Lean Startup and AI: testing assumptions, metered funding, and building small

Executive overview

Most AI projects fail not because the technology doesn't work, but because teams skip validating what customers actually want before building. The Lean Startup framework — identify leap-of-faith assumptions, run minimum viable experiments, fund incrementally — applies directly to AI development, where costs are real and the technology is genuinely uncertain.

Generative AI changes the economics of company-building: cost of goods sold is back, open-source is closing the gap fast, and most firms are still copying the SaaS playbook onto a fundamentally different stack.

The antidote to wasted AI investment is treating every project as a portfolio of small, credibly-funded bets with explicit success criteria and automatic cancellation.

Identifying the right assumptions to test

• A leap-of-faith assumption is any input that, if it went to zero, would destroy the business model — find it in your spreadsheet's most critical cells.
• Metrics are people: translate each number into the human behaviour it represents (referrals, repeat use, willingness to pay).
• Test customer assumptions — will they adopt it? Does it threaten their job? — before testing technical feasibility.
• AI stretches what's imaginable, but "technically possible" does not mean "desirable" or "profitable".
• Cross-functional teams need a data specialist and AI specialist from day one, not later — feasibility is no longer obvious.
• The Venn diagram of "what we can build" and "what customers want" has no guaranteed overlap.
• Large companies pre-constrain the solution space; explicitly open it up, then test.

Minimum viable products in an AI context

• An MVP is the least work required to discover the truth of a situation — not cutting corners, but cutting scope.
• Reducing scope often increases quality: a single perfect unit reveals more than a mediocre mass rollout.
• Concierge MVP: a human works behind the scenes to ensure output quality while you validate that customers value the result.
• A whiz-bang demo is fine as long as there's a sign-up page — can the customer buy it right now?
• Wrapping a foundation model (OpenAI, Anthropic, etc.) is acceptable for validation; it's dubious as a permanent strategy.
• Blind taste tests against off-the-shelf models are fast, cheap, and often reveal that expensive specialised tools are outperformed by Claude or GPT out of the box.
• Offer the product inside a course: customers pay, use the tool in a structured context, and their post-course behaviour tells you actual retention value.

The dangers of upfront funding

• ROI calculations in AI proposals are typically made-up number ÷ made-up number; results almost always come in far worse than the plan.
• Previous-era software cost overruns were bounded by headcount; AI adds runaway operational costs on top.
• Annual budget cycles treat funding as a subscription — committing a million dollars means a million a year, growing indefinitely.
• Rational middle managers delay shipping: delay preserves budget, shipping and failing cancels it.
• AI makes the illusion of progress cheap — new demos, new research papers, new capabilities — without any real customer traction.
• Companies fund one big bet to signal conviction; this conflates commitment with concentration risk.

Metered funding as the antidote

• Metered funding gives a team a fixed resource pool (time or money) and explicit criteria for unlocking more — default answer is no unless results are shown.
• Scarcity creates necessity: constrained teams are dramatically more creative and productive than teams with open-ended budgets.
• Y Combinator runs 10–12 weeks; startup weekends produce remarkable results in 48 hours. Most corporate timelines are unjustifiably long.
• The executive advantage: a portfolio of small bets lets you cover more surface area. Ten winners require far more starting bets than most companies budget for (even at an optimistic 50% pass rate per stage).
• Double down quickly when something works — waiting a month to approve the follow-on experiment hands the opportunity to a competitor.
• Pre-agree at funding time that a successful first experiment automatically triggers the second; remove the political decision point.
• When teams see the first wave of projects cancelled on schedule, every subsequent team immediately asks "what do we need to demonstrate?" — accountability becomes self-reinforcing.

What executives get wrong about accountability

• Employees unconsciously warp their behaviour to match what they believe executives actually reward — not what executives say they want.
• Telling people to care about results has almost no effect unless the funding system makes results the only path forward.
• Executives who complain that teams don't care about success criteria usually set up incentive structures that reward delay and punish shipping.
• Once metered funding is credible, teams self-organise around the criteria without being asked.

Three insights for AI builders

• Understand where value will actually accrue. The AI stack is not SaaS. Cost of goods sold is real. Open-source models are closing the gap fast. Michael Porter's Five Forces apply — revisit them.
• Most AI-native companies still scale like SaaS companies — large human sales forces, standard pricing. The contrarian bet: a 12-person generalist team augmented by AI can outperform traditional org structures at a fraction of the cost.
• Question every vendor assumption. Cloud hosting, security audits, sys-admin — AI can substitute for many of these at near-zero marginal cost. First-principles thinking about the stack reveals that many assumed dependencies are no longer necessary.