AI Moats: From Software Advantage to Operating Advantage

No, it's not AI generated. I actually spent last weekend at 'Moat Cottage' at Wilderness Reserve in Suffolk (UK) and, although I was supposed to be relaxing, it got me thinking about moats in the age of AI. 

Every major technology shift begins with the wrong question.

When the internet arrived, companies asked whether they needed a website. When cloud arrived, they asked whether they should move their servers. When mobile arrived, they asked whether they needed an app.

Now, with AI, the question is: “What tool should we use?”

It is the wrong question again.

The better question is this: what becomes defensible when intelligence is cheap, software is easier to build, and every company has access to broadly similar foundation models?

That is the real question behind moats in the age of AI.

A moat is not a feature. It is not a good demo. It is not a clever prompt. It is not a temporary lead. A moat is an advantage that remains when competitors notice what you are doing and try to copy it and they struggle to replicate it, or overcome it.

That distinction matters because AI is changing the economics of software. It is becoming easier to build products, easier to automate tasks, and easier to create something that looks impressive in a demo. At the same time, enterprise demand is no longer theoretical. Menlo Ventures estimates that enterprise generative AI spend reached $37 billion in 2025, with the application layer alone capturing $19 billion. McKinsey’s 2025 global survey found that 88% of organisations are now using AI in at least one business function, but most have still not embedded it deeply enough into workflows and processes to realise material enterprise-level benefits. (Menlo Ventures)

That is the paradox.

The opportunity is large. The adoption curve is real. But many of the products being built today are not very defensible. They are interfaces. Wrappers. Prompts. Copilots. Assistants. Chatbots. Agent demos. Useful, perhaps. Valuable in some cases. But not necessarily durable.

As a Co-founder of Implement AI, defensibility is something I think about, even on a weekend break! The test is simple: does the advantage survive when models get better, cheaper and more widely available? If the answer is no, it is probably not a moat.

I wrote a newsletter on the end of competitive advantage in February 2025. 

The Moat Is Moving From Software To Doing Work

For the last two decades, the strongest B2B software companies were built around familiar sources of defensibility: systems of record, switching costs, workflow depth, integrations, proprietary data, procurement trust, brand and distribution.

Salesforce owned the customer record. Workday owned the employee record. ServiceNow owned operational workflow. HubSpot owned the marketing and sales motion for a generation of growing companies.

AI does not make these advantages disappear. But it does change where the centre of gravity sits. The old software moat was often about where information lived. The new AI moat will increasingly be about where work gets done and how AI-driven solutions are implemented and deliver measurable ROI.

That is a profound shift. Traditional SaaS mostly helped people record, organise, search and report information. AI-native systems increasingly perform the work itself: qualifying leads, resolving tickets, processing documents, analysing calls, reconciling invoices, updating systems, escalating exceptions and coordinating activity across tools.

Andreessen Horowitz has framed this as software moving from seats to outcomes, with AI agents accelerating the long shift in software from recording work to performing it. Bain has made a similar point in the context of SaaS, arguing that generative and agentic AI can automate tasks, replicate workflows, and push software companies towards owning data, setting standards and pricing around outcomes rather than log-ons. (Andreessen Horowitz)

This is the move from software advantage to operating advantage.

A company does not buy an AI system merely because it is clever. It buys it because it can accelerate revenue, increase capacity, reduce friction, improve speed, lower risk or create measurable value. Once software starts doing work, the moat moves closer to the operating model.

A Practical Guide to AI Moats

The market often talks about “AI moats” as if there is one answer. There is not.

There are several possible moats. Some are strong. Some are weak. Some are useful but insufficient. The strongest AI businesses will usually build a stack of moats rather than rely on one and some will be more robust than others. 

What I am confident about is this: the future will not be dominated by a handful of frontier model providers, because the underlying intelligence layer is already fragmenting, commoditising, and specialising. Plus distribution (see below) is key. 

The Model Moat

If a company owns a frontier model, trains it at massive scale, controls specialised infrastructure and has access to scarce technical talent and capital, that can be a serious advantage. However, this is not the moat most B2B AI companies will have. Most application companies rent the model layer. They may use OpenAI, Anthropic, Google, Meta, Mistral or an open source model, especially if frontier capability is not required. They may switch between models depending on performance, cost, latency or customer requirements. 

That flexibility is useful, but it is not defensibility. “We use AI” is not a strategy. “We use the best model” is not much of a strategy if competitors can use the same model next week. "We have fine-tuned and manage a model that has been trained on your customer data and operating model and culture" may be more of a moat and one built on delivery as much as the technology.

The Interface Moat

A better user experience matters. A clean workflow, a well-designed dashboard, a natural conversational interface or a simple onboarding journey can help a company win early customers. But interface alone is now just not enough. AI has made software easier to build. It has also made interfaces easier to copy. A feature that feels distinctive in January may be table stakes by June. A chatbot, document summariser, copilot or voice agent may win attention, but attention is not the same as protection. The interface is often the doorway. It is rarely the moat.

Claude Design has just been launched at the time of writing, and it is disruptive not just because it replaces design tools and their interfaces with conversation, but because it compresses the entire design workflow into a single loop where intent becomes execution. 

The Workflow Moat

This is where things become more interesting, as a workflow moat exists when a product becomes embedded in how a company actually operates. It is not used occasionally. It becomes part of the rhythm of work. This is why enterprise software has always been more defensible than it looks from the outside. The value was never just the code. It was the process encoded into the software. a16z argues that the bear case against software misunderstands where value lives: code has never been the whole value of software companies, otherwise they would have been commoditised long ago. (Andreessen Horowitz)

AI makes workflow ownership even more important. A tool that answers a question can be replaced. A system that sits inside sales, support, finance or operations and coordinates real work is harder to remove. It understands what happens first, who approves what, which system must be updated, where exceptions go, what should never be automated and how success is measured. That is the difference between a clever assistant and an operating layer.

The Context Moat

Generic intelligence is becoming abundant. Business-specific context is not. A context moat exists when an AI system understands the real operating state of a customer’s business: its customers, policies, documents, systems, permissions, exceptions, tone, history, commercial rules and institutional memory. 

This is not just “data”. Data alone is often messy, stale or unusable. Context is data with meaning, structure, permissions and relevance to a task. A generic model may understand sales in general. That is not the same as knowing how a particular company qualifies leads, which customers are high priority, what language the brand uses, what discounts are allowed, which accounts need human attention, its cultural mores, and what risk signals matter. Competitors cannot easily copy that from the outside.

In the model moat section I used the example of fine-tuned models that add material value and which can 'learn' over time and be used by AI agents within the business to improve performance and reduce hallucinations. 

The Execution Moat 

There is a major difference between AI that suggests and AI that acts. A system that drafts an email is useful. A system that researches the account, drafts the email, checks the CRM, personalises the message, routes it for approval, sends it, updates the record and learns from the response is much more valuable. 

The execution moat appears when AI is not simply producing outputs but completing work across systems. This requires integrations, permissions, approval flows, monitoring, exception handling and reliability. It also requires clear boundaries around what the AI can and cannot do. Many AI products will fail here. They can generate. They can summarise. They can answer. But they cannot reliably operate inside the messy environment of a real organisation - yet.

The Trust Moat

The more powerful AI becomes, the more valuable trust becomes. In B2B markets, capability is not enough. Enterprises buy risk reduction. They need to know what the system did, why it did it, who approved it, what data it used, what rules constrained it and how mistakes can be corrected. 

Trust is not a brand slogan. It is architecture. It means permissions, audit trails, role-based access, human-in-the-loop controls, policy libraries, escalation paths, monitoring, evaluation, security and clear accountability. As more data is used by agentic systems, it also increasingly means external validation through recognised standards and regulation such as ISO 27001, SOC 2, GDPR compliance, and adherence to frameworks like the EU AI Act.

Gartner has warned that more than 40% of agentic AI projects may be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. It has also warned about “agent washing”, where vendors relabel assistants, RPA tools and chatbots as agents without meaningful autonomous capability. (Gartner)

That warning should not be read as a rejection of agents. It is a rejection of theatre. The enterprise does not need more AI language. It needs systems that can do real work, safely.

The Implementation Moat

This is the one most people miss. 

Deploying AI into a real organisation is not just a software problem, but an operating model, or business transformation, problem. The technical challenge is only one part of it. The harder problems are usually process ambiguity, fragmented systems, poor data quality, edge cases, compliance, adoption, incentives, decision-making authority, employee fears, and organisational trust. This is why so many AI pilots never scale. The pilot proves that something is possible. Production requires proving that it is repeatable, safe, measurable and adopted.

My experience is that many business leaders still view AI as another IT or technology project, when in reality ongoing implementation and optimisation will become a permanent, feature of operations and competitive advantage.

Implementation know-how compounds because every deployment teaches what breaks. It shows which use cases are worth automating, where failure modes live, which integrations matter, how humans should supervise agents, how quickly autonomy can be expanded and how value should be measured. But there is a critical distinction. Implementation is only a moat if it becomes codified. If it remains trapped in the heads of a few experts, it is skilled services. Valuable, but not deeply defensible. If it becomes frameworks, playbooks, reference architectures, test suites, governance templates, rollout methods and structured feedback loops, it becomes institutional capital. 

In the AI era, implementation is not the work around the product. It is part of the product.

The Outcome Data Moat

Many companies talk about data moats. Fewer have them. I spent years selling cloud when 'big data' was the next tech hurdle to overcome. A large dataset is not automatically a moat. If the data is generic, stale, unstructured, easily licensed or not tied to outcomes, it may not defend much at all. The data that matters most in AI is outcome data. 

Which recommendation was accepted? Which lead converted? Which support answer resolved the issue? Which workflow saved time? Which decision required escalation? Which response created risk? Which action did a human override? Which agent performed better in which situation? This data compounds because it connects action to result. A shallow AI tool may only see prompts and responses. A deeply embedded AI operating layer sees context, action, approval, exception, outcome and improvement. That creates a learning loop: more work creates more data, more data improves performance, better performance earns more trust, and more trust allows the system to handle more work.

The Distribution Moat

Even in the age of AI, it is a mistake to underestimate the power of distribution and ownership of corporate relationships. AI has made software easier to build. It has not made enterprise distribution any easier. 

Selling into businesses still requires trust, timing, credibility, procurement navigation, stakeholder alignment and a clear route into existing workflows. This is especially true when the product touches sensitive data or performs operational work. B2B sales are inherently complex, involving multiple stakeholders and long decision cycles built on relationships and proven value.

A distribution moat exists when a company has privileged access to buyers, users, partners or ecosystems that competitors cannot easily replicate. Frontier model companies often appear intent on selling directly into organisations, but that underestimates how entrenched enterprise relationships are. IT and supplier relationships can span decades, and trusted advisers are unlikely to recommend technologies they do not understand or that do not align with their commercial incentives.

That access typically comes through channel partners, managed service providers, system integrators, vertical expertise, comms dealers, marketplaces or existing workflow platforms. In a noisy and crowded market, distribution becomes more valuable, not less.

The Standards & Category Moat

New markets are confusing. Buyers do not just need tools. They need language, benchmarks, risk models, deployment frameworks and decision criteria. The company that helps define what “good” looks like can become more than a vendor. It can become the frame through which the buyer understands the category. In traditional SaaS, the prize was often to become the system of record. 

In AI, the prize may be to become the trusted standard for how digital work is deployed, governed, measured and scaled.

Other Moats?

There are also weaker moats that matter but should not be mistaken for structural protection. Speed matters, because it helps a company capture a market window and learn faster. Funding matters, because it buys time, talent and distribution. Feature velocity matters, because customers want to see progress. But none of these is enough on its own. They are accelerants. 

They help a company build a moat. They are not the moat itself.

The strongest AI companies will use speed, capital and product velocity to build deeper advantages: context, workflow ownership, execution, trust, implementation capability, outcome data and distribution.

The Uncomfortable Conclusion

AI will make software easier to build and it will not make durable companies easier to build.

In fact, it may do the opposite. When features are easier to copy, the value shifts to what is harder to copy: workflow ownership, operational context, trusted execution, implementation capability, proprietary outcome data, distribution and the ability to deploy reliably inside complex organisations.

The moat is not the model. The moat is not the prompt. The moat is not the interface. The moat is not the demo. The moat is the system that turns intelligence into reliable work.

That is the strategic shift. Software advantage was about owning the record, the workflow or the user interface. Operating advantage is about owning the method, context, controls, fine-tuned intelligence, and learning loops that allow AI to perform work safely and repeatedly.

Most companies will buy AI tools. Some will redesign work.

A smaller number will build the operating layer that others come to depend on.

That is where the next generation of moats will be built.

Thank you for reading.