For the owner of a product business who has started using AI and is asking the harder question: if a competitor can build the same stack next week, what is actually left that is yours.
The short answer. Not the software, and not the data sitting in it. Your moat is a learning loop that compounds over time: what you know about your customers, your products and your margins that took years to build, the relationships nobody can prompt their way into, and the accountability the machine cannot carry. The tools are a commodity now. What sits on top of them is not.
For thirty years the most expensive thing a business could do with software was leave it. The switching cost was the whole point. Your processes, your team's muscle memory and three years of workarounds were poured into the system like concrete, and pulling it out felt like open-heart surgery at full speed.
AI changes the economics of that lock-in. An agent does not need your data mapped into its own schema first. It reads your spreadsheets, your emails, your invoices and your accounting ledger directly, without the six-month integration project that used to be the price of admission. The system of record, the one everybody called the single source of truth, is quietly becoming a storage layer that something smarter reads from and writes back to.
For a product business at our scale the practical version is simpler. Shopify, Amazon Seller Central, Xero, whatever your 3PL calls its warehouse system. Each of those used to feel like a strategic commitment with a five-year shadow behind it. Now they are plumbing. The intelligence sits in the layer that reads from all of them, spots the pattern and acts on it. Who handles your invoicing is about as interesting as who supplies your electricity.
That sounds like freedom. And it is, until you notice the same freedom applies to every competitor across the road.
The first thing most businesses do when they see the old software losing its grip is run straight at the newest thing and wire everything to one AI provider's API. Every prompt tuned to one model. Every workflow dependent on one vendor's function calling format. The entire business logic living inside a system that somebody else owns, prices and can change on a Tuesday afternoon.
I have heard this described as digital autonomy. It is the opposite.
Traditional code runs the same way every time you call it. A language model is probabilistic, which means when the provider updates the model, deprecates a version or changes the terms, a process that ran perfectly last week can produce different answers today. And if your business only knows how to speak to one model, you have no negotiating position when the invoice goes up. What are you going to do, rebuild everything? That is the old lock-in wearing new clothes.
What does this mean in practice for a business with five or ten people? Keep your business rules in your own files, in a format any model can read. Use open integration standards where they exist, so the engine underneath is replaceable without rebuilding what sits on top. The work of switching a model should cost a day of configuration, not a quarter of engineering.
The models are commodities and they get cheaper every six months. Your own context, your own process logic and your own customer data are your capital. Keep them separate from the engine, or you will find out the hard way that you traded Walldorf for San Francisco and the rent is just as high.
Traditional SaaS software was an investor's dream because each additional user cost almost nothing to serve. Gross margins for mature SaaS companies sit around 80 per cent, which is why every venture fund loved the category. Once the code was written the money kept arriving.
AI does not work like that. Every reasoning step, every data retrieval, every chain of thought costs compute, and the bill comes in tokens. Route every routine task through the most capable model on the market and the compute bill will eat the productivity gains. Quite possibly within weeks.
The fix is routing, and it is not complicated. Three tiers, each at a different cost.
Deterministic code handles the strict rules. Matching a purchase order total against its line items, checking whether a margin falls below a threshold, validating a listing against a specification. This costs nothing, runs instantly and is right every time. No model needed.
Small specialised models handle routine classification. Sorting customer enquiries by type, extracting dates and amounts from emails, flagging an invoice that looks unusual. These run cheaply, often on your own infrastructure, which also means customer data does not leave your hands.
The expensive reasoning models handle genuine complexity. Synthesising a quarterly review from ten different sources, resolving an ambiguous supplier dispute, answering a question nobody has documented before. This is the work worth paying the premium for.
In my own stack most of the daily volume sits on the first two tiers. The expensive model handles a small fraction of the total, and the cost stays manageable because the work was sorted before it got there. I would like to say this was a design choice. It was not. We ran the expensive model on everything until the invoice landed, and then we learnt to sort. That is probably a more useful lesson than the tidy version.
If the tools are copyable, raw data is not that special, and the cost of building an AI stack drops every quarter, then what is left?
Michael Porter wrote in 1996 that operational effectiveness is not strategy. If every competitor in an industry optimises the same processes with the same widely available tools, the result is convergence. Everybody gets faster. Nobody gets ahead. The gains flow through to customers as lower prices. Running your operations with AI is becoming the new baseline, and having a baseline does not mean having an advantage.
And data sitting passively in a database is not a moat either. Transaction data ages fast, has weak legal protection in most places, and improving models need less of it every year to reach useful conclusions. So owning a pile of historical records is not the same as owning something defensible.
Four things are harder to copy than a tool stack.
Data is not valuable in storage. It is valuable when it moves. When the decisions your agents make in live operation, and the consequences of those decisions, continuously refine what the system knows about your specific products, your specific customers and your specific cost structure, you build something an outsider cannot replicate by hiring the same developer. The loop takes months and then years to produce its advantage, and that time is the moat. What does the loop know after a thousand real purchase orders that it did not know after ten? That answer is yours.
Internal efficiency is half the story. The real lock-in happens when your business is embedded in your customer's daily workflow deeply enough that switching away from you is disruptive, uncertain and just not worth the risk. This is as old as commerce. AI changes the tools you use to deliver the integration, but the principle underneath has not moved.
An agent can calculate, recommend and draft. It cannot sign a contract, carry the liability or pick up the phone when the delivery went to the wrong warehouse. What do you actually pay a premium for in B2B? Not the calculation. The calculation is free now. You pay because somebody has a name, a track record and a balance sheet behind the answer, and they are personally accountable when the answer turns out wrong. An AI stack with nobody standing behind it is a toy. One with the right person behind it is a business.
Hamilton Helmer calls this process power and considers it the rarest competitive advantage of all. It lives in culture and in the tacit knowledge between people, not in any procedure manual or prompt library. If a process can be fully described in a prompt, it can be copied with the same prompt. The real advantage sits in the questions you have learnt to ask, the instinct for which supplier to trust, the feel for when a margin looks too healthy to be real. Those things came from years of doing the work wrong before you did it right, and they do not transfer by copying a configuration file.
More relevant than to a larger one. A company with a hundred engineers can afford to rebuild if it picked the wrong stack. At GBP 3 million and three people, the wrong dependency can cost a quarter's margin before anybody notices. The architecture decisions, keeping your logic portable and your compute costs sorted, matter more when the budget is tight, not less.
Ask one question: if this provider doubled its price tomorrow, how long would it take you to move? If the answer is months, the architecture needs work. If it is days, you are in a reasonable position. The first step is separating your business rules and product data from the model layer, so the model becomes a replaceable component rather than the foundation. More on the practical side of this in what AI is actually for in a product business.
Simple rules for simple checks. Small cheap models for routine sorting. The expensive model only for genuine reasoning. Most daily work in a product business, matching invoices, classifying correspondence, monitoring stock levels, does not need the largest model at all.
No. The tools are the table stakes and falling behind on them is expensive. But invest knowing that the tools themselves will not protect you. What protects you is what sits on top: the process knowledge, the customer relationships and the learning loop that compounds quarter after quarter.
Anyway. The moat was never the software. Not thirty years ago when the enterprise bought SAP because leaving felt impossible, and not now when a startup builds everything on one model because it is too busy to think about portability.
The moat is what the people in the business know, who they know it for, and whether they are willing to carry the responsibility when the answer turns out wrong.
That part has not changed.
Michael Porter, "What Is Strategy?", Harvard Business Review, November–December 1996. Hamilton Helmer, "7 Powers: The Foundations of Business Strategy", 2016. SaaS gross margins: observable from public company filings, Salesforce FY2026 annual report (approximately 77%), Adobe FY2026 (approximately 88%). All retrieved 9 September 2026.