Built a coaching Gem or custom GPT? Here's why employees won't use a coach they think the company can see into, and what to fix before you scale.
%2520(1).png)
If someone on your team has already built their own AI coach as a Gemini Gem or custom GPT, they've done the hard thinking. They've shown people want somewhere to talk through their work. The harder question is whether people will keep using it once it's rolled out to everyone.
This guide is for the L&D or People leader making that call. We sell AI coaching, so we've tried to be straight about when building is the right answer.
A home-built coach is a smart first step. It costs almost nothing, you can shape it around your leadership model in an afternoon, and it tells you quickly whether people will talk to an AI about work. If a few dozen people use it and come back, keep that evidence. It's the strongest part of your business case.
A prototype's first users are usually volunteers who trust the person who built it. Everyone else meets it differently. To them it's the company's bot, in the company's workspace, set up by the same function that runs performance reviews.
People already expect workplace AI to watch them. When Pew Research asked US adults how workers would feel if employers used AI to analyze how they do their jobs, about eight in 10 said workers would feel inappropriately watched. Deloitte's TrustID Index, reported in Harvard Business Review, found trust in company-provided generative AI fell 31% between May and July 2025.
That suspicion isn't paranoia, because the enterprise tools most coaching bots are built on are designed to be searchable:
Those controls exist for good reasons, like legal holds and regulated industries. They also change how people behave. A new manager who wants to say "I don't think I'm cut out for this" will say something safer, or skip the session entirely. The coach never hears the real problem, and usage quietly fades after launch.
Before you scale, ask your admin one question: if an employee used our coaching Gem for a year, what could we retrieve? Then ask how you'd explain that answer to your people. Adoption depends on both.
A Gem or GPT waits for a prompt, and the people who need coaching most are often the least sure what to ask. A coach runs a structured arc through the week, following up on last time and nudging you before the moment that matters. That's much of what separates coaching from a tips engine, as we cover in What is AI coaching?
Chat tools now remember things like the tone you prefer. Coaching needs a different kind of memory: the goal you set, the conversation you've been avoiding, and what you said you'd do by Friday.
Pasting your competency model into a Gem works for one Gem with one owner. It gets harder when the model changes and hundreds of people are using copies made months apart.
A chatbot only knows your version of events. Real development needs input from the people you work with, brought into the coaching itself.
The first version is quick to build. Within 90 days of rolling it out, you'll have your answer:
Breadth is usually the hardest gap to close. A prompt tuned for one leadership cohort now has to support a new frontline manager, a senior engineer, a salesperson before a tough renewal, and an executive preparing for the board. Each needs different context and coaching, plus consistent handling when a conversation turns sensitive. That takes a product team, and it competes for the same people who run your leadership programs.
Keep building if the group is small and already trusts the person who built it.
Buy when you want every manager and every person they lead coached in parallel, in a space they believe is theirs. By the 90-day mark, the numbers usually make that call for you.
Either way, your builder's work carries forward. The leadership language they wrote down and the questions people actually asked are exactly what a good coach needs to know about your company.
With Huckleberry, you give every manager and every person they lead a voice-first AI coach, on a team plan at $20 a seat a month. You load your handbook and competency model once, and every coach works from them. Each person's coach builds on their recent sessions and commitments, and brings in voice-based feedback from colleagues.
Privacy is built into the architecture, which is what lets people be honest. There's no admin path to session content, and HR sees aggregate themes, never transcripts. The coaching relationship belongs to the person and goes with them if they change jobs. See why generic AI assistants aren't safe for employee coaching, or read how Huckleberry works for HR leaders.
Q: Why don't employees use AI coaching tools their company builds?
A: Because they assume the company can see what they say. Enterprise AI assistants keep conversations searchable for compliance, and a coach built on them inherits that. People avoid it or keep their answers safe, so the coach never hears the real issue. Adoption follows trust, which means privacy has to be designed in and explained clearly.
Q: Is a custom GPT or Gemini Gem private from our admins?
A: Usually not. ChatGPT Enterprise workspace owners can grant access to conversation messages through the Compliance Platform, and Google Workspace admins can search and export Gemini app conversations through Vault. Confirm your settings with your admin team.
Q: What does it cost to build our own AI coach?
A: More than the prototype suggests. Add up the salaries of the people who build and maintain it, plus the AI usage costs, which climb with every conversation and every employee you add. Then compare that with a per-seat price where the engineering and privacy design are already paid for.
Your prototype showed that people want a coach. Scaling it means giving them one they trust enough to be honest with.
Book a demo to see how Huckleberry keeps coaching private while working from your company's context.
%20(1).png)
