From RAG to Riches

1,630 Words — A 9 Minute Read

Each day, I have the pleasure of meeting founders building incredible new AI applications and businesses.

Yet, as I meet with these founders and dig deeply into their businesses, there is a question that continuously plagues me: Do any of these companies have enduring defensibility without “owning” the underlying model they built on top of?

Many of these companies are solving complex yet wonderful use cases. I speak with their customers and hear things like, “I don’t even want to think about returning to the 'old way' of doing things.”

While that’s an expression of real value, in most cases, the startups do not “own” the underlying model that is the “engine” of their product. Rather, they are building on top of frontier models.

In practice, these startups often use retrieval-augmented generation, or “RAG,” a pattern in which a language model retrieves relevant external information at query time and uses it as context to generate a response, rather than relying solely on its trained weights.

In other words, startups use RAG to feed the underlying frontier model business- or customer-specific context at the moment a question is asked, so the answer is grounded in the right information.

Instead of: “Answer this from your [the model’s] general training.”

It becomes: “Answer as if you had just read these specific documents.”

That is what makes the response useful, accurate, and contextualized.

So then, back to my question: Is there any real and enduring value being created by these companies when they don’t own the “engine?"

The rest of this article dives into a framework I am developing to help answer this question:

(It’s developing, but is far from developed)

1. Start with a hard truth: model ownership is rarely the moat

2. Why “proprietary data” is an incomplete answer

3. Where defensibility actually comes from without owning the model

• Moat 1: Workflow entanglement, not intelligence
• Moat 2: Domain-specific abstraction layers
• Moat 3: Distribution and trust, not capability
• Moat 4: Human-in-the-loop learning loops
• Moat 5: Economic moats via cost and margin structure

4. Why OpenAI and Google usually do not win these battles

5. A litmus test for determining enduring value

The Framework in Ernest

1. Start with a hard truth: model ownership is rarely the moat

• Owning the base model is neither necessary nor sufficient for durable advantage.
• Empirically: Frontier models commoditize quickly. GPT-3.5 → GPT-4 → GPT-4.1 type jumps compress differentiation windows to 6–18 months. Training frontier models requires capital (and a lot of it), data scale, and infrastructure that only a handful of actors can sustain.
• Most economically valuable companies historically have not owned the “core intelligence” layer. Oracle did not invent SQL. Salesforce did not invent databases. Bloomberg did not invent terminals.
• So the question is not “can OpenAI build this?” because the answer is almost always yes. The correct question is “Would OpenAI win economically if it tried?” Or, “Is it worth their time?”

2. Why “proprietary data” is an incomplete answer

RAG does not compound as model weights do. Retrieval latency grows with corpus size. Long-context reasoning degrades with shallow grounding. You cannot internalize customer-level patterns into the model weights. Feedback loops remain brittle, external, and slow.

This means most “data moats” built via RAG are static, not compounding.

Inability to update weights limits learning velocity. If you cannot fine-tune or continually train, you are stuck at the application layer. Your system primarily improves through prompt engineering, workflow tuning, and intelligent, optimized batching, not through true representation learning, gradient descent, or backpropagation. This caps the slope of performance improvement over time.

So, “we have proprietary data” is usually overstated and often fragile.

3. Where defensibility actually comes from without owning the model

(and why founder-market fit is the prerequisite)

For AI companies building atop frontier models, enduring defensibility does not emerge from the model layer itself. It emerges from how intelligence is applied inside real-world systems. In practice, each durable moat below is enabled or accelerated by strong founder–market fit, specifically founders with lived experience facing the same problems as their customers.

There are five real moats that do persist even when the model is outsourced.

Moat 1: Workflow entanglement, not raw intelligence

The most defensible AI companies do not merely assist users; they embed directly into core workflows and become difficult to remove.

Founder–market fit is critical here. Founders who have personally operated within the target domain understand:

• Where work actually breaks down,
• Which steps are mission-critical vs superficial?
• Where automation creates leverage vs risk,
• Which handoffs and exceptions cannot be skipped?

As a result, they design AI systems that sit inside irreversible workflows, not alongside them. This depth of integration is difficult for generalist platform providers to replicate, creating high switching costs even when the underlying models are commoditized.

Moat 2: Domain-specific abstraction layers

Defensibility increasingly lives in the abstractions a company builds on top of models, not the models themselves.

Founder–market fit enables founders to encode tacit, hard-won domain knowledge into software abstractions that:

• Hide model complexity from users,
• Reflect real operational constraints,
• Anticipate edge cases before they surface,
• Convert unstructured chaos into narrow, actionable decisions.

Without lived experience, teams tend to expose model behavior rather than compressing domain complexity. With it, the abstraction layer becomes proprietary even if the underlying intelligence is not.

OpenAI could replicate the intelligence, but not the encoded understanding of how work actually happens in a vertical.

In other words, this encoded understanding is wisdom. As our Managing Partner and founder, Mark Phillips , recently wrote in his article Wisdom Economy, “Wisdom is how to live. It is the residue of mistakes, metabolized by time and reflection. It cannot be rushed or coded. It is an embodied – as in felt in the body – experience, guidance from the inside.” You can read that article here.

Moat 3: Distribution and trust, not capability

In enterprise and operational markets, customers do not buy “the best model.” They buy trustworthy systems that will not jeopardize their job, compliance posture, or uptime.

Founders with lived experience:

• Understand the real buyer and the buying committee,
• Speak the customer’s language rather than the model’s,
• Anticipate regulatory, security, and reputational concerns,
• Design adoption paths that feel safe rather than experimental.

This leads to faster, more durable distribution and trust that compounds over time, creating a moat that frontier model providers often struggle to replicate due to misaligned incentives and brand risk.

Trust compounds slower than technology, but it compounds longer.

Moat 4: Human-in-the-loop learning loops

Even without updating model weights, companies can compound advantage through organizational and workflow-level learning.

Founder–market fit enables teams to:

• Identify which feedback signals matter and which are noise.
• Classify errors correctly (model failure vs workflow failure),
• Design escalation and override paths that mirror real practice,
• Improve outcomes without retraining models.

This creates a learning system that improves with usage, even when the underlying model remains static.

Moat 5: Economic moats via cost and margin structure

Owning the model is often not the economic advantage; deploying intelligence efficiently is.

Founders who have lived the problem understand:

• Where high-cost intelligence is justified,
• Where cheaper heuristics suffice,
• Which decisions generate ROI and which do not?

This leads to selective model usage, tighter unit economics, and margin expansion at scale. Large platform providers, by contrast, are structurally poorly suited to narrow, cost-optimized vertical software.

4. Why OpenAI and Google usually do not win these battles

They face real constraints:

• Channel conflict with partners.
• Brand risk from vertical failures.
• Incentive misalignment between platform revenue and vertical depth.
• Organizational gravity toward generality.

They win when the abstraction is horizontal.

They lose when the abstraction is operational. Mostly because it’s not worth their time, at their scale, to fight specialized, vertical use cases.

This is why:

• AWS did not kill Snowflake.
• Salesforce did not kill vertical CRMs.
• OpenAI will not kill most vertical AI companies that actually embed into work.

5. A litmus test for determining enduring value

When evaluating a company building atop frontier models, I ask:

• If the model got 10x better overnight, would this company still matter?
• If the model got 10x cheaper, would this company’s margins expand or collapse?
• If OpenAI launched a competing product, would customers churn immediately or hesitate? Why?
• Does value accrue from usage, or only from intelligence?
• Is the company learning faster than the model provider, even if the model stays fixed?

If the answers are strong, model ownership is irrelevant.

Wisdom Over Weights

What should your takeaway from all of this be?

1. Models are not the source of defensibility - Owning the underlying model is neither necessary nor sufficient for enduring value. Frontier intelligence is becoming faster, cheaper, and more broadly accessible. As a result, model quality alone rarely determines long-term outcomes. Asking whether a startup “owns the model” is usually the wrong question; the model is infrastructure, not advantage.

2. Defensibility lives in judgment, not prediction - As intelligence commoditizes, advantage shifts to how that intelligence is applied. Enduring companies own the judgment layer: the workflows and intelligence embedded in them, the abstractions that compress complexity, the trust and distribution that enable adoption, the economic structure that determines ROI, and the learning loops that improve outcomes over time. Prediction is cheap; deciding when, where, and how to deploy it inside real systems is not.

3. Wisdom precedes judgment; founder–market fit is the gate - The ability to build and own this judgment layer comes from lived experience. Founders who have faced the problem themselves understand where work breaks, where automation creates leverage vs. risk, and where trust is earned. That tacit knowledge cannot be learned from data alone or replicated by a platform provider. Founder–market fit is not a nice-to-have; it is the prerequisite for every durable moat in an AI-first world.

Special thanks to:

• Thank you to Matthew Black and Rohit Nair - speaking with you two recently has illuminated many of these questions in my mind.
• Thank you Cheng Nie -- Dr. Nie was one of my favorite professors I had at Iowa State University - Ivy College of Business - I posed the central question of this article to him in an email.
• And of course, thank you to my team at 11 Tribes Ventures for helping review each of my research articles