AI-Native Moats Under Frontier-Model Improvement

Answer#

Frontier-model improvement is a moat stress test. It destroys advantages made of generic workflow know-how, prompt scaffolding, migration friction, or raw "we use better AI" claims. It does not automatically recreate elapsed customer history, external networks, exclusive resources, scale economies, or incumbent reluctance to cannibalize itself.

The surviving AI-native moats are therefore:

External structural powers: network effects, scale economies, and cornered resources from Seven Powers Applied to AI.
Time-locked proprietary behavioral data: the Compounding Data Moat where real users refine workflows inside the product over calendar time.
Domain edge-case artifacts: vertical-specific evals, tests, integrations, and exception maps built from actual customer use, not generic domain descriptions.
Deep workflow/platform embedding: customers building automations, APIs, webhooks, trained routines, and operating procedures around the product.
Counter-positioning against legacy incumbents: the The AI-Native Safe-Choice Inversion where "legacy incumbent" becomes the unsafe choice because it cannot credibly become AI-native without retraining, rebuilding, and cannibalizing itself.

The weaker moats are:

Product velocity: Product Velocity as Moat wins trust and lands customers, but it is a treadmill unless it converts into data, workflow, network, or platform lock-in.
Narrow wedge: Narrow Wedge into a Legacy Market is a high-leverage entry strategy, not a final defense.
Custom model / best AI claims: durable only if tied to proprietary data, proprietary workflow distribution, or cost/latency economics. If the claim is merely "our model is better," frontier-model improvement commoditizes it.
Generic switching costs and process power: Seven Powers Applied to AI says these erode because agents can port data, rebuild integrations, and hill-climb processes.

Survival Table#

Moat	Survives model improvement?	Why
Network effects	Strong yes	A better model does not recreate the user's network, marketplace liquidity, or ecosystem participation.
Scale economies	Strong yes	Better models may lower operating costs, but they do not remove capital intensity, utilization advantages, or fixed-cost spreading.
Cornered resources	Strong yes	AI does not grant access to exclusive contracts, proprietary data streams, regulatory position, or scarce talent.
Time-locked behavioral data	Yes, if real	A competitor can copy features but cannot buy months of customer workflow history already generated inside the product.
Domain edge-case artifacts	Partial yes	Survives when based on rare, local, customer-specific exceptions. Erodes when it is just public vertical knowledge the next frontier model can internalize.
Workflow/platform embedding	Yes, if deep	Generic switching costs erode; customer-built APIs, automations, trained routines, and operating procedures are harder to agent-port cleanly.
Counter-positioning	Yes, but time-bounded	Improves while incumbents cannot credibly become AI-native. Weakens once the category's largest AI-native vendor becomes the new safe choice or incumbents gain credible AI.
Product velocity	Weak alone	It lands deals and signals trust, but frontier models make everyone faster. It must compound into another moat.
Narrow wedge	Weak alone	It focuses the attack and creates learning, but it is a wedge, not a wall.
Custom foundation model	Unclear	Durable only if it encodes cornered data, workflow signal, or cost structure. Otherwise the frontier labs catch up.
Process power	No alone	Better models can infer and hill-climb workflows. Process must attach to scale, data, network, or external rights.
Generic switching costs	No alone	Agents reduce migration pain by porting data, rebuilding integrations, and regenerating configuration.

The Real Data Structure#

Bad framing: "AI-native companies have a moat because they use AI."

Better framing: an AI-native moat is a stack with different half-lives:

Layer	Function	Durability
Narrative inversion	Makes the buyer willing to switch	Medium, category-cycle dependent
Narrow wedge	Finds the first segment where incompleteness is acceptable	Low, GTM mechanism
Velocity	Keeps the wedge expanding faster than customer complexity	Low to medium, relative to competitors
Behavioral data	Converts usage into proprietary learning	High if captured from real workflows
Domain artifacts	Encodes edge cases generic models miss	Medium to high if private/local
Platform embedding	Makes the product part of customer operations	High if customers build on top
External powers	Network, scale, cornered resource	Highest

The first three layers win the land. The last four defend the expand. Treating velocity, AI branding, or the wedge as the moat is confusing the entry motion for the defensibility.

Campfire Read#

Campfire is the useful case because it contains both durable and fragile claims.

The durable stack:

The AI-Native Safe-Choice Inversion: boards and executives now give finance buyers air cover to choose AI-native ERP over NetSuite.
Narrow Wedge into a Legacy Market: Campfire did not need all of NetSuite; it only needed to be best for high-growth tech companies outgrowing QuickBooks and using a small slice of NetSuite.
Product Velocity as Moat: customers buy the trajectory - "you will scale with us" - not only today's feature set.
Compounding Data Moat: multi-entity accounting workflows, financial operating patterns, and integrations can become proprietary behavioral history and workflow lock-in.

The fragile claim:

"Best AI" and "custom foundation model" are not automatically durable. If the custom model is just a better model, frontier-model improvement attacks it directly. If it encodes Campfire-specific financial workflow data, edge cases, agent platform behavior, and customer interaction traces, it becomes a Compounding Data Moat or cornered resource.

So the answer for Campfire is not "custom foundation model survives." The answer is: custom model survives only as the carrier of proprietary workflow signal; product velocity survives only if it creates that signal before competitors catch up.

What Model Improvement Actually Erodes#

Harness Shrinkage as Models Improve gives the mechanism. Capabilities that used to require scaffolding move into the model. The business equivalent is:

prompt tricks stop differentiating;
generic AI workflows stop differentiating;
agent-portable integrations stop differentiating;
"we can ship fast" becomes table stakes among AI-native peers;
process knowledge that can be inferred from public examples stops differentiating;
vertical software that depends only on "the incumbent is clunky" loses urgency once all entrants are AI-native.

This is why Seven Powers Applied to AI is the right frame. Model improvement is not evenly good or bad. It is good for companies whose moat is outside the model and bad for companies whose moat is a temporary model capability gap.

Practical Test#

For any proposed AI-native moat, ask what happens the day a much stronger frontier model ships.

If the stronger model lets competitors copy the advantage faster, the moat is weak:

generic process power;
prompt scaffolding;
broad feature catch-up;
migration friction;
"our AI is better" without proprietary signal.

If the stronger model makes the incumbent's position worse, the moat may strengthen:

counter-positioning where the incumbent's old stack and org cannot absorb the new model cleanly;
product velocity against a legacy regression surface;
narrow wedge expansion before the incumbent can reorganize.

If the stronger model does not grant access to the underlying asset, the moat survives:

network effects;
scale economies;
cornered resources;
proprietary behavioral data;
customer-built workflow embedding.

If the stronger model improves the product because the company has private feedback loops competitors lack, the moat compounds:

accepted/rejected outputs;
customer-specific workflow traces;
vertical edge-case test suites;
integrations with niche systems;
APIs and automations customers build on top.

Bottom Line#

The moats that survive frontier-model improvement are not "AI" moats. They are old structural moats rebuilt with AI-native acquisition and compounding loops.

The durable pattern is:

Use The AI-Native Safe-Choice Inversion, Narrow Wedge into a Legacy Market, and Product Velocity as Moat to land customers; use Compounding Data Moat and the persistent powers in Seven Powers Applied to AI to make those customers hard to copy.

Anything else is probably a treadmill.