Venture investors built their evaluation playbooks in a world where shipping took months. In a world where any builder can ship a working product in a weekend, the old signals. UI polish, SaaS ARR, prototype sophistication. no longer locate durable value. Here is the framework that does.
A seed-stage founder walks into a VC pitch with a polished demo, a clean SaaS dashboard, and three months of ARR climbing 20 percent month over month. In 2019, that was a company worth funding. In 2026, it is a weekend project that one of their competitors is shipping right now. The investor who funds it on the strength of those signals is not evaluating a moat. They are evaluating a temporary head start measured in days, not years.
The problem is not that the founder is a bad bet. The problem is that the evaluation framework has not updated to account for what has changed. Every traditional signal of product defensibility. UI sophistication, feature completeness, prototype maturity, even early ARR. can now be replicated by a determined competitor with API access and a sprint. The moat has moved underground, below the interface, into the intellectual depth and proprietary architecture that the demo does not show and the pitch deck does not explain. Most investors are not looking there yet.
This post gives investors a new signal stack for the AI era, rooted in three original frameworks: the Moat Inflation Index, the Research Depth Coefficient, and the Replication Surface. These are not replacements for financial diligence. They are the layer that goes underneath it. the layer that determines whether the revenue you are seeing will still exist in 18 months.
The traditional VC evaluation heuristic was built on a world with high build costs. A polished UI meant months of design and engineering investment. A SaaS architecture meant a team with deep technical capability. A working demo meant you had survived the valley of death between idea and execution. These signals worked because they were expensive to fake.
AI-assisted development has compressed those costs by an order of magnitude. A sophisticated UI is now a prompt away. A working SaaS backend is an afternoon project. A polished demo can be assembled in a week by a solo founder with no engineering background. The cost of reaching the surface layer of a product. the layer that investors historically used as a proxy for depth. has dropped to near zero.
This creates what this post terms Moat Inflation. Every barrier that once required months of specialized labor to reproduce is now accessible in days. The surface layer is no longer a signal of anything except speed and access to tools. Investors who continue reading surface signals as depth signals will fund companies that look like moats but are not.
Moat Inflation. The structural devaluation of pre-AI competitive barriers as AI-assisted development reduces the time and cost required to replicate them to near zero. A capability that constituted a durable moat in 2021. a sophisticated UI, a working SaaS architecture, multi-model API integration. has been inflated to a commodity by the availability of AI coding tools, foundation model APIs, and AI-assisted product design. Moat Inflation is not a cyclical phenomenon. It is a permanent structural shift: once a class of capability becomes replicable in days rather than months, it does not become a moat again.
The misreading runs deeper than UI. Revenue itself. the signal investors trust most. is subject to Moat Inflation in the AI era. A company with $2M ARR built on a narrow integration layer over a foundation model API is not a defensible business. It is a distribution head start. The question is not whether the revenue exists. The question is whether the architecture that generates it can survive a competitor who builds the same integration in a sprint and charges less for it.
Before funding any AI company in 2026, ask: could a well-resourced team with API access and three senior engineers replicate the core product in eight weeks? If the honest answer is yes, the revenue you are evaluating does not yet represent a moat. It represents a window.
If UI polish, demo quality, and early ARR are no longer sufficient, what are investors supposed to evaluate? The answer is below the interface. Durable value in the AI era lives in three places: the intellectual depth embedded in the product, the proprietary feedback loops it has built, and how much of the core logic is replicable by a competitor without access to the company's unique assets.
Research Depth Coefficient (RDC). A qualitative measure of how much proprietary intellectual architecture sits below the user interface of an AI product. High RDC means the product's core value derives from original algorithmic work, domain-specific model training, novel evaluation methodology, or formal frameworks that a competitor cannot reproduce by calling the same APIs the product uses. Low RDC means the product's core value is an integration, a wrapper, or a UX layer over a foundation model. replicable in days by any competent team with API access. RDC is assessed by asking one question: if the company's codebase disappeared tomorrow, how long would it take a well-resourced competitor to rebuild the product to functional parity? Days indicates near-zero RDC. Years indicates high RDC.
Replication Surface. The exposed perimeter of a product through which a competitor can build to functional parity using only publicly available resources. foundation model APIs, open-weight models, open-source tooling, and public documentation. A wide Replication Surface means the product can be rebuilt from scratch using nothing the competitor does not already have access to. A narrow Replication Surface means the product depends on proprietary training data, original research, domain-specific tuning, or behavioral signal that exists only inside the company. Replication Surface is the inverse of moat depth. The narrower it is, the more durable the competitive position.
The third signal is one most investors already understand but rarely apply with enough rigor in AI contexts. Signal Moat. the competitive advantage that accrues when a product generates proprietary behavioral signal from real users. is often the most durable moat available to AI companies because it compounds. A competitor can clone the UI. They cannot clone two years of user behavior that has been used to continuously fine-tune domain-specific models. That gap widens with time, not narrows.
The architecture diagram above is the most important thing in this post. Most venture evaluation. the demo, the walkthrough, the ARR dashboard. reads Layers 1 and 2. Those layers are what the founder can show. They are also what a competitor can ship. The question every investor should be asking is: how deep does this company go? What lives below the interface that a competitor cannot buy with API access?
A company operating at Layer 4 does not look dramatically different from one operating at Layer 1 during a 30-minute pitch. That is the problem. Both have a demo. Both have users. Both have ARR. The Layer 4 company has original research, behavioral signal that compounds, and proprietary architecture that cannot be replicated in a sprint. The Layer 1 company is a head start waiting to be erased.
A founder who ships fast, iterates aggressively, and demonstrates strong product velocity is impressive. In the AI era, this is table stakes. AI-assisted development has made shipping fast the baseline, not the differentiator. An investor who funds a company primarily because the founder moves quickly is funding the ability to ship Layer 1 and Layer 2 products faster than competitors who are also shipping Layer 1 and Layer 2 products faster than last year. The signal is real. The moat it implies is not. Early warning signal for this failure mode: the company's pitch is organized primarily around product roadmap and velocity, with minimal discussion of the intellectual architecture below the features.
Revenue is not wrong as a signal. It is incomplete. A company with $3M ARR built on a thin integration layer over a foundation model API may have great revenue. But if its gross margins are being compressed by inference costs, if its core product can be replicated by a better-capitalized competitor who undercuts on price, and if its users have low switching costs, then the revenue is measuring the size of a window, not the depth of a moat. Early warning signal for this failure mode: the company cannot clearly articulate what a competitor with the same API access would need to build to reach functional parity, or the answer is "less than six months."
The most dangerous funding scenario in 2026 is the investor who is impressed by a demo of something they cannot build themselves. The demo is sophisticated. The UI is polished. The outputs are impressive. The investor has no technical context to evaluate whether the impressive outputs are the result of original research or a well-crafted prompt chain over a foundation model. These look identical in a demo. They are not identical in a competitive landscape. Early warning signal for this failure mode: the founder cannot name a single technical decision their architecture makes that a competitor would not also make given the same problem statement.
Illustrative framework. Use this depth map when assessing where a company's core value actually lives before committing capital.
Below is a decision matrix for assessing AI company moat depth. It is built around the three original frameworks introduced above: Moat Inflation Index, Research Depth Coefficient, and Replication Surface. The matrix gives investors a structured way to move the evaluation below Layer 2 before they see the ARR chart.
| Evaluation Variable | Low Depth Signal | High Depth Signal | Weight |
|---|---|---|---|
| Research Depth Coefficient | Product is an integration or wrapper over foundation model APIs. No original research. No citable IP. | Product depends on original algorithms, domain-specific fine-tuning, or proprietary evaluation methodology. Citable contributions exist. | High |
| Replication Surface | A well-resourced team could reach functional parity in under 8 weeks using only public APIs and open-source tooling. | Replication requires proprietary training data, behavioral signal, or original model architecture unavailable to a competitor. | High |
| Signal Moat | No systematic collection of user behavioral signal. Model behavior is static. No compounding feedback loop. | User interactions continuously improve model behavior. Behavioral signal is logged, labeled, and used in fine-tuning. The product gets better as it scales. | High |
| Revenue Quality | ARR is growing but gross margins are compressing. Inference cost as a percentage of revenue is rising. Users have low switching costs. | ARR is growing with stable or improving margins. Inference costs are managed through optimization. Switching costs are high due to integration depth or data lock-in. | Medium |
| Founder Technical Depth | Founder can describe the product but cannot explain the technical decisions that differentiate its architecture from a competitor with the same API access. | Founder can articulate the specific technical choices that would take a competitor months to reproduce, with the reasoning for each. | Medium |
| Domain Data Ownership | All training and evaluation data comes from public sources or the foundation model provider. No proprietary dataset. | Company has assembled or generated a proprietary dataset that a competitor cannot access, covering domain-specific edge cases the foundation model was not trained on. | Medium |
A Series A company presents with $4M ARR, 35 percent month-over-month growth, and a polished diagnostic imaging interface that impressed three radiologists at a demo. The GP applies the Replication Surface test: the core product calls a foundation vision model API with a custom prompt template and a proprietary UI layer. The proprietary piece is the UI and the prompt design. Both are replicable in weeks. The GP then asks for the Research Depth Coefficient assessment: has the company fine-tuned on domain-specific imaging data? Do they have a labeled dataset of cases the foundation model fails on? Do they have citable research demonstrating their accuracy on the specific pathology class the product targets? If the answer to all three is no, the $4M ARR is measuring a distribution window, not a moat. The GP passes and looks for a diagnostics company whose core IP is a proprietary labeled dataset assembled over three years of hospital partnerships.
A pre-seed company with $200K ARR pitches a legal research platform with impressive citation accuracy on a narrow set of case law. The Principal applies the Signal Moat test: does the product log what attorneys fix, what citations they reject, and what queries return no useful results? Is that signal being used to continuously improve retrieval accuracy on the specific case law categories the product targets? The company has no systematic signal collection. Every interaction is stateless. A better-capitalized competitor can reach functional parity in under three months and immediately begin building the signal flywheel the company has not started. The Principal passes. They invest instead in a legal research company that has been logging attorney corrections for 18 months and using them to fine-tune a retrieval model that improves by 12 percent per quarter on the cases law firms actually bring them. That 18-month head start in behavioral signal is a real moat.
An angel sees a polished recruiting AI that generates structured candidate summaries from resumes and interview transcripts. The demo is impressive. The ARR is $80K and growing. The angel applies the Moat Inflation test: what specifically would prevent LinkedIn from building this in a quarter? The founder lists the UI, the prompt design, and the workflow integrations. All three are replicable by LinkedIn in weeks. The angel then looks for the Research Depth Coefficient: does the product have original research on which resume features predict successful hires in specific roles? Does it have a proprietary labeled dataset of hire outcomes? Does it have a feedback loop from recruiters that improves its ranking model? The answer is no on all three. The angel recognizes they are evaluating a workflow tool with a short window before a well-capitalized competitor ships the same thing. They pass and look for a recruiting AI with three years of hire outcome data and a documented accuracy improvement curve.
Funding a Layer 1-2 company at a Layer 3-4 valuation. When the inevitable competitor ships, the portfolio company loses pricing power, margins compress, and the exit multiple collapses. The capital was funding a window, not a moat.
A fund with ten AI portfolio companies that are all integration layers over the same foundation model APIs has concentrated risk in a single technical layer. One model provider changing its API pricing or deprecating a capability can stress the entire portfolio simultaneously.
The companies building at Layer 3-4 often look slower and less polished in early diligence. Their demos are not as slick. Their ARR growth is not as fast. But their moat widens with time. Missing them to fund faster-shipping Layer 1 companies is the defining opportunity cost of outdated evaluation signals.
When too much capital flows to Layer 1-2 companies, it distorts the signal for founders building at depth. Original research-backed companies that grow slower get passed over in favor of faster-shipping wrapper products. The fund that corrects for this first has a structural advantage.
The evaluation framework above is not a replacement for financial diligence. It is a layer that goes underneath it. run before you look at the ARR chart, not after. Here is how to integrate it into a standard process.
Apply the Replication Surface test before requesting financials. Ask the founder a single question: what would a team with the same API access need to build to reach functional parity with your core product? Score the answer on a 1-4 scale mapping to Layers 1-4. If the score is 1 or 2, request a technical deep-dive before proceeding to financial review. Go/no-go gate: score of 3 or 4, or a clear path to 3-4 in the next 12 months with the current team.
Assess Research Depth Coefficient. Request code review or architecture review focused specifically on what is proprietary. Ask for evidence of the signal flywheel. logs, fine-tuning cadence, accuracy improvement curves. Ask for the domain dataset and its provenance. Bring a technical advisor who can assess whether the core algorithms are original or are prompt engineering over public models. Go/no-go gate: at least two of the three high-weight criteria in the evaluation matrix score positively.
Evaluate not just current moat depth but moat trajectory. A Layer 2 company with a clear, resourced roadmap to Layer 4 over 18 months. one that depends on proprietary data the company is actively accumulating. may be a better bet than a Layer 3 company with no flywheel. The question is: does the moat widen or narrow as the company scales? Success criteria: a documented technical roadmap in which the Research Depth Coefficient increases as ARR grows, not decreases.
The framework above is not designed to make investors more cautious. It is designed to make them more accurate. The companies worth funding in the AI era are not the ones with the best demos. They are the ones whose demos are backed by intellectual architecture that cannot be cloned by a well-resourced competitor over a long weekend. Those companies exist. They are being missed by investors who are still reading surface signals.
For a deeper look at what makes an AI product defensible from the product leader's perspective, the framework in The Moat Is Not the Model addresses Model-Product Fit, Capability Ceiling, and Signal Moat from the builder's side of the table. The investor evaluation framework in this post and the product strategy framework there are complementary lenses on the same question: where does the durable value actually live?
The broader context matters too. When 89 percent of enterprise AI pilots never reach production, the companies that can navigate the gap between prototype and deployment are structurally more valuable than the companies that build better prototypes. An investor who can distinguish between a company with a polished demo and a company with a repeatable deployment methodology is looking at a different level of the stack. and a different level of the moat.