Salesforce Trailhead

Trailhead is Salesforce's free online learning platform, taking millions of professionals from beginner to certified expert. It was built for the climb, and it loses the experts the moment they reach the top.

Over 10 months, our team of five partnered with the Salesforce Trailhead product team to research, design, and build for the platform's most experienced users: the power users who built their careers on Trailhead and then quietly stopped coming back.

Salesforce's most valuable power users were disengaging, and for a company built on community and advocacy, that is an expensive thing to lose

Salesforce's ecosystem runs on its community: experienced users who answer questions, mentor newcomers, and champion the platform. Those power users were the engine of it, and they were showing up less and contributing less.

They weren't unhappy users churning, they were quietly drifting. So the problem we set out to solve was re-engagement, not retention: winning back the attention of the platform's most valuable advocates.

I ran research operations end to end, reaching 650+ Salesforce professionals across the US, India, Canada, Europe, and the Middle East, qualifying 72 as power users, and interviewing 12 across two rounds

Round one focused on opinion: how people felt about Trailhead and what would make it more inspiring. It gave us warm, agreeable answers and no real direction.

Round two focused on behavior: what people actually do, where they drop off, whether they use the community and why not, whether certifications even matter to their goals. That is where the real pattern surfaced.

Power users dropped off right after milestones like certification. And between those moments, they took their questions to ChatGPT and YouTube for a faster answer, despite a community on the platform built for exactly those questions.

Of everything we explored, two ideas stood out, so we skipped wireframes and vibe-coded them into working proofs of concept

We explored a wide range of directions for bringing power users back: community engagement, status, skill visibility. The two strongest earned a quick build, enough to try each idea for ourselves. Neither held up, and killing them shaped the two features we shipped.

We almost gave power users an AI version of themselves to run their community presence. We killed it to protect Trailhead's trust.

The idea was status plus leverage: power users become group leaders, and an AI trained on their posts answers community questions on their behalf.

We took it down two paths, and both broke. Fully automated meant undisclosed AI speaking as a real person, which damages trust. Disclosed or approve-first clashed with Trailhead's warm, human brand and the real anti-AI anxiety among users, and the approval version stopped saving anyone time, which was the entire point. Underneath all of it, our research was clear that these users didn't want to engage with the community anyway.

We designed a public skill profile meant to pressure power users into staying current, then killed it because it only worked if employers bought in

The idea borrowed from GitHub's contribution graph. Make someone's skill currency public, and the visibility itself pressures them to keep it up. A shareable Trailhead profile would show certifications, badges, and skill levels tied to how current your learning is, replacing the scattered cert posts people drop on LinkedIn. Keep it public, and people come back to keep it current.

We killed it for two reasons. First, it only works if the outside world treats the profile as a credential, and that adoption was entirely outside our control. Second, even if it caught on, it would drive engagement for the sake of engagement, not the meaningful, lasting reason to return we were after.

We bet that plain-language answers drawn from Salesforce's own data would beat any general AI tool, and shipped two connected experiences to prove it

ChatGPT runs on public data. Slack answers depend on whoever happens to reply. We bet Salesforce's proprietary knowledge could beat both. Access to that data, and to Agentforce, was never on the table, so we built on OpenAI with curated Salesforce documentation and focused on proving the experience.

That bet became two features: Solve catches power users at the moment they would otherwise open ChatGPT, and a proactive system that resurfaces what's relevant next so returning is worth it. Neither works alone.

Solve answers a real Salesforce problem in plain language, on the spot, calibrated to your expertise and backed by sources you can check

How it works:

• Ask a problem in plain language and get a specific, actionable answer that cites Salesforce documentation.
• Set the depth to Beginner, Intermediate, or Advanced, so you are never over-explained or underserved.
• When Solve can't fully help, it drafts a ready-to-post Trailblazer Community question from your conversation, so you get an answer either way.

That last one mattered to us. We designed for the moment the AI can't help, not just the moment it can.

Without being asked, the system surfaces what's relevant to each expert: skill gaps against the market, follow-ups from their questions, and renewals before they lapse

How it works:

• Before a certification lapses, it flags the renewal on its own, closing a notification gap users complained about.
• Unprompted, it measures a user's skill profile against what the market is hiring for and builds a path to close the gap it finds.
• It reads the questions they asked Solve and recommends what to learn next, so the follow-up finds them.

None of this waits for a request. The platform does the noticing and brings the next step to the user.

Figma MCP took each component straight into Cursor and Claude Code, where we wired it to the OpenAI API and shipped Solve as a working app, not a mockup

After designing each component in Figma against Trailhead's design system, we built it one piece at a time, no handoff, no rebuild from a spec. The result was a functional product: a React front end, an Express and SQLite back end, OpenAI for responses, voice input, expertise calibration, and community-post drafting, all running.

Taking a concept from design to working software ourselves is what let us test it like a real product instead of a slideshow.

Real power users were hard to get hold of, so we built 12 AI agents from our research data to catch the obvious problems before spending anyone's real time

We turned our research into 12 AI agents and ran them through the flows to surface usability problems first. By the time real users arrived, the rough edges were already smoothed, and their time went to what actually needed a person.

We tested whether Solve could replace the reflex to open ChatGPT. Given the same Salesforce question, 6 of 8 power users preferred Solve.

Each participant took a question about the Salesforce product they work with, asked it to both ChatGPT and Solve, and compared.

The prototype pulled power users back into the content, lifting session duration 35% and cutting drop-off 22%

This was the half of the problem we most needed to prove: that experts who used to leave after a milestone would stay.

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

For most of my career, code was something I designed around. I had never prototyped with it and never felt at home in it. This project left me no room to avoid it, and using AI coding agents to build and ship a working product turned that limitation into something I can now do.