Steve Blank's post-MVP startup; Aaron Levie's deployment layer
Steve Blank's rebuilt startup class and Aaron Levie's deployment layer, side by side: what proves an AI product is worth building once a working demo costs a weekend, and who does the work of getting it into a business.
Steve Blank, whose Lean LaunchPad class has run for fifteen years and is taught at Stanford 1, closed a four-part essay series on September 30 2 with a diagnosis the series had been building toward: the demo a team can now generate with AI no longer proves that the team can build or that anyone wants what it built 3. Aaron Levie, the co-founder and chief executive of Box, wrote on September 30 that the value in AI is collecting in the layer that puts a model to work inside a company 4, and on October 4 that agent adoption outside coding has barely begun 5.
Both pieces answer a question a founder faces this week, now that a capable model is something any team can rent. What counts as proof that you have found something worth building? And who does the work of turning a model into a finished job?
Steve Blank: a working product stopped being evidence
Blank designed the Lean LaunchPad class in 2011 around a premise that was heretical at the time: on day one, nearly every important fact in a business is a hypothesis 3. The class taught founders to search for a business model, and it gave the search a physical artifact. A working minimum viable product carried two jobs at once: it showed that a team had the technical competence to build something, and it produced the evidence that customers wanted it 3.
AI took both jobs away. Blank states the change in one sentence:
"With teams capable of vibe coding a product, minimum viable products (MVPs) did not represent proof of technical competence. MVPs were also no longer reliable evidence of customer discovery, critical thinking, hypothesis testing, product/market fit or even commitment." 3
When a polished product costs a weekend, the artifact stops separating a team that understood a customer from one that did not. Blank renamed what AI builds on day one the Initial Untested Product 2. Then he did the harder thing. "You can't know your future if you don't understand your past," he wrote, and went back to the class's 2011 design and its ten founding elements; two of the ten, the two jobs the MVP had carried, turned out to be obsolete 3.
He also looked outside the classroom. In the September 28 post he named the four things that had moved since 2011 and that decide whether a founder's search works at all: how much venture capital a startup can raise and from whom; how products get built; what building costs in time, team size and capital; and how customers evaluate, buy, deploy and use software 1. His summary of the shift runs to one line: "Bottlenecks for adoption and scale still exist, they've just moved elsewhere." 1
The redesign in the September 30 post is where the transferable material sits. Blank split the class into two halves — weeks one to four for thinking deeply about stakeholders, problems, hypotheses and connections, weeks five to nine for acting quickly to land design partners — and then replaced five things 2.
- The artifact. The MVP becomes the Initial Untested Product, and teams test one value-proposition hypothesis at a time with a "snippet": "a small portion of the IUP used to test one value-proposition hypothesis without showing the full product." 2
- The fit target. Product/market fit becomes Product-Outcome Fit, defined per category: outcome and agent fit for enterprise software, mass creation and testing for consumer products, digital twins for hardware, simulated clinical endpoints for life sciences 2.
- The other boxes. A Minimal Tested Canvas replaces the MVP as the instrument for the parts of the business model beyond the product: "the smallest set of experiments that meaningfully tests the non-product parts of the Business Model Canvas." 2
- The evidence. Customer validation moves to a design partner who commits staff time, data, integration work, permissions and test environments, and the class's objective becomes a signed partner rather than a demo 2.
- The gates. Because AI removed the pacing that building an MVP used to impose, teams now face explicit verify-or-pivot gates, "so a polished product can't substitute for evidence." 2
Blank added a second set of rules aimed at what AI does to a team's judgment 2. Teams run synthetic interviews and grade them against live ones: "the live interviews grade the synthetic ones, never the reverse." Any team member should be able to explain the customer, the workflow, the evidence, the economics and the key contradictions without opening a laptop, and AI-generated work has to be labelled as the model's or the team's — a test Blank calls "learning debt". Three items of canvas economics that did not exist in 2011 now do: what running the model costs the startup, what it costs the customer, and which actions a customer will let an agent take at what level of reliability and reversibility 2.
Two of those additions carry the strategic weight. Blank tells teams to treat model dependence as a business-model risk rather than an engineering choice, because "a startup should assume that its preferred model, pricing, latency, terms, or relative quality will change" 2. And he argues that in many markets the evidence a customer needs before deploying an agent is itself part of product/market fit 2.
For a founder, the redesign reads as a set of replacements for the thing builders reach for first 2. When AI can produce the demo, the proof has to come from elsewhere: a hypothesis narrow enough to test with a piece of the product, a fit target specific to the category, and a partner who risks something real before the product is finished.
Aaron Levie: the value is stuck in the deployment layer
Levie's September 30 post opens with a claim about where the next businesses are: "There's a huge opportunity right now in being the deployment layer for AI into the economy." 4
Loading content card…
His reason is that an agent is a different object from a piece of software. Software deployment ended when the customer's team learned to run the tool; an agent delivers work output inside a process, which keeps the vendor involved in whether the work gets done. "You're delivering actual work augmentation to the organization, which has a completely different set of complexities associated with it," he writes 4.
That difference produces a long list of jobs, which Levie names: moving legacy systems to the cloud, reorganising data and access, connecting software to agents in new ways, reengineering workflows, settling the human-in-the-loop decisions, generating and maintaining evaluations, and updating the whole system as new models arrive. He adds that the list "may even be longer" 4.
Who does that work inside a large company? On October 1 Levie pointed to internal forward-deployed engineers, moved into departments to connect AI to the workflows underneath 6. The role needs, in his description, "a high level of technical expertise, an understanding of AI, and the ability to understand the workflows and processes that the enterprise is trying to automate." He quotes Jake Stauch's name for the job — the Automation Engineer — and Stauch's note that "nobody has ten years of experience doing this yet, because ten years ago the tools that made it possible didn't exist (it didn't even exist two years ago)" 7.
On October 4 Levie described how far the transition has actually gone: "AI agent adoption is still very bimodal right now. You have coding and coding adjacent tasks which have taken off, and then everything else." 5 The shape follows from the same list, because every workflow outside coding still needs new practices for accountability and liability when an agent is in the mix, plus governance, compliance and security changes 5. His multiplier for what that implies: "you can basically expect 100X more agent adoption from what we've seen so far." 5
In a second post that day Levie read the same shift through jobs. AI "requires significant technical work and surrounding services to deploy into the economy," he wrote, which is creating applied-AI engineering roles, forward-deployed engineers inside companies, and a new kind of services firm 8. He quotes figures from The Kobeissi Letter, which attributes them to LinkedIn estimates: AI-linked roles have accounted for more than 750,000 new jobs in the United States since 2023, led by 282,000 data annotators, 117,000 data-centre jobs and 105,000 AI engineers, at a median salary near $180,000 against $80,000 across all jobs 9. Those are one platform's estimates of a category it defines itself, and Levie's own reading stays narrow: "It's a lot easier to picture what AI can replace vs. what it creates until it starts happening." 8
For a founder, the two posts mark out where the unmet work sits. The layer that turns a model into finished work is large enough to support services firms alongside products, and Levie expects the openings to sort by industry, by company size and by problem inside a company 4. The skills it needs are scarce because the job is new, and his advice to software people moving into AI is to go deep on that combination 6.
Four decisions this week
- Demote the demo. Treat what AI built as an untested product, and write down the single value-proposition hypothesis it exists to test. Blank's class now grades teams on the quality of their customer outreach rather than on proving the snippet right 2.
- Set the fit target for your category. Name what outcome fit means in your category — agent behaviour in a workflow for enterprise, repeated testing for consumer, digital twins for hardware, simulated endpoints for life sciences — and use it in place of a generic version of product/market fit 2.
- Sign a design partner and ask for real commitments. Staff time, data, integration work, permissions and a changed workflow are the evidence; a partner who offers only enthusiasm is still a prospect 2.
- Write the deployment list for one customer. Take Levie's list — data access, integration, workflow, human-in-the-loop, evals, updates — and mark which items your product carries and which it leaves to the buyer 4.
A decision to run now
Take the last thing your team built with AI and answer three questions about it on one page: which customer hypothesis it exists to test, what a customer would have to commit for their use of it to count as evidence, and which parts of their workflow your product leaves for them to run 24.
Blank and Levie work at opposite ends of a company's life, the search for a business model and the delivery of one. Both hand a founder the same task this week: treat the demo as the beginning, find one customer who will commit something scarce, and write down the work that still has to happen around your product before it does anything for them.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.