The Business Moat After AI

Imagine two small businesses with access to the same capable model. Both can draft a proposal, summarize documents, create a product page, and answer ordinary questions. One consistently wins useful work at a sustainable margin. The other struggles. The difference cannot be explained simply by saying that the first business “uses AI.”

A business moat is an advantage competitors find difficult to reproduce or overcome while the business continues delivering value. It can arise from costs, capabilities, relationships, rights, scale, or a particular combination of them. An impressive demonstration is not necessarily a moat. Neither is a complicated system that customers do not need.

AI changes the moat question by reducing the difficulty of reproducing some visible work. A rival may imitate a page, a feature, or a generic service explanation more easily. That puts greater weight on the parts the imitation does not capture: verified outcomes, specific data, reliable execution, distribution, integration, and an operator’s ability to learn. Which of those actually protects a business requires a concrete test.

Begin with the rival’s simplest route

A moat analysis should start with a capable competitor trying to serve the same customer. What would that competitor need to reproduce, and what could it avoid?

A business that has spent years creating custom software may assume the software is the barrier. The competitor might use an ordinary tool and a better process. A business with a large archive may assume the archive is essential. The competitor might answer the customer’s narrower question with a smaller, more accurate collection. A business with expensive premises may face a rival that delivers the service remotely.

This exercise is valuable because owners naturally see the work they have invested. Customers see the result they need. The competitor only needs enough capability to meet that need at acceptable quality and cost. It does not need to duplicate the owner’s biography, sunk expenses, or preferred architecture.

Define the job clearly. “AI support for businesses” is too broad. “Resolve ordinary warranty questions for this product family using verified order and policy records” is specific enough to examine. Once the job is clear, the analysis can separate necessary assets from historical clutter.

A model advantage can be real and temporary

A specialized model may perform a task better than a general tool. Access to scarce compute or training expertise may matter. It would be wrong to declare every model interchangeable simply because common drafting has become easier.

The durability question is what happens when a rival gains similar capability. Does the business retain a useful customer relationship, proprietary outcome history, distribution route, or lower-cost operating process? Or does the advantage disappear with the performance gap?

The OECD’s 2025 analysis of AI infrastructure examines competition concerns around inputs such as advanced chips and computing resources. Cheap downstream access can coexist with concentrated upstream infrastructure. That means operators should examine both model substitutability and supplier dependence instead of assuming universal commoditization. OECD, Competition in AI Infrastructure.

For a small business, a sensible response is to keep the model advantage where it is useful while investing in complements that remain useful if the model changes. An evaluation set of real customer problems, lawful outcome records, and a reliable delivery process can help the operator select and use a different model later. These assets may also improve today’s service.

Data becomes a moat through a decision

Proprietary data matters when it changes a decision in a way a rival cannot cheaply reproduce. Its value does not depend mainly on the number of rows or the word “proprietary.”

Consider a hypothetical repair service that records symptoms, diagnostic steps, parts installed, actual labor time, and whether the repair held after a follow-up period. A useful record could help estimate a new job, detect a recurring fault, or avoid an unnecessary part. The moat would be the combination of specific records and the ability to turn them into better service.

A competing firm might still reproduce the benefit by hiring experienced technicians, partnering with a manufacturer, or collecting its own evidence. The existing data creates a head start, not an eternal barrier. If equipment changes, old observations may become less useful.

Data can also impose costs. Keeping sensitive records requires appropriate permissions and safeguards. Poorly labeled cases can produce wrong recommendations. A history containing only accepted jobs may omit the difficult cases the business declined. The moat test asks whether the data improves an outcome enough to justify those costs.

This is the distinction developed in ownership in the AI economy. Possessing a record and deriving a defensible advantage from it are separate achievements.

Workflow integration can make replacement costly

A tool becomes harder to replace when it fits into a customer’s actual work. An assistant that merely drafts text may be replaced easily. A service that correctly handles identifiers, permissions, exceptions, handoffs, and records can become more useful over time.

Integration produces value by reducing repeated coordination. The customer does not need to re-enter the same information, reconcile inconsistent records, or teach a new supplier every ordinary case. These saved efforts may support retention even when a rival has a similar model.

However, switching costs have two moral and commercial forms. One comes from useful accumulated work: a well-maintained service history, familiar procedures, and trusted support. The other comes from obstruction: inaccessible exports, unnecessary complexity, or contractual traps. Both may discourage exit, but only the first aligns retention with continuing value.

A business should be able to explain its integration advantage without boasting that customers cannot leave. “We know the workflow and preserve its records” is a more durable claim than “migration is painful.” If a customer decides to leave, a responsible transition can preserve reputation and future referrals.

Reliability lives in difficult cases

Many demonstrations show the ordinary path. The customer asks a common question; the tool answers; the work appears complete. A moat often appears in the cases that do not fit.

A shipment arrives damaged. Two records disagree. A warranty has an unusual exclusion. A customer needs an accommodation. The supplier is late. The operator must recognize the exception, preserve the relevant evidence, and arrange a useful remedy.

A rival can imitate the interface quickly while lacking the capacity behind it. Reliable exception handling depends on clear authority, access to the correct records, practiced procedures, and people able to judge when the procedure is inadequate. AI may help classify or investigate the case, but reliability comes from the entire system.

This does not justify overbuilding for every imaginable event. Identify the exceptions that materially affect service or loss. Keep their history and test their handling. A handful of carefully designed cases can reveal more about a claimed advantage than a large collection of easy examples.

Distribution and trust are complements, not substitutes

A business with strong reach can introduce a new offer cheaply. A trusted business can reduce the buyer’s checking burden. Together, those capabilities can make an otherwise ordinary product easier to sell and support.

Yet both can weaken. An audience can be poorly matched to the offer. A reputation earned in one category may not transfer to another. A business known for careful educational content may lose credibility when it makes exaggerated commercial claims.

Distribution concerns the route to appropriate attention. Trust concerns justified expectations and recourse. A moat analysis should identify their separate contributions. Which customers arrive because of the channel? Which proceed because they have evidence about the business? Which return because delivery was satisfactory?

The distinction helps allocate effort. If interested buyers cannot discover the offer, improve access. If they discover it but need repeated reassurance, improve evidence and explanation. If they buy once and do not return, investigate the result before expanding the channel.

Scale can help or become a burden

Scale may reduce unit costs, broaden an evidence base, improve purchasing terms, or support specialized staff. It can also create coordination costs, rigid processes, and slower responses. A large organization is not automatically protected from a small rival using capable tools.

A hypothetical document-processing service might spread its infrastructure cost across many clients. That creates an advantage if the clients’ needs are sufficiently similar. If each client requires extensive custom work, volume may increase exception handling faster than it reduces unit costs.

AI can change both sides. It may make adaptation cheaper, strengthening scale. It may let a small operator provide a previously expensive capability, weakening a large firm’s advantage. The answer depends on the task, data, quality standard, and responsibility for mistakes.

Measure the relevant unit. Cost per accepted case is more useful than cost per generated response. Include review, rework, integration, support, and the expense of failed cases. A supposed scale moat can disappear when the business includes the costs its dashboard omits.

A local advantage can remain difficult to digitize

Some business advantages depend on proximity and physical capability. A person who can inspect an item, make a repair, deliver a needed part, or coordinate locally may offer something a remote model cannot supply.

The advantage is still specific. Being local does not guarantee skill or value. A customer may prefer a cheaper distant alternative when timing and inspection do not matter. The operator must identify the situations where location changes the outcome.

Salars’s local advantage economy explores this in grounded commercial terms. For the AI moat question, the relevant mechanism is complementarity: cheaper analysis may improve the local operator’s estimates, scheduling, and explanations while the physical service remains scarce.

A hypothetical rural equipment service could use AI to organize repair history and prepare parts questions. Its advantage would still depend on reaching the site, understanding the machine, carrying appropriate equipment, and doing the work safely. The tool broadens capacity; it does not move the truck or assume the operator’s obligations.

Stress-test a claimed moat

A moat should survive an unfavorable question. Assume that the rival obtains the same model at the same price. What remains? Assume that the main traffic source becomes less effective. What remains? Assume that the largest customer leaves. What remains?

The answer need not be impressive at first. A young business may have little durable advantage. It can still compete through disciplined execution, a narrow offer, and a low-cost learning process. Pretending the moat is already deep prevents the operator from building the capabilities it needs.

Use a short evidence table:

The table is an operating aid, not a valuation formula. It forces each claim to identify what would support it and how it might fail.

Learning speed can matter more than a static feature

A rival may copy what the business offers today. It is harder to copy a working process that repeatedly discovers customer needs, tests an improvement, records the result, and changes the offer responsibly.

This learning advantage requires more than frequent experimentation. A business can run many tests and retain little useful knowledge if it changes several variables at once, forgets the baseline, or records only favorable results. It can also learn the wrong thing by measuring convenient proxies rather than accepted customer outcomes.

A useful process preserves the question, expected mechanism, comparison, budget, result, and decision. It distinguishes a measured improvement from a plausible explanation. When the evidence is weak, the business can keep the conclusion narrow and choose a better test.

Learning also requires authority to stop. A project that consumes attention indefinitely can damage the advantage it was supposed to build. The kill engine provides a place for explicit stopping rules; the moat analysis asks whether the learning process produces assets worth maintaining.

The strongest objection: perhaps no moat is needed

A small business may earn a useful living without a formidable barrier to competition. Its market may be local, fragmented, or too specialized to attract aggressive rivals. Good service and reasonable costs can be enough.

That objection is sound. Moat language can make ordinary enterprise sound like a competition to become a monopoly. The more useful question is whether the business can continue delivering value at sustainable economics as conditions change.

A modest advantage maintained carefully may be better than an expensive attempt to manufacture exclusivity. A repair shop does not need a unique frontier model if its customers value dependable work and a reachable operator. A publication may benefit more from clearer navigation and accurate articles than from a proprietary recommendation engine.

The moat framework earns its place when it helps the owner identify a real vulnerability or a productive investment. If it becomes an excuse to collect unnecessary data, trap customers, or overbuild software, it has stopped serving that purpose.

The operator should preserve the rejected alternatives too. They explain why the current advantage was worth building and prevent the next review from repeating an expensive dead end.

A departing customer can supply useful counterevidence. Ask what alternative they chose and which switching difficulties were genuinely valuable context rather than obstruction. Record the answer without assuming every departure reflects a flaw or every retention reflects satisfaction. That evidence can reveal whether the claimed moat is better service, habit, limited alternatives, or something the business has misunderstood.

AI Leverage in Practice

What changed: similar cognitive tools are available to more rivals. Some visible output is easier to imitate, while infrastructure, real-world evidence, and reliable execution can remain difficult.

What to do today: state one claimed advantage in a sentence tied to a customer outcome. Identify the evidence supporting it, the simplest rival response, and the cost of maintaining it. Test the claim against one adverse scenario.

Choose the complement that most improves actual delivery. Build a better outcome record, repair an integration, make an exception process reliable, or create a more useful route to the offer. Set a completion test that a customer or operating record can confirm.

Repeat the analysis when a major model, channel, supplier, or customer changes. A moat is maintained through changing conditions. A document describing last year’s advantage does not establish this year’s.

What may come later: new tools can make previously difficult tasks easier and move competition toward new complements. Treat that as a reason to monitor the business’s bottleneck and retained assets, rather than predict one permanent winner.

An advantage worth maintaining

The business moat after AI is whatever still lets the business serve a worthwhile need better, more reliably, or more economically when others can buy similar tools. It may be a specific dataset, a practical integration, a physical capability, a trusted relationship, or a disciplined combination.

The test is concrete. Can the owner explain the advantage, show what it changes, and maintain it without misleading or trapping customers? Can the business adapt if one input becomes ordinary?

A strong answer does not need grand language. It needs evidence of useful work that the business can keep doing.

Return to The Age of AI Leverage, or explore the AI section.

Sources

• OECD, Competition in Artificial Intelligence Infrastructure, 2025.
• FTC, Staff Report on AI Partnerships & Investments, January 2025. Background for examining supplier dependence rather than assuming complete model interchangeability.