Your AI Startup Has No Moat (Yet)
Why Data, Models, and Features Are Not Enough to Build Enduring Companies
AI is making it easier to build a useful product. It is not necessarily making it easier to build an enduring business.
A capable team can now assemble in weeks what once required years of research, specialist engineering, and substantial capital. Models improve, inference becomes cheaper for a given level of capability, development tools become more powerful, and distribution channels can bring an unfamiliar product to thousands of users almost overnight. This is extraordinary progress. It also creates an uncomfortable question for founders: if building the product is becoming easier for you, what prevents someone else from building an equivalent one?
At Atlas AI VB Fund, we believe the answer increasingly lies outside the model itself. The strongest opportunities will often belong to companies that own critical workflows, accumulate proprietary operational knowledge, learn from verified outcomes, and control meaningful access to their customers. They will not simply use AI to produce answers. They will become better at getting work done.
Our thesis is straightforward: a competitive advantage is not defined by how difficult a product is to build, but by how difficult it is for competitors to capture the economic value it creates.
This is not an argument that models never have moats. Nor is it a claim that every application will become a durable software company. It is a framework for identifying which advantages are likely to survive as intelligence becomes more accessible.
1. Intelligence is becoming abundant. Differentiation is not.
By the third quarter of 2026, the economics of model access had shifted dramatically. In a September 2026 analysis, Epoch AI estimated that the cost of achieving a fixed level of AI performance had fallen approximately 47% per quarter since 2023 across five mathematics, science and game benchmarks, equivalent to around a 13-fold annual decline. This is an estimated trend for comparable benchmark performance, not a claim that the latest frontier releases or real enterprise workloads become 47% cheaper every quarter. [1]
The 2026 AI Index also found substantial convergence among leading models on some comparative benchmarks. At the same time, it reported that agents still failed roughly one in three attempts on a structured computer-use benchmark. Capability is improving rapidly, but reliability and real-world execution remain difficult. [2]
Founders should take two conclusions from these facts. First, a capability that appears unique today may become broadly available tomorrow. Second, transforming that capability into reliable, measurable work remains an unsolved operational challenge in many domains.
The decline in fixed-capability costs coexists with a counterintuitive shift in spending. Menlo Ventures' 2026 State of Enterprise AI report estimates that frontier models captured 63% of enterprise model-API spending in Q3 2026, up from 35% in Q3 2025. Their share of tokens increased from 18% to 42%. Buyers are trading up to more capable models, even as cheaper options proliferate; Menlo also estimates that the blended price per closed-model token fell 24% over the year. These are survey-informed market estimates, not audited industry accounts. Taken together, they contradict the simplistic claim that AI intelligence has already become an interchangeable commodity. [3]
Cheaper benchmark capability also does not guarantee a smaller AI bill. Agentic workflows may make many calls, use longer contexts, invoke paid tools, and require human verification. A business can spend more on inference while achieving a lower cost for each individual model capability. The metric that ultimately matters for an application provider is the contribution margin per verified, completed task, not the headline price per million tokens.
The more useful distinction is between temporary technical advantage and durable economic power. A company can benefit enormously from a new model capability without owning a lasting advantage. The question is what it builds around that capability before competitors catch up.
2. Five popular AI moats that are usually weaker than they look
Model access is not ownership
Being early to a model, obtaining a favorable API agreement, or using a sophisticated prompt stack can create a commercial head start. Unless the access is genuinely exclusive and durable, it is not a structural barrier. Other developers can eventually purchase similar intelligence. The relevant test is whether the business remains differentiated when a competitor receives the same model.
'We have proprietary data' is not enough
Companies routinely describe every collection of customer documents, interaction logs, and uploaded files as a proprietary dataset. But access is not ownership, volume is not usefulness, and a customer permitting data processing is not the same as granting permission for cross-customer learning.
A genuine data advantage must answer four questions. Is the information difficult for competitors to obtain? Can the company legally use it for the relevant purpose? Does it improve performance in a measurable way? Does the advantage grow as the product is used?
If a dataset cannot improve the product, it may be an operating input rather than a moat. If every competitor can obtain equivalent information, it may be useful without being exclusive. If a customer can export it and give it to another vendor, the startup's control may be more fragile than its pitch suggests.
Fine-tuning is a capability, not an automatic barrier
Specialized training may make an application faster, cheaper, safer, or more accurate in a particular domain. That is valuable. But the moat is not the fact that a team has fine-tuned a model. It is the scarce training signal, domain knowledge, distribution, or validated performance advantage that other teams cannot reproduce economically.
We would ask whether the tuned system beats the latest general model after accounting for the cost of maintaining it. If the difference disappears with a model upgrade, the company has built an optimization, not necessarily a defensible position.
Integrations can become standardized
Connecting an AI assistant to a CRM, enterprise database, document repository, or ticketing tool is not trivial. But integration protocols, APIs, and development tooling tend to reduce the cost of building connectors.
Integrations are more valuable when they encode authorization rules, compliance procedures, exception handling, and the sequence of actions that actually makes a process work. Ten generic connectors are less defensible than deep responsibility for one mission-critical process.
Features are not franchises
An elegant interface, faster summarization, automatic research, or a promising agent demo can attract customers quickly. None guarantees an enduring business. Features can be copied by neighboring applications, enterprise incumbents, internal development teams, and the providers of the models themselves.
This risk is not hypothetical. In McKinsey's 2026 survey, 32% of respondents reported that their organizations had decided against purchasing at least one software product or feature because they could build it internally with agentic coding tools. That does not mean most software will be replaced. It does mean that the cost of reproducing basic functionality is becoming an increasingly credible negotiating weapon. [4]
3. The most defensible applications own important workflows
The strongest application-layer opportunity is not necessarily to answer the user's question better. It is to complete the user's work better.
Consider an AI system that summarizes a supplier contract. The user reads the answer, checks it, and continues working elsewhere. Another product can reproduce that feature with an equally capable model.
Now consider a system that receives the contract, determines the relevant approval policy, identifies deviations from previously accepted terms, obtains the required human sign-offs, updates the procurement system, records the final outcome, and learns which exceptions the organization consistently accepts or rejects.
The second system is embedded in an operating process, not merely a screen. It knows which decisions matter, who is authorized to make them, what happens when an exception arises, and how successful completion is verified.
This is the transition from a system of information to a system of action.
Workflow ownership creates several possible advantages: customer switching becomes operationally costly; the product observes data that never appear in generic documents; performance can be measured against completed outcomes; and the company may earn the right to expand into adjacent decisions.
But it is essential not to confuse integration with control. A startup that merely reads data from another platform may be at the mercy of that platform. A startup that becomes accountable for an entire outcome may occupy a much stronger position.
A 2026 analysis by Andreessen Horowitz describes the value of embedded organizational processes as process power. The argument is compelling, but the conclusion is conditional: process knowledge matters only if competitors cannot cheaply recreate it and customers continue to reward the vendor for using it. [5]
4. The real data moat is operational knowledge
The most valuable proprietary data is often not a large document archive. It is a record of decisions and outcomes that explains how real work gets done.
Imagine two companies operating in the same regulated industry. Both can access identical public rules and similar customer documentation. One simply retrieves information and generates recommendations. The other, with appropriate permissions, records how experts resolve ambiguous cases, which recommendations were overturned, how exceptions were approved, and whether the final decision produced the intended result.
The second company has the possibility of building a distinctive operational knowledge base. This can include edge cases, context, authority structures, verified mistakes, and domain-specific standards of quality.
Possibility is the important word. It is not yet proof of a moat.
Customer data frequently remain inside the customer's environment for privacy, security, or contractual reasons. Training across customers may be prohibited. Some feedback is subjective, inconsistent, or impossible to label reliably. A system may accumulate thousands of interactions without learning anything that improves future decisions.
The strongest data thesis therefore rests on rights, quality, and measured marginal value, not on storage volume.
Our preferred test is a controlled comparison: give the same model and tools to a baseline system, then add the startup's claimed proprietary context and decision knowledge. Does accuracy, completion rate, speed, or economic value improve materially? Does that improvement persist across customers and as baseline models advance?
If the company cannot answer those questions, the dataset should be treated as an asset under development, not a proven competitive advantage.
5. Compounding learning is powerful, but only when feedback is trustworthy
The ideal AI learning loop is easy to describe.
More customer usage creates more observable decisions. More decisions create more verified feedback. Better feedback improves evaluation and orchestration. Better performance attracts more users, generating new evidence and making the system harder to replicate.
That is how operational learning can compound.
Unfortunately, the loop can also run in reverse. Incorrect outputs can become incorrect labels. Automated approval can be mistaken for correctness. Unrepresentative customers can distort a model's judgment. A startup may end up learning to optimize its own metric rather than the result its customer actually wants.
This is why evaluation infrastructure matters. In a 2026 industry survey of more than 1,300 professionals, approximately 89% of respondents reported some form of agent observability, but only about 52% reported offline evaluations on test sets and 37% reported online evaluations. The survey is not a representative census of all businesses, but it illustrates the gap between observing behavior and systematically improving it. [6]
Reliable evals need difficult cases, human judgment where appropriate, clear definitions of success, and regression testing whenever a model, tool, or workflow changes. Research on agent evaluation emphasizes how easily poorly designed tests can reward the wrong behavior or reject a correct outcome. [7]
The moat is not 'we collect feedback.' It is 'we possess a repeatable and legally usable mechanism that turns verified real-world outcomes into better performance faster than competitors can.'
The distinction is substantial. Logging activity is cheap. Accumulating good judgment at scale is not.
6. Distribution may matter more than technical elegance
The venture industry understandably admires technical breakthroughs. Customers, however, do not buy novelty in isolation. They buy confidence, outcomes, compatibility, and solutions to painful problems.
An AI startup with modest technical differentiation can become a formidable competitor if it owns a trusted distribution channel, wins a high-frequency workflow, or is the default choice within a professional community. Conversely, a superior model implementation may struggle if the customer acquisition economics are unsustainable.
In regulated and complex enterprise markets, distribution is inseparable from trust. Buyers need evidence of performance, security guarantees, implementation support, and a credible answer to who is responsible when the system fails.
The nuance is that distribution is not simply a large contact list or a founder's network. A durable distribution advantage lowers acquisition costs, raises conversion or retention, or provides privileged access that competitors cannot easily purchase.
The current market is unusually open to new entrants. Menlo Ventures estimates that AI-native providers captured 77% of enterprise application-layer AI spending in its 2026 market model. Even excluding the frontier-model providers, independent AI-native companies accounted for an estimated 61%. These figures suggest that incumbent distribution is not an impenetrable wall. They do not show that today's winners will retain their positions. [3]
For founders, the implication is uncomfortable but productive: product velocity can create the opening, but the company still needs to convert adoption into an enduring customer relationship before a better-funded competitor arrives.
7. A workflow moat is strongest when the customer can measure the outcome
There is a dangerous gap between AI usage and AI value.
A tool may have many weekly active users, generate millions of tokens, and produce enthusiastic employee feedback while creating little measurable financial benefit for the organization.
McKinsey's 2026 survey found that 80% of respondents reported improved individual productivity from AI, yet only 37% reported a positive contribution to organization-level EBIT. Around 6% qualified as high performers under the survey's definition. These are respondent reports, not controlled causal estimates, but the difference matters. [4]
The same study found that nearly three-quarters of AI high performers reported fundamentally redesigning workflows, compared with about one-quarter of other respondents. Correlation is not causation; companies that are already effective operators may be better at AI adoption as well. Still, the association supports the idea that workflow redesign matters more than installing isolated AI features. [4]
Experimental research provides another useful benchmark. A peer-reviewed study in The Quarterly Journal of Economics examined the introduction of generative AI assistance to more than 5,000 customer support workers and found an average productivity increase of approximately 15%, with substantial differences across experience levels. It demonstrates that AI can create measurable operational value, not that the vendor capturing that value automatically owns a moat. [8]
The best products make the chain from action to outcome visible. They can tell a customer not just how many tasks were attempted, but how many were completed correctly, how many required human intervention, what changed in throughput or quality, and what economic benefit was realized.
A startup that cannot distinguish activity from verified value is vulnerable to being treated as a discretionary software expense.
8. Creating value is not the same as capturing value
This is where many AI investment theses become financially careless.
A system may save a customer hundreds of hours and still have unattractive economics. The provider may pay substantial inference costs, rely on expensive human review, require custom integrations for every deployment, or be forced to lower prices as competitors obtain the same underlying intelligence.
We therefore care about contribution margin per completed, verified task, not merely gross revenue or the number of agents deployed.
A practical calculation begins with revenue attributable to a successful outcome. From that, subtract model inference, retrieval and infrastructure, third-party tools, variable implementation work, human verification, failure remediation, and ongoing customer-specific service. Then examine how the calculation changes as volume increases and model prices move.
This is not a claim that every AI application should have traditional SaaS margins. Some outcome-based businesses will resemble software-enabled services and can still be attractive. But the valuation must reflect the economics of the actual business, not a convenient label.
McKinsey found that approximately 20% of survey respondents considered AI-related operating costs a constraint on usage in 2026. This reinforces the need to model workload-specific costs rather than assume they will fall away. [4]
A strong company should have a credible mechanism to retain value even as a cheaper model makes its core technical task less expensive. It might do this through superior distribution, trusted process control, more complete outcomes, or the ability to perform work that previously could not be automated.
The key is economic power. If customers can reproduce the same outcome at a fraction of the price, or a platform can absorb the functionality into an existing product, then impressive revenue growth may prove temporary.
9. The platform threat cannot be ignored
AI application companies face competition from three directions at once: other startups, established software platforms, and the providers of underlying intelligence.
The third direction is particularly difficult because the supplier may control capabilities, pricing, tools, and customer access while also developing its own applications. A business that depends on one model provider and offers little beyond that provider's core capabilities may be exposed to both commercial and technical displacement.
The correct response is not to avoid external models. Building every layer of the stack internally would consume enormous capital and slow product development for little guaranteed advantage.
Instead, founders need to know which dependencies are strategic and which are replaceable. They should understand how performance changes when substituting another capable model, whether proprietary workflow logic survives the substitution, who owns customer relationships, and whether contracts protect the right to use accumulated knowledge.
Portability itself can create negotiating leverage, although maintaining multiple models has a real engineering cost. More importantly, the company should own what cannot be bought through a standard API: customer trust, validated workflows, rights to useful data, embedded decisions, and operational accountability.
A founder should be able to answer one question without rehearsed jargon: if the underlying model becomes substantially better and cheaper next year, does our company become more valuable or less necessary?
10. This thesis does not apply equally to every layer of AI
A simplistic version of the argument would say that models have no moats and all value inevitably moves to applications. We reject that framing.
Frontier research, specialized hardware, novel architectures, scientific intellectual property, proprietary measurement systems, and capital-intensive infrastructure can create genuine technical and economic barriers. Their advantages must be assessed using different criteria, including research capability, scale economics, access to physical resources, intellectual property, and time to replication.
At the same time, application-layer moats are not guaranteed to be stronger. A software company can own an important workflow today and lose it tomorrow if the underlying platform absorbs the process, customer switching costs decline, or the market shifts to a new interface.
We should expect competition among layers, not a predetermined winner. Falling costs can expand demand and make previously impossible products viable. They can also transfer bargaining power to customers and compress margins.
The relevant investment question is not 'Which layer of AI always wins?' It is 'Where can this particular company build and retain economic power as the technology changes?'
11. What we look for at the earliest stages
Pre-seed companies rarely have fully formed moats. Demanding one would be incoherent. The important issue is whether the founders have a credible path toward creating one.
We focus on the evidence behind that path.
A painful and recurring customer problem: not a feature in search of a budget, but work that somebody must get done.
A differentiated starting point: unusual domain expertise, a technically exceptional team, privileged access to a workflow, or a more effective route to customers.
A theory of proprietary learning: a clear explanation of which decisions or outcomes the product will observe, which rights it will possess, and why those signals become more valuable with usage.
An explicit measurement system: benchmarks against strong alternatives, real customer outcomes, documented failures, and a process for detecting regressions.
A credible path to economics: realistic assumptions about inference, human-in-the-loop costs, onboarding, retention, and the ability to price an outcome.
A platform-risk assessment: a thoughtful answer to what happens if the model provider, incumbent system, or customer builds the same functionality.
These are not six boxes that every startup must tick on day one. They are six lines of inquiry that separate a plausible product demonstration from a plausible venture-scale company.
We are particularly interested in how the evidence changes over time. Does onboarding become faster? Does the percentage of tasks completed correctly improve? Does the system need less human review for comparable cases? Do customers expand usage because the product becomes a natural part of their operations? Do margins improve with scale, after accounting for real service costs?
Equally important, what would falsify the moat thesis? A founder who can articulate the conditions under which their supposed advantage disappears is often better prepared than one who claims to be impossible to copy.
12. The companies that endure will get harder to replace as they grow
The next generation of valuable AI businesses will not be defined only by how much intelligence they can access. They will be defined by what they make that intelligence responsible for.
Some will win through unique science or infrastructure. Others will win through distribution, trust, deep workflow control, and accumulated operational knowledge. The most compelling application companies may combine several of these forces, allowing initial product adoption to develop into stronger economic advantages over time.
But founders should resist the industry's favorite shortcut: attaching the word 'proprietary' to whatever currently appears in the product architecture.
A collection of prompts is not necessarily a moat. A database is not necessarily a moat. A large number of integrations is not necessarily a moat. And rapid growth, while essential to many startups, does not prove long-term defensibility.
What matters is the direction of travel. Does every new customer make the company meaningfully better? Does every verified outcome deepen knowledge competitors cannot cheaply reproduce? Does expanding usage strengthen the customer relationship and increase the share of economic value the company can retain?
At Atlas AI VB Fund, we do not expect founders to have all these answers at inception. We do expect them to recognize the distinction between a product that can be built and a company that can endure.
The model may make you useful. What you build around it must make you difficult to replace.
That is where the moat begins.
Research and references
Selected original reports, peer-reviewed research, and industry surveys. Survey results and market-sizing estimates are identified as such in the text. References are current to 12 November 2026.
Comments