Small, Sharp, and Hard to Kill

There has rarely been a better time to build a product, or a worse time to mistake building one for building a business. AI can take a team from idea to working software at startling speed. It can also do the same for every competitor. Capital is abundant in the headlines and scarce in the rooms where most founders pitch: the PitchBook–NVCA report recorded $412.7 billion in U.S. venture investment through June 2026, with 87.5% going to rounds of $100 million or more. Silicon Valley Bank describes record revenue growth alongside one of the hardest fundraising markets in decades. PitchBook–NVCA^[1]; SVB^[2].

That is the peculiar knife edge of this moment. The cost of making something has fallen. The cost of being chosen, trusted, and kept has not. Here are ten things we think a startup needs to stay alive on it.

1. A problem someone can name without saying “AI.” If the pitch only works after a model demo, the customer may be buying a novelty. Start with an expensive delay, a missed sale, a tedious obligation, or a result people already struggle to get. AI can be the engine; the problem must be the reason to drive. In SVB's analysis of roughly 9,000 VC-backed companies, 42% of those using AI language in their descriptions showed little descriptive evidence that AI was central to their offering. That is a warning about the label's limits, not a technical audit of those companies. SVB^[2].

2. A buyer and a path to that buyer. “Everyone who works” is an audience too wide to reach. Know who feels the pain, who can approve a purchase, and how a user finds the product. Menlo Ventures estimated that product-led adoption accounted for 27% of AI application spending in 2025, but that does not erase procurement. A startup needs a way to move from one person's enthusiasm to an organization’s budget. Menlo Ventures^[3].

3. Proof from the work itself. A beautiful demo is a beginning. Show what happens on a real task: time saved, errors avoided, revenue gained, or quality improved. Test against the old way of doing the job, not against an empty screen. Stanford’s 2026 AI Index reported 66.3% agent accuracy on OSWorld, a structured computer-use benchmark. That result is not a production failure rate, but it is a reason to ask for receipts when a product claims autonomy. Stanford HAI^[4].

4. A business that knows its own math. Count model calls, human review, implementation, support, sales, and the cost of mistakes. Price for the complete service, not the cheapest API request. Track gross margin and cash runway under slower sales and higher usage. SVB projected that 2,345 VC-backed companies were on pace to fail in 2026 (a forecast, not a year-end count). SVB^[2].

5. A place in the workflow that survives a model upgrade. In March 2026, Stanford found several leading models clustered closely on one model leaderboard. That does not mean they perform equally on every task. It does remind us that a feature built around one model's current limitation may have a short life. A startup gets harder to replace when it understands the customer’s sequence of decisions, the data and approvals involved, and the result that must make it back into the system of record. Stanford HAI^[4].

6. Freedom to change the machinery. Build so the team can compare models, swap providers where sensible, and revise prompts or tools without tearing out the whole product. The point is not to chase every release. It is to benefit when capability, cost, or reliability changes. Open standards such as Agent2Agent address communication across agent systems; they may reduce some integration work, but they do not make models or providers interchangeable by themselves. Linux Foundation^[5].

7. Trust built into the product. Know what data enters a model, who may see it, when a person must approve an action, and how an error is caught and reversed. Keep a record of what the system did. These are product choices, especially when software acts rather than only answers. NIST’s AI risk framework calls for governing, mapping, measuring, and managing risks across the system’s life. NIST^[6].

8. A fast learning loop. Ship a narrow promise, watch it in use, and ask why customers stay, leave, or work around it. Measure repeat use and the outcome delivered, not just sign-ups or generated output. Menlo's market model estimated $37 billion in enterprise generative AI spending in 2025; Stanford's 2026 report found agent deployment still in the single digits across nearly all business functions it surveyed. There is a large distance between spending, experimentation, and durable use. Menlo Ventures^[3]; Stanford HAI^[7].

9. A team that can cross boundaries. Small AI-native companies need people who can move between product, engineering, design, data, operations, and the customer’s actual job. They also need people who know when a model answer is insufficient. The scarce skill is not merely getting AI to generate more work. It is turning that work into a dependable result someone will pay for.

10. The judgment to partner. A startup does not need to own every capability its customer needs. One team may understand a hospital’s workflow; another may be excellent at secure data connections; a third may know how to evaluate model behavior. A focused partnership can offer the whole solution sooner than each team can build it alone, if ownership, economics, customer responsibility, and failure handling are explicit.

We think more partnerships between small AI-native companies may emerge for a simple reason: these firms can be concentrated pools of talent at the frontier, while many of the skills that make AI useful can travel between industries. Evaluating outputs, connecting tools, designing human approvals, managing model costs, and turning a prototype into a reliable workflow are examples. Open protocols for connecting agents and tools may reduce some technical friction between teams. The Linux Foundation said more than 150 organizations supported its Agent2Agent protocol by April 2026. That is a project-reported count of support, not a count of active integrations or proof that startup partnerships are already surging. Linux Foundation^[5].

The best pairings will still have a specific reason to exist. Two teams with the same demo and no distribution do not become stronger by shaking hands. A useful partnership combines different assets: customer access and technical depth, domain knowledge and execution capacity, or a trusted product and a missing capability. Start with one customer problem, one joint offer, and a test that both sides can measure. Then decide whether the relationship deserves to grow.

A more flexible shape of work

We think independent contractors, and agencies built around independent contractors, will become a larger part of the future of work, to a certain extent. The capabilities a company needs can change faster than a permanent job description. A startup may need a specialist in model evaluation this quarter, a designer who understands a particular customer workflow the next, and someone who can make the whole system reliable after that. A small agency can bring those people together around a specific result without pretending every skill has to live under one roof forever.

There is a human reason this model may appeal, too. When people are unsure how long their full-time role will last, building portable expertise and a range of client relationships can feel more resilient than relying on one employer. But independence does not automatically mean security. Contractors take on gaps between projects, benefits, sales, and the risk of being treated as interchangeable capacity. The strongest agencies will give specialists shared standards, a dependable way to collaborate, and a clear owner for the customer's outcome. Full-time teams will still matter wherever continuity, trust, and deep institutional knowledge are essential. The shift we expect is toward more deliberate combinations of both.

The old startup fantasy was to look bigger than you were. Fall 2026 may reward a different move: know exactly what you are good at, prove it in the customer's world, keep enough cash to learn, and assemble the rest with people who are just as sharp. Small can be an advantage. It has to be a deliberate one.