is great for: Preparing for a uncomfortable conversation, getting a script for giving feedback, understanding a policy, or finding the right questions to ask about career growth. It helps with the "how do I start?" moments.
   • The People team is essential for: Situations involving an employee's health or wellbeing, persistent performance issues, interpersonal conflict, or anything with potential legal or compliance implications. Human judgment are irreplaceable here.
   • Let's make this concrete with an example: a performance concern.

     
3. Automate: only then did we layer on AI as an intelligent, conversational interface to make that knowledge instantly accessible and actionable.

To make it concrete, here’s a common scenario where a lead might turn to Lumo for help.

A Lead asks Lumo

How Lumo responds

Great question. Approaching this with curiosity is key. Here’s a plan to help you prepare:

1. Objective & mindset: Your goal is to understand, not to judge. Assume positive intent. Frame the conversation around support: "I want to make sure you have what you need to do your best work."

2. Key questions to ask: Observation-based opener: "I've noticed your energy seems a bit lower in our team syncs recently. How are things going?" Workload check: "How are you feeling about your current projects and workload?" Blockers inquiry: "Are there any challenges or blockers that are getting in your way?"

3. Sample script starter: "Thanks for meeting. I wanted to create some space to check in. I've noticed [mention a specific, factual observation, e.g., 'you were quieter than usual in yesterday's planning session']. I want to understand what’s going on and how I can best support you."

4. Alan resources to keep in mind: Remember our Feedback Principles on being direct and empathetic. If the conversation touches on wellbeing, here is our Mental Health Support Guide

🚨 When to Loop in People: If you learn the disengagement is tied to a health issue or a conflict with another team member, it's a good time to bring in the People team for guidance.

Trust isn't given; it's earned. We built Lumo on a foundation of transparency and security.

• The right data: Lumo is built on Dust and uses a Retrieval-Augmented Generation (RAG) model. In simple terms, it doesn't just guess; it retrieves answers directly from our approved People knowledge base.
• Context is key: through a Slack workflow, Lumo can be given high-level context about an Alaner (like their level, recent feedback topics, or leave information) to provide more tailored advice.
• Clear boundaries: we are explicit about what Lumo should escalate to the People team*: eg anything related to a potential departure.*

"Lumo is asking the exact right questions to trigger smart discussions." — People Team Member

Launching a tool is easy; getting people to use it is hard. We focused on targeted, organic adoption.

• Narrow your audience first: we began with new coaches and leads, integrating Lumo training directly into their onboarding.
• Meet people where they are: we built Lumo into Slack, the tool our teams already use all day, every day.
• Show, don't just tell: we used live demos on real-world cases during training sessions and shared success stories in our internal "Coaching Gazette."

While we're still in the early innings, the initial signals from our first semester of testing have been incredibly encouraging:

• 100+ Alaners have given Lumo a try.
• Over 2,000 messages have been sent, each one a chance for us to learn what our users need.

"Time-saver to prepare coaching sessions and share feedback with the right tone and alan’s vocabulary." — A Coach at Alan

Building Lumo has been a masterclass in iteration. Here are our key learnings:

1. Problem: early answers were too verbose.

   • Fix:We're constantly refining our prompts for conciseness, but it's a delicate balance between being thorough and being overwhelming.
2. Problem: a prompt optimized for one task often degraded its performance on another.

   • Fix:We learned that a prompt tuned for direct, factual answers (like policy questions) could feel robotic when handling sensitive coaching moments. To fix this, we now maintain a "golden set" of diverse test scenarios—from performance feedback to wellbeing check-ins. Every prompt change must be tested against this set to ensure we don't sacrifice empathy for efficiency, or vice-versa, maintaining a consistent "fil rouge" for the user.
3. Problem: manual context was a chore.

   • Fix:We built the Slack integration to automatically and securely provide the necessary context. The integration is helping, but it's not a silver bullet. We're still working to make the experience feel seamless.

Building with AI means being honest about what you haven't solved. Here are the challenges we are still actively wrestling with:

• Nailing the tone. How can an AI be an empathetic coach one moment and a direct policy expert the next? We're still working to make Lumo’s tone feel right for every situation.
• Measuring what truly matters. It's easy to count users, but much harder to know if our tool led to a better, more meaningful conversation. We are still searching for the right ways to measure real-world impact.
• Ensuring Lumo is a partner, not a crutch. We want to empower our people, not create a dependency on an AI. Fostering critical thinking and encouraging smart handoffs to the People team is a design challenge we are still tackling.

A huge thank you to Aude Vantyghem, Sebastien Delarquier, Juliette Raimbault, Camille Leflour, Paul Sauveplane, and Lucas Picot for bringing this project to life!