AI Growth Defensibility

Core thesis

No startup is born with defensibility, yet every successful company dies without it. Speed buys distribution. Distribution buys the opportunity to build defensibility. Defensibility buys time to build the next defensibility. You cannot skip stages. You cannot rest at any stage. In the AI era, this sequence compresses to months instead of years. The companies I see winning are not the ones with the best model. They are the ones that understand defensibility is a process, not a feature, and they build toward the next stage while operating in the current one.

Why AI-era defensibility is harder

The LangChain State of Agent Engineering survey, published in early 2026, found that 57.3 percent of organizations have AI agents in production. That number jumps to 67 percent for enterprises with more than 10,000 employees. The barrier to building has collapsed. Anyone with an API key can launch a product in a weekend. This means defensibility cannot come from the product alone. It has to come from what accumulates around the product: data, relationships, learned preferences, workflow integration, brand trust. The Writer 2026 Enterprise AI Adoption survey found that 79 percent of organizations face significant AI adoption challenges, a double-digit increase from 2025. The gap between building and adopting creates the window where defensibility forms. Companies that close the adoption gap while competitors ship features are the ones that build real moats.

The Speed Trap

Speed is the most seductive false moat in AI. I have watched teams raise on speed, hire for speed, and optimize for speed, only to discover that speed alone is a bailey. Hopin is the canonical example. At its peak during the pandemic, Hopin raised over $1 billion across multiple rounds, scaled to more than 800 employees, and hosted 80,000 events per month. It became the fastest-growing European startup in history. When the pandemic ended and the virtual events market normalized, the company sold its core assets for approximately $15 million in 2023. Speed had created the illusion of defensibility. There was no structural moat underneath. The lesson for AI companies: if your only advantage is that you shipped first or grew fastest, you have no advantage. Speed must convert into something durable before the window closes. Deloitte's 2026 State of AI report underscores this: worker access to AI rose by 50 percent in 2025 alone, but the number of companies with more than 40 percent of AI projects in production remains below 20 percent. The gap between access and production is where speed alone fails.

Five misconceptions that kill AI companies

• "Speed is the only moat": Hopin proved speed collapses without structural defensibility. In the AI era, development speed is a commodity. Every team has access to the same models, the same coding agents, the same deployment infrastructure. The team that ships three days faster does not win. The team that converts speed into a data advantage, a network effect, or a brand moat wins.
• "Distribution is the only moat": HubSpot's content marketing playbook built a category-defining company, but once the tactics became public, competitors copied them within months. Distribution without proprietary data or network effects is replicable. AI distribution channels are becoming more crowded, not less. Gartner projects that by 2027, 60 percent of B2B companies will use AI-generated content in their marketing. When everyone has AI content, distribution returns to the same bottleneck: trust.
• "Data is the moat": TripAdvisor had millions of reviews, but when Google entered the travel market with its own review aggregation, TripAdvisor's data alone did not protect it. Data is only a moat when each additional unit of data produces marginal improvements that competitors cannot replicate. If a competitor can train on public data and reach 80 percent of your quality, your data moat is thin. The question is not "do you have data?" but "does each new data point make the product meaningfully better than the last?"
• "Moats are permanent": BlackBerry dominated the smartphone market until it did not. Yahoo was the internet's front page until Google replaced it. Blockbuster had 9,000 stores and 60,000 employees at its peak. Moats erode. The AI era accelerates erosion because model capabilities jump every six to nine months. A moat built on model quality last quarter may not survive the next model release. The only durable moat is one that compounds: each user makes the product better for the next user.
• "AI made previous moats obsolete": Network effects, brand, economies of scale, regulatory capture, and switching costs still work. AI changes the speed at which moats form and erode. It does not change the categories of moat. The companies that win in AI will have the same structural advantages that winners had in previous technology waves. They will just build them faster, and they will lose them faster if they stop reinvesting.

The Motte-and-Bailey Framework explained

The Motte-and-Bailey framework, adapted from medieval castle design, distinguishes between shallow and deep defensibilities. The bailey is the outer wall: easy to build, easy to breach, but useful for buying time. The motte is the inner keep: hard to build, hard to breach, and where you retreat when the bailey falls. In AI, bailey defensibilities include speed to market, temporary distribution advantages, early brand buzz, and public-data curation. These are real advantages for 12 to 24 months. But they erode. Motte defensibilities are structural: direct network effects (more users make the product better for all users), cross-side network effects (more buyers attract more sellers), data network effects (each interaction improves the model for everyone), brand (trust accumulated over years), and economies of scale (unit costs that decline with volume). The rule I give every team I work with: identify your bailey. Identify your motte. Spend 80 percent of your strategic energy building the motte while the bailey holds. If you cannot name your motte, you do not have one.

The Defensibility Sequence: the full map

Speed leads to Distribution. Distribution leads to Engagement. Engagement leads to Data. Data leads to Network Effects. Network Effects lead to Platform. Each arrow is a bridge that must be deliberately built. Skip a bridge and the sequence breaks. Most teams I work with are stuck between Speed and Distribution. They shipped fast, got some users, and plateaued. They never built the bridge from speed to sustainable distribution. The sequence is not automatic. It requires active engineering. At each stage, the team must ask: what is the bridge to the next stage, who is building it, and what is the metric that tells us we have crossed it?

Stage 1: Speed to Distribution

The first bridge is converting development speed into customer acquisition. This is not about faster shipping. It is about shipping things that acquire users. A team that ships 10 features per week but acquires zero users per feature is moving fast in the wrong direction. The bridge from speed to distribution has two components. First, every feature must have a distribution hypothesis: who will hear about this, and how will they find it? Second, the team must measure distribution velocity: users acquired per unit of development time. The LangChain survey data is relevant here. Fifty-seven percent of organizations have AI agents in production, but the quality gap remains the number one barrier at 32 percent. Shipping broken AI features burns distribution channels. Speed without quality burns the channel before it opens.

Stage 2: Distribution to Engagement

Getting users is not the same as keeping them. The bridge from distribution to engagement is the activation experience. I have seen AI products acquire 100,000 users in a week and retain fewer than 500 after month one. The distribution worked. The engagement failed. The bridge requires three things. First, a clear activation metric: what does a user need to do to experience the core value? Second, an onboarding path that gets them there in under five minutes. Third, a measurement system that tells you what percentage of new users cross the activation threshold. If that percentage is below 30 percent, distribution is pouring water into a leaky bucket. Fix the bucket before you widen the hose.

Stage 3: Engagement to Data

Engaged users generate data. But not all data is defensibility-grade. The bridge from engagement to data requires deliberate instrumentation. The product must capture the signals that improve the model for every user, not just the individual. A user who generates 1,000 interactions that never feed back into the model is generating waste. The bridge has two components: signal design and feedback loops. Signal design means every user action that could improve the product is captured, labeled, and routed. Feedback loops mean the captured signal is used to improve the product within a timeframe the user notices. If a user corrects the AI on Monday and the same correction is still needed on Friday, the loop is broken and the data is not compounding.

Stage 4: Data to Network Effects

Data alone is not a network effect. Data becomes a network effect when each additional user's data makes the product better for every other user. This is the hardest bridge to build because it requires two conditions most teams never achieve. First, the data must be proprietary. Competitors cannot replicate it by scraping the web or buying a dataset. Second, the data must have increasing marginal returns. Each new data point must produce a meaningful improvement in product quality. If quality plateaus at 10,000 users, the data network effect stops at 10,000 users. I tell teams to measure the data network effect coefficient: for every doubling of users, what is the percentage improvement in a key quality metric? If the coefficient approaches zero, the data moat is not compounding and the company is running on a bailey.

Stage 5: Network Effects to Platform

Platform is the terminal defensibility stage. A platform is a product that third parties build on top of. When other companies depend on your product to serve their customers, switching costs become structural. OpenAI is building toward platform defensibility. Developers build on the API. Startups depend on the models. Enterprises integrate the tooling. Each integration is a switching cost. But platform defensibility is fragile in AI because models are increasingly commoditized. The platform bridge requires more than an API. It requires workflows, integrations, data residency, compliance certifications, and SLAs that a thin API wrapper cannot provide. The companies building true platforms in AI are not the ones with the best models. They are the ones with the deepest integrations.

Three Self-Diagnostic Questions, expanded

• Can a competitor clone your core value in under 90 days? If yes, your moat is shallow. The test is not whether they can build a similar product. It is whether they can replicate the experience, the data, the integrations, and the trust that your users depend on. If the answer is yes, you are in the bailey. Build the motte.
• Are you shipping meaningful improvements weekly? This is not about commit count. It is about whether users notice the improvement. If you ship 10 changes per week and users cannot tell the difference, you are not shipping. You are rearranging deck chairs. Meaningful improvement means a user who used the product last week has a measurably better experience this week.
• Does losing your top 10 percent of users break the product? This tests whether your network effects are structural or cosmetic. If losing your power users makes the product worse for everyone else, you have early network effects. If the product works the same with or without them, you do not have network effects. The test is not about revenue concentration. It is about whether losing users degrades the experience for the users who remain.

Measuring defensibility: the moat scorecard

I use a simple scorecard to measure defensibility across the six-stage sequence. For each stage, rate your company from 1 (no progress) to 5 (fully built). Speed: how fast does the team move from idea to shipped feature? Distribution: how many users do you acquire per unit of engineering investment? Engagement: what percentage of new users return in week two? Data: is your data proprietary and does it compound? Network Effects: does each new user improve the product for existing users? Platform: do third parties depend on your product? Score each dimension. The total is out of 30. A score below 10 means the company is running on a bailey. A score above 20 means structural defensibility is forming. Most AI startups I evaluate score between 8 and 14. The gap between 14 and 20 is the distance between a company that might survive and a company that probably will.

The playbook for getting started

Start with the three self-diagnostic questions. Answer them honestly with your team. Do not sugarcoat. The moment you realize your moat is shallow is the moment you can start building a deeper one. Next, map your current position on the defensibility sequence. You cannot skip stages. If you are in the speed stage, your job is to build the bridge to distribution, not to fantasize about platform. Pick the one bridge you need to build next. Assign an owner. Give them a metric. Set a six-week deadline. Review progress weekly. The sequence takes discipline, not genius. The teams that build real moats are the ones that treat defensibility as a product they are shipping, not a concept they are discussing.