Module 10: Startup Thinking with AI — Problems, MVPs, Business Model Canvas & Lean Methodology | College AI Track

Learning Objectives

By the end of this module, you will be able to:

Frame ideas as falsifiable hypotheses about pain, behavior, and timing.
Apply BMC and Lean Canvas with evidence per block, not vibes.
Design MVPs that test riskiest assumptions with clear metrics and kill/pivot rules.
Assess when AI is load-bearing vs ornamental—TCO including eval and safety.
Pitch with problem-insight-demo-traction-ask aligned to Module 03 citation norms and Module 09 eval habits.

Starting from “we have GPT” mirrors HARKing in science. Start from stakeholder evidence: interviews, tickets, public complaints. Hypothesis template: “We believe [segment] experiences [pain] when [context], causing [measurable consequence].”

10.2 Jobs-to-be-done

Users “hire” products for progress—time saved, anxiety reduced—not for “using AI.”

10.3 Business Model Canvas — evidence standard

Block	Ask
Segments	Who adopts? Subsegments? Interview counts?
Value prop	vs next best alternative including spreadsheets
Channels	How they discover you; trust/compliance effects
Revenue	Even $0 projects: who might pay and why
Key resources	Data rights, APIs, compute
Key activities	RAG maintenance, eval, support
Partners	IT, disability services, faculty sponsors
Costs	API $/user, labeling, legal

10.4 Lean Canvas

Emphasizes problem, solution, metrics, unfair advantage—state the riskiest assumption explicitly.

10.5 Build–measure–learn

Assumption → instrument (landing page, concierge, prototype) → metric → threshold → persevere/pivot/kill—document negative results.

10.6 MVP types

Concierge; wizard-of-oz (ethics: disclose in research); single-scenario slice; landing smoke test.

10.7 When AI is differential

High-variance language/perception tasks; personalization over permitted data—with guardrails. Weak when stakes are catastrophic without pros, data is tiny, or audit needs exceed black-box tolerance.

10.8 Unit economics (conceptual)

Variable cost ≈ tokens × sessions × price; fixed = curation, eval, compliance. Compare to substitutes (tutor hourly × time saved).

10.9 Pitch arc

Hook vignette → problem + metrics → insight → solution + AI boundary → why now → market (assumptions stated) → model → traction → team → ask.

10.10 Ethics

Map stakeholders/harms; avoid dark patterns and data hoarding. Competitive analysis like a mini systematic review—AI drafts tables, you verify websites.

10.11 Traction vs vanity

Useful: retention, task success, human-evaluated quality on golden sets—not raw message counts.

Code and Computational Examples

Key Example: Transparent assumptions beat trillion-dollar TAM hand-waving—plug your campus numbers from real sources.

python

[object Object], dataclasses ,[object Object], dataclass

,[object Object],
,[object Object], ,[object Object],:
    segment_population: ,[object Object],
    adoption_rate: ,[object Object],
    annual_price_usd: ,[object Object],
    citation_for_population: ,[object Object],

,[object Object], ,[object Object],(,[object Object],) -> ,[object Object],:
    ,[object Object], m.segment_population * m.adoption_rate * m.annual_price_usd

ugrad = MarketAssumptions(
    segment_population=,[object Object],,
    adoption_rate=,[object Object],,
    annual_price_usd=,[object Object],,
    citation_for_population=,[object Object],,
)
,[object Object],(,[object Object],)

Tables instead of code: Fill BMC and Lean Canvas in markdown; use YAML experiment pre-reg (hypothesis, metric, stopping rule); JSON pitch outline with ai_boundary and citation fields.

Weak hypothesis	Stronger sketch
“AI affects learning.”	“Among course X students, weekly draft assistance correlates with rubric scores controlling for prior GPA.”

Try This! Interview two people for five minutes each; write one problem sentence that includes a number or frequency they actually said.

Fun Fact: Fabricated LOIs and user counts are integrity violations—zero users + honest next experiment beats fake traction.

Practice Exercises

Segmentation matrix with evidence; BMC v1 → v2 after three conversations; AI necessity test (what breaks without AI?); 12-row unit economics sensitivity; competitive matrix with sourced footnotes; five-minute pitch storyboard; ethics pre-mortem; MVP one-page pre-reg.

Mini-Project

Lean + BMC + evidence pack: Problem brief with three anonymized quotes; both canvases with weakest three blocks flagged; MVP spec with kill rule; stakeholder map; pitch outline JSON; 300-word reflection on surprises.

Key Takeaways

Evidence beats eloquence in early-stage work.
Hypotheses parallel research hypotheses; MVPs parallel pilots.
AI accelerates drafting; you own facts, users, and ethics.

Key Takeaway

Anchor every canvas block in interviews, proxies, or citations—“unknown—test plan” is allowed.
Make the riskiest assumption one sharp sentence with a timeboxed experiment.
AI necessity test: if removing AI does not break value, narrow the scenario until it does—or drop the gimmick.
Unit economics and ops (RAG refresh, support) belong in the story early.
Honest negatives are learning velocity; vanity metrics are debt.

References and Cross-Modules

Ries, Lean Startup; Osterwalder & Pigneur, Business Model Generation; Maurya, Running Lean; Fitzpatrick, The Mom Test; Modules 01, 09, 11, 12.

Module 10: Startup Thinking with AI — Problems, MVPs, Business Model Canvas & Lean Methodology