Bo3 Lab connected factories with suppliers of Industry 4.0 solutions, from first match to a pilot. I owned the MVP from onboarding to PoC initiation, then cut the survey the team believed in, because the data showed it was where buyers quit.
Buyers quit halfway through a long survey, wouldn't share documents before a pilot, and drowned in detailed solution pages. Many who finished saw their matches and never came back.
More questions didn't buy better matches. It bought fewer people reaching a match at all. And a match list is a dead end unless it pulls you toward a pilot.
AI-guided onboarding that asks only what your use case needs, top-5 matches with a match % and editable keywords, documents only after PoC initiation, and nudges back toward a pilot.
+27% onboarding, +70% engagement and +10% retention in 2024.
Industry 4.0 is manufacturing getting smarter and more connected: machines that predict their own failures, production that adjusts to demand, less downtime. The catch for buyers is that these systems have to prove themselves on their floor first. A plant tests a failure-prediction tool on two machines for a month before it commits to a rollout.
Bo3 Lab was a two-sided marketplace built around that reality. Suppliers list solutions, buyers describe a problem, and the platform carries both sides through discovery, AI matching, starting a PoC, the pilot, feedback and purchase. I joined to build the first version.
BuyerSigns up, describes the use case, answers onboarding
SupplierLists solutions on a supplier profile
Quit halfway through a long survey that also wanted documents
Describe it in a sentence; answer only what's relevant
BuyerReviews matched solutions
SupplierSurfaces when the fit is real
Saw a wall of detailed matches, left, never came back
Top 5 with a match %; refine keywords any time
BuyerPicks a solution to pilot
SupplierReceives the request
Rarely reached
Documents requested here, by email. Nudges lead back here.
BuyerTests on a few machines
SupplierRuns the pilot on site
BuyerJudges results against the promise
SupplierGets structured feedback
BuyerCommits to a rollout
SupplierCloses the deal
My scope MVP · onboarding → PoC initiation
GA4 funnels showed the drop plainly: people started onboarding and left midway. Walking the flow end to end, I found three things pushing them out.
Lots of required questions that didn't matter for the match. Every buyer answered everything, whatever their problem.
Buyers were wary of sharing sensitive plant details before they'd even chosen a supplier to pilot with.
Every solution came with full detail. Evaluating them one by one became its own long project.
One long form with many required fields, most irrelevant to any one buyer.Describe the problem in a sentence; a text classifier picks the questions that matter for it.
→ Shorter path to a first matchUploads sat inside onboarding, before buyers trusted anyone with them.Upload left onboarding. Initiating a PoC triggers an email asking for what's needed.
→ Fewer reasons to quit midwayDetailed listings that each needed their own evaluation.Top 5 ranked by similarity, a match % on each, and keywords you can edit.
→ Refine, don't re-readLonger surveys give the matcher richer data, so matches get better.
Buyers dropped off halfway. A rich profile is worth nothing if it never gets finished.
Ask for a short use-case description, and let a text classifier choose only the questions that matter for it.
Thinner data per buyer. The hedge was a keyword feedback loop: if the matcher guessed wrong, the buyer could correct it in a tap.
“This went against our initial belief that ‘more questions = better data’, but the results proved it right.”
AI-guided onboarding. Buyers describe their use case in a sentence. An ML text classifier reads it and selects only the relevant questions, so a vision buyer never answers maintenance questions.
Top 5, with a reason. NLP similarity against the solution database ranks the matches. Buyers see five, each with a match % (“87% match”), plus the keywords behind it, which they can edit to refine.
Documents after commitment. Upload left onboarding entirely. The moment a buyer initiates a PoC, an email requests what the pilot needs.
Describe a use case, or pick an example, to see matches.
Onboarding improved, but plenty of buyers onboarded, looked at their matches and left. Nothing pulled them back: the first visit was also the last step.
Two changes. Matching became an ongoing tool: come back, tweak the use case or keywords, get refreshed recommendations. And because PoC initiation was the real moment of value, follow-ups and email nudges pointed buyers at starting one instead of just browsing. Retention rose about 10% over the next few cohorts.
For the supply side I launched the supplier profile ecosystem: an internal CRM dashboard where suppliers list and track products, and we see everything in one place. I worked with design on wireframes and flows, and wrote the PRDs, including the REST APIs connecting it to our CRM. The team built it in React, Node.js and PostgreSQL.
Success was concrete: how many suppliers successfully listed products, listing completion rates, and fewer manual steps between sign-up and live.
Instrumentation — GTM, GA4 and heatmaps on the product; Airtable + Zapier to track CRM/CMS journeys and email engagement.
Why end onboarding at the matchThat's the moment a buyer experiences what the product does: the bridge between “I just signed up” and “now I get why this is useful.” A finished survey with no match isn't an onboarded user.
The team's original fear. The classifier kept the questions that mattered for that use case, and editable keywords let buyers correct the matcher themselves.
Moving upload out of onboarding couldn't mean losing it. Initiating a PoC triggered the request, at the moment the buyer had a reason to share.
B2B buyers won't pilot off a ranking they can't read. Every match showed a %, and the keywords behind it were visible and editable.
Close to a release, a stakeholder brought a new feature request. Instead of pushing tickets, I checked feasibility with the team first, then took the trade-offs to stakeholders. We agreed to move it to the next sprint. It's not about saying yes or no immediately.
Every onboarding question costs you users. The best data is the data people actually finish giving you.
“More questions = better data” was never tested. It was simply believed. Now I write beliefs down as beliefs, so when one breaks we know which, and the data gets the final word.