Where AI Creates Durable Advantage: AI Explained

Layer 1: Surface

If your AI advantage is “we use a frontier model and our competitor doesn’t,” you have a six-month head start, not a durable advantage. Foundation models are available to everyone with a credit card. Any capability you build on top of a public API can be replicated by any competitor with the same access.

Durable competitive advantage from AI comes from four sources, and only one of them is the model itself:

Proprietary data: You have data that others cannot replicate: transaction history, user behaviour, proprietary labels, domain-specific records built up over years. A model trained or adapted on that data can do things no competitor can match until they acquire similar data.
Workflow integration: Your AI is so deeply embedded in how your operations run that replacing it would require restructuring how work gets done. This is sticky not because the AI is hard to replace technically, but because the processes built around it are.
Network effects: Your AI improves as more users interact with it: more queries, more feedback, more corrections. A competitor starting from zero has worse AI not because their model is worse, but because their training signal is weaker.
Speed of iteration: Your organisation can ship, measure, and improve AI faster than competitors. This is an organisational capability, not a technical one. It requires strong evaluation practices, a deployment culture, and a feedback loop from users to engineers.

The honest question to ask of any AI investment: if a well-resourced competitor called the same API providers tomorrow and matched our features in six months, what would we still have that they don’t?

Why it matters

Organisations that mistake temporary differentiation for durable advantage over-invest in features and under-invest in the infrastructure, data, feedback loops, operational integration, that actually compounds.

Production Gotcha

Common Gotcha: The most common strategic mistake is investing heavily in AI features that are immediately replicable by any competitor with API access, while under-investing in the proprietary data infrastructure and feedback loops that would create a real moat. The AI feature is the visible part; the data flywheel is the defensible part.

The assumption: “We built this feature and our competitors haven’t: that’s our moat.” The reality: they will build it in the next product cycle.

Layer 2: Guided

Evaluating an AI investment for durability

For any proposed AI investment, run it through this durability test:

Question	If yes →	If no →
Could a competitor with the same model access replicate this feature within 12 months?	Thin moat: the value is time-to-market, not durable advantage	The moat comes from below the model: data, integration, or network effects
Does the feature get better as users interact with it?	Network effect potential: invest in the feedback loop	Consider whether you can add a feedback mechanism
Is the feature deeply embedded in a critical workflow?	Integration moat: switching is costly regardless of AI quality	Feature is additive but not sticky
Does the feature depend on data only you have?	Data moat: invest in protecting and expanding that data asset	Consider whether proprietary data is achievable

The four sources of durable advantage: in depth

Proprietary data is the most fundamental moat. It takes three forms:

Exclusive data: Data you own that no one else has access to: historical transactions, proprietary measurements, confidential records. This is the strongest form.
Behavioural signal: User interaction data that accumulates as your product is used. A product with ten million users generating feedback signals has a training advantage over a new entrant.
Labelled domain data: Expert annotations of domain-specific decisions, correct outputs, or error corrections. This is especially valuable in specialised fields (medical, legal, financial) where labelling requires expensive expertise.

Workflow integration is underrated as a moat because it is not glamorous. When an AI feature is deeply embedded in how a team runs its daily operations, their approval workflows, their data entry, their reporting, it becomes difficult to remove even if a better alternative emerges. The cost of switching is not the technical cost of replacing the API call; it is the organisational cost of retraining people and restructuring processes.

Network effects in AI manifest when your model or system improves as a function of usage. Examples:

A recommendation system that learns from click-through and purchase data: more users → more signal → better recommendations → more users
A classification system where users correct errors: more corrections → better model → fewer corrections needed
A search system where queries and click patterns improve relevance: more queries → better ranking → more satisfied users → more queries

Not all AI features have inherent network effects. Building in a feedback loop, even a simple thumbs up/down or correction mechanism, can convert a feature without network effects into one that compounds.

Speed of iteration is the most actionable advantage for most organisations, because it depends on process and culture, not on proprietary assets. Organisations that can ship, measure, and improve AI features faster than competitors maintain an advantage that grows over time. The key capabilities:

Strong evaluation infrastructure (you know immediately whether a change made things better or worse)
Short deployment cycles (changes go to production in days, not weeks)
Production monitoring (you learn from real user behaviour, not just dev testing)
A culture of small bets over large projects (many experiments beats one big one)

AI as enabler vs AI as moat

Most AI is an enabler: it makes existing things faster or cheaper without creating a structural advantage. This is not bad; enablers that genuinely improve user experience or operational efficiency are valuable. But they should be evaluated on their operational ROI, not on their strategic value.

A moat requires one of the four sources above. If none of them apply, be honest: you are buying a capability improvement, not building a competitive barrier.

# A simple durability scoring heuristic — pseudocode
from dataclasses import dataclass

@dataclass
class DurabilityAssessment:
    name: str
    replicable_within_12m: bool   # can a funded competitor match it?
    has_network_effects: bool     # gets better with more users?
    workflow_embedded: bool       # operationally sticky?
    proprietary_data: bool        # depends on data only you have?

def durability_score(a: DurabilityAssessment) -> str:
    moat_factors = sum([
        a.has_network_effects,
        a.workflow_embedded,
        a.proprietary_data,
    ])
    if a.replicable_within_12m and moat_factors == 0:
        return "Thin moat: time-to-market value only"
    if moat_factors >= 2:
        return "Strong moat: multiple durable factors"
    return "Moderate moat: one durable factor, monitor competitors"

Layer 3: Deep Dive

The commoditisation curve

Foundation model capabilities follow a commoditisation curve. A capability that required a specialised model in year one becomes achievable with a prompted general model in year two, and is a standard feature in commodity SaaS tools in year three. Understanding where a capability sits on this curve is essential to timing investment correctly.

Investment in a capability that is about to commoditise produces a short-lived advantage at a high cost. Investment before commoditisation, in the proprietary data and feedback loops that will persist after the model capability is widely available, produces a durable advantage.

Practically: map your AI features against the commoditisation curve. Any feature on a third-party API that could be replaced by a standard SaaS feature in 18 months is not a strategic investment: it is a tactical one. Treat it accordingly.

The data flywheel: a worked example

Consider a business that builds an AI-assisted contract review tool. At launch, the AI uses a general foundation model and is not meaningfully better than a competitor’s equivalent tool.

Over 24 months:

Lawyers use the tool and correct errors, generating labelled examples of correct and incorrect clause identification
These corrections feed back into the model, improving accuracy on the specific contract types the firm handles
The improved model attracts more firms, generating more corrections
After 24 months, the tool is substantially more accurate on the firm’s specific contract types than any general-purpose tool: not because the base model changed, but because the feedback loop ran for two years

The competitor who enters in month 25 with a better base model starts from zero corrections. The data flywheel took time to build but is very difficult to replicate quickly.

Structural advantage in regulated industries

In regulated industries (financial services, healthcare, legal), structural advantage from AI often comes from compliance infrastructure rather than model quality. The first organisation to build a compliant AI workflow, with the audit trails, human oversight checkpoints, and regulatory approval, has a significant head start, because building that infrastructure is expensive and slow. The advantage is not the AI; it is the compliance architecture around it.

Where AI Creates Durable Advantage