Work/Trailhead Solve
Case study ( 01 ) · Trailhead
An AI-powered re-engagement strategy Trailhead power users preferred over ChatGPT.
Improving power-user retention with an agentic AI RAG bot: Trailhead Solve, a dashboard redesign and a RAG assistant.
Overview
Built for the climb, and it loses experts at the top.
Trailhead is Salesforce's free learning platform, taking millions of professionals from beginner to certified expert. The Trailhead product team brought us one question: why are the platform's most experienced users showing up less?
What we found changed the brief: power users weren't unhappy with Trailhead. They were taking their hardest questions to ChatGPT. We ran it like a product bet: kill ideas fast, ship working software, test with real experts. This was a team of five; the research, prototype and media are shared work.
Problem: Experts took their questions elsewhere.
Salesforce's most experienced users were taking their questions to ChatGPT and Slack instead of Trailhead's own community, and between milestones, nothing gave them a reason to come back.
Solution: Answers, plus prompts.
Plain-language answers to real Salesforce problems inside Trailhead, paired with proactive prompts that surface what's relevant next without being asked.
Outcome: 6/8
Given the same Salesforce question, 6 of 8 power users preferred Solve over ChatGPT.
Insight
The certification path ended, and ChatGPT caught them on the way out.
Salesforce's ecosystem runs on power users who answer questions, mentor newcomers and champion the platform. They weren't churning in anger. They were quietly drifting, and nothing on the platform noticed.
We reached 650+ professionals, qualified 72 power users and interviewed 12. A first round on opinion gave warm, useless answers, so we switched to behavior: where people drop off and where they go instead. Engagement fell exactly where certification ended, and between milestones experts took their questions to ChatGPT.
Findings: Why experts left
64% post-milestone
Disengaged after major milestones like certifications.
45% workarounds
Supplemented Trailhead with ChatGPT, YouTube and Slack.
45% depth
Found the content too shallow for their experience.
36% personalization
Were unhappy with the existing personalization.
Reframe: Re-engagement, not retention
Question
How might Trailhead become the first place an expert goes when they're stuck, not only when they're studying?
Success metric
Preference over ChatGPT on a real question, session duration, and drop-off after milestones.
Guardrail
Protect Trailhead's warm, human brand. No AI speaking on behalf of a real person.
What we killed
Two ideas we built on purpose, so the real bet had room.
We explored community status, leverage and skill visibility, then skipped wireframes and vibe-coded the two strongest into working proofs of concept. Trying them ourselves was faster than debating them.
Neither held up. Killing them told us what not to build, and shaped the two features we shipped.
Two proofs of concept
An AI version of themselves
An AI trained on their posts answers community questions on their behalf.
A public skill profile
A shareable Trailhead profile showing certifications, badges, and skill levels.
Killed 01: An AI version of themselves
The idea
Power users lead groups while an AI trained on their posts answers community questions on their behalf.
Fully automated
Undisclosed AI speaking as a real person damages trust.
Disclosed or approve-first
Clashed with Trailhead's warm, human brand, and approving every reply stopped saving anyone time.
Underneath
Research was clear these users didn't want to engage with the community anyway.
Killed 02: A public skill profile
The idea
Borrowed from GitHub's contribution graph: make skill currency public, and visibility pressures people to keep it current.
Adoption was out of our hands
It only works if employers treat the profile as a credential.
Engagement for its own sake
Even if it caught on, it gave no meaningful, lasting reason to return.
What we shipped
Catch the ChatGPT moment, then give experts a reason to come back.
We bet that plain-language answers drawn from Salesforce's own knowledge would beat any general AI. Proprietary data and Agentforce were never on the table, so we built on OpenAI with curated Salesforce documentation and focused on proving the experience.
The bet became two connected features. Solve catches power users at the moment they would otherwise open ChatGPT; proactive prompts make returning worth it. Neither works alone.
01 · Solve: A plain-language answer, calibrated to you
Ask in plain language
A specific, actionable answer that cites Salesforce documentation.
Set the depth
Beginner, Intermediate or Advanced, so you are never over-explained or underserved.
Never a dead end
When Solve can't fully help, it drafts a ready-to-post Trailblazer Community question from your conversation.
02 · Proactive prompts: The platform does the noticing
Renewals
Flags a certification renewal before it lapses, closing a notification gap users complained about.
Skill gaps
Measures a skill profile against what the market is hiring for and builds a path to close the gap.
Follow-ups
Reads the questions asked in Solve and recommends what to learn next.
Build & test
Working software and synthetic users before we spent real people's time.
Each component was designed in Figma against Trailhead's design system, then taken through Figma MCP into Cursor and Claude Code and wired to OpenAI. No handoff, no rebuild from a spec: a working app, so we could test it like a product instead of a slideshow.
Real power users were scarce, so we turned our research into 12 AI agents and ran them through the flows first. By the time 8 real power users arrived, the obvious rough edges were gone and their time went to what needed a person.
What was running
React front end, Express and SQLite back end, OpenAI for responses, voice input, expertise calibration and community-post drafting.
12 agents, then 8 users
Synthetic users built from research data caught usability problems before real sessions.
Impact & learning
Preferred over ChatGPT, with honest limits on what we proved.
Each participant asked the same question about the Salesforce product they work with to both ChatGPT and Solve. 6 of 8 preferred Solve.
The prototype also pulled experts back into the content, lifting session duration 35% and cutting drop-off after milestones 22%, the half of the problem we most needed to prove.
The honest limit: curated public documentation, not Salesforce's proprietary data or Agentforce, and 8 power users. The next step is a larger test on real support data, where the bet can be proven at scale.
6/8 preferred Solve over ChatGPT
Same Salesforce question, side by side.
4.5/5 confidence
In Solve's answers.
+35% session duration
Power users stayed in the content.
−22% drop-off after milestones
Experts who used to leave after a milestone stayed.
“I had never been comfortable in code, and had never prototyped with it, so teaching myself to ship this as working software is the barrier I'm proudest of clearing.”