Your MVP looks finished. Your customer discovery isn't.

What Stanford Lean LaunchPad 2026 is really testing this season — and what mentor rooms in Hong Kong keep seeing.

Most founder decks I see this month open with a polished prototype. The demo loads. The canvas looks complete. Someone has already asked ChatGPT to tidy the interview script. From the outside, it looks like progress.

From the inside of a mentor session, I often ask a quieter question: which customer changed your mind this week?

That gap — between a finished-looking MVP and unfinished discovery — is the story of this season. And the expensive part is rarely “we never interviewed.” It is we interviewed early… then stopped.

A semester where building got cheap

Steve Blank’s Stanford Lean LaunchPad Spring 2026 (the course’s 16th year) put eight teams through 978 interviews. For the first time, every team used AI for canvases, MVPs, interview questions, and summaries. Build velocity jumped. Learning velocity did not keep up. Polished MVPs started to feel like success, so pivots arrived late. Blank’s line is worth sitting with: it wasn’t the AI hallucinating — it was the teams.

Poets&Quants’ write-up of his rethink adds a second warning. Synthetic “customer interviews” can invent feedback that sounds right. In one comparison, roughly half of a synthetic set hallucinated plausible answers. Ground truth is still outside the building. Course design is drifting toward design partners, instrumented prototypes, and clearer learning criteria — not prettier demos.

If you mentor in Hong Kong, none of this should feel foreign. We have been watching the same pattern in LLP rooms for years. AI just poured accelerant on it.

Four ways discovery dies after a good start

These are the failure modes I see when teams do talk to customers — then quietly quit the habit:

1. Early interviews, then radio silence. Week 1–2: ten calls, a tidy problem statement, a green “validated” box on the canvas. Week 6: zero new conversations. The market moved; the team didn’t. Discovery is not a phase you complete — it is a cadence you keep until the evidence stops changing (and often after).
2. MVP polish as a proxy for learning. The UI gets sharper every sprint. Screenshots replace interview notes in the mentor update. Stanford 2026’s teaching takeaway maps here: when the product looks finished, teams confuse a deliverable with stakeholder understanding — and pivot late because polish feels like product/market fit.
3. Synthetic interviews that never leave the laptop. AI-drafted personas and chatbot “customers” produce fluent, confident answers — including ones nobody ever said. Poets&Quants’ reporting on Blank’s classroom experiment is blunt: about half of a synthetic interview set hallucinated plausible feedback. Use AI to prep questions and summarise real calls. Do not outsource the customer.
4. Late pivots after sunk demo pride. The deck is beautiful. Investors smiled. Classmates clapped. Only then does a design partner say the workflow is wrong. Changing direction after a polished demo costs more ego and calendar than changing after interview #3. Blank’s 2026 cohort felt that lag: speed to build outran speed to learn.

Same root cause every time: customer discovery was treated as a checkbox, not a continuous experiment.

What I listen for in mentor hours

I have spent a long time with Lean LaunchPad cohorts across PolyU, HKU SPACE, HKSTP, and related DeepTech tracks in Hong Kong. Clever prototypes are rarely the scarce resource. What teams under-price:

• the hypothesis this week’s interviews were meant to falsify
• a design partner who will still take the call after Demo Day energy fades
• a definition of “done” that includes learning notes, not only features shipped
• the courage to kill a polished demo that nobody asked for
• a standing interview habit after the first “validation” sprint

Better features matter. They are also the easiest story to tell investors and classmates. The scarcer story is: we know how evidence accumulates in Hong Kong or Asia — and we priced the interview hours every week, not only in week one.

That is why 978 interviews across eight Stanford teams can outweigh eight shiny MVPs. Learning still has to outrun the demo.

Four questions before the next sprint

Use these in your next mentor check-in. Answer in plain language — no slide required.

1. What claim are we testing? Problem, segment, willingness to pay, or channel — pick one.
2. Who can say no this week? Buyer, beneficiary, influencer, or channel partner.
3. Where does learning land between builds? Interview notes, kill/keep decisions, design-partner follow-ups.
4. Can you defend the evidence without the AI draft? If the canvas came from ChatGPT, can you still narrate why a real person cares?

If the room goes quiet on those four, you do not have ongoing customer discovery. You have a prototype with a story — and a closed interview calendar.

Closing

Asia will keep getting faster builds. The teams that compound will be the ones who treat listening, hypothesis tests, and design partners as core product work — the boring, repeatable kind that continues after the first ten calls.

Mentor ask for this week: how many real customer interviews did you complete — not how many features you shipped? And did you book next week’s calls before you closed the laptop?

For a Lean LaunchPad teaching enquiry, email [email protected].