Trying to keep up with the pace of change in software in the world of AI is dizzying. Whether it is new drops of state of the art models, new harnesses that extend what you can do with AI, AI native products, or new products and features that have been built over a weekend, it is a lot.
As a product guy, I have spent a lot of time thinking about what every product leader and company should be asking right now: how can I use AI to build a moat and drive real competitive advantage?
Most people think they know how they should be using AI to do this. I certainly did when I started. But I was wrong. Adding an AI layer into your product or using it to automate some of your most time consuming tasks intuitively makes sense, right? I thought that if I added AI into my product or service, I could do more with less, automate things that I had never been able to automate before, and this would mean more money for me.
But here is what I missed. For this to be true, we can't be AI washing or using AI just for the sake of it. One of two things needs to be true:
I can use the time freed up and the tools to leverage my existing distribution channels or create new distribution channels to increase revenue.
My use of AI is able to drive a competitive advantage and increase my moat, locking people into my product or service and driving growth.
This isn't a discussion on distribution, so for this piece I want to focus on the second point.
The SaaSpocalypse is an example of investors not believing in the ability for Software as a Service companies to maintain their competitive advantages when the barriers to building competitive products or in-house solutions have shrunk to almost nothing. Previously, attempting this would require a skilled engineering team, great product management, and potentially years of effort to get the product right.
So before looking at competitive advantage in the age of AI, it is probably worth looking at what a moat actually is.
A company's moat refers to its ability to maintain the competitive advantages that are expected to help it fend off competition and maintain profitability into the future. Whilst this concept was popularised by Warren Buffett in his famous shareholder letters, when investors look for moats, they are usually looking at the five sources that Morningstar defines:
Intangible Assets – things like brand power, IP, proprietary data and knowhow
Switching Costs – how locked in are customers and how hard is it for them to move away, taking into account not just cost but time, effort, and risk
Network Effect – does each additional user, business, or partner increase the value of the product by growing the network?
Cost Advantage – can you undercut your competitors whilst maintaining your profit margins?
Efficient Scale – does the company own the industry so completely that there is no point in competitors entering the market?
From these, it is clear to me that AI is drastically eroding three of these five moats.
Switching costs? AI commoditises expertise and reduces the uncertainty of every meaningful barrier to switching like time, effort, and risk.
Intangible assets? Patents on prompts are worthless and switching to the knowhow of AI removes any institutional knowledge.
Cost advantage? Everyone accesses the same APIs at scale, so everyone can have ChatGPT in their product.
So if AI erodes traditional moats at an unprecedented rate, where does that leave us?
The problem isn't that moats are disappearing. It's that the kind of moat that exists around our intangible assets has changed. The old moat was your knowhow, your algorithm, your secret sauce. The new moat is your ability to control your data, your governance, and how you deploy AI against what you uniquely know.
In the following sections, I'm going to explore some areas where AI is creating pressure on traditional competitive advantages, and some ways you might be able to protect yourself and your company.
Regression to the Mean
This is particularly easy to see when it comes to content—advertising, positioning, copy, or any other creative output that your team produces. As more and more people use the same set of base models to do similar things, we are going to get similar results. Research by Graphite shows that around 50% of articles on the internet are primarily AI generated.
AI generated content carries telltale signs of heavy emoji use, bullet lists, and em dashes (yes, I wrote this one, but got it AI proofed). We can of course use custom instructions and other context to augment the AI output to be more like what we might come up with, but the value of individuality and creativity could be getting lost.
This is highlighted by a fascinating paper on cognitive debt in AI essay writing. The data shows homogenised structure, styles, and even topics among LLM users, whereas the brain-only and search engine groups had distinctive writing styles. Human assessors were able to identify who wrote what based on their distinctive work.
Does the creativity of your people create a distinctive edge for you and your business? If the answer is yes, it might be time to start putting in place some things to help you retain this value rather than surrendering it to speed. Here are a couple of ideas:
Let your team know how important their individuality is for your business and ask them to prioritise this over speed of delivery.
Put in place some processes that foster creativity. This might be spending the effort designing and planning first, then using AI as a co-pilot to complete it. Or it might be crafting it by hand before using AI to refine the output.
The companies that stand out aren't the ones that sound like everyone else. They're the ones that sound like themselves.
Common Base Model
As the majority of businesses converge on using the same hyperscaler base models (ChatGPT, Claude, Gemini), we are effectively all using the same underlying data to do our work. It is easy for us to look at this and think that we are growing our competitive advantage because we are able to streamline some parts of our work, reducing costs or increasing throughput. But we could accidentally be surrendering parts of our expertise to exactly the same base model as our competitors are using.
Let's take an example of underwriting in banking and insurance. For decades, these institutions have built institutional IP in data, people, and algorithms that they have leveraged in their underwriting processes to determine who they underwrite in order to maximise their profit while remaining competitive in the market. This institutional knowledge is the "secret sauce" to the success of these companies, having been built from hard won experience.
Here's the problem: as AI is being adopted into these processes, the underwriting process is becoming a prompt with a thin context layer and some tooling, with the heavy lifting being done by the LLM. This is incredibly easy for any competitor to replicate. Offloading all the judgement work to the same base model that your competitors are using is potentially throwing your moat away and outsourcing it to a commodity product everyone on earth has access to.
But we really don't want to stop using AI. The genie is out of the bottle, so we need to know how to use AI to preserve our unique value.
Instead, we can use AI to build repeatable processes that embed your expertise, not to replace it. Use AI to speed up the creation of repeatable, deterministic processes that create and preserve value in your business. AI has taken off in development because it is able to dramatically speed up the creation of software into a testable, verifiable output that can be used over and over again without the use of AI, using the intelligence and speed of AI development once for something that can be repeated N times. For our underwriting example, this means using AI in the right places—not to make the judgement call on who to underwrite, but to help you create models that can make this decision, log any exceptions, and develop the model further as you learn more.
Renting Intelligence
Most of us are renting intelligence. We are paying per token to get knowledge work done. Renting can be great, just like renting a house. If you're not ready to commit to where you want to live, you can pay to use someone else's property and try it out. There are some great advantages to this:
- Try before you buy
- Lower upfront costs
- Someone else is worrying about the maintenance and upkeep
But there are also some real downsides:
- Costs can increase unexpectedly
- Sufficient demand could mean no more capacity is left
- Every dollar that you are spending is not going into your "equity"
These are exactly the same whether it is a house, AI infrastructure, or cloud computing. If your goal is to build an asset or competitive advantage with AI, then it might be better to own your own intelligence and not expose yourself as much to the price of a token or access to specific models and providers. We saw with the export controls imposed by the US Government on Anthropic's latest model in July 2026 that our access to private American AI company models is not guaranteed. This model was 2x more expensive than their previous most expensive model on release, and also came with a new tokeniser that meant it consumed 30% more tokens on like-for-like requests. The costs ramped up significantly.
But there is also another problem. Picture this: 100+ million people using standalone AI platforms every day. 1+ billion more interacting with AI every month through digital products and services that they use, and generally all through the same model providers.
With that kind of scale of data, just imagine what you could learn. According to Anthropic, whilst your data is not used in training, they still maintain "a system that enables us to understand important trends and behaviours".
This matters because model providers can now see where demand is building and build products to capture it. Anthropic has been releasing products such as Claude Design, Claude Legal, and Claude Code, which all directly compete in a vertical. With almost infinite resources, the model providers are able to do this across any vertical in any country. It is easy to see a scenario where 100 banks are running their risk assessments through the same models as each other, and the model providers are able to spot patterns in how they are doing this, giving them an insight to out-compete you or selling this to your competitors. You might just be handing over the keys to your castle without knowing it.
So you're leaking your patterns to vendors with infinite resources, paying per token with no equity, and locked into their pricing and roadmap. This is the real cost of building on shared infrastructure.
How to Protect Your Edge
The points I have explored so far have been mostly about how to apply AI in circumstances that might already be present in your business—people using AI to do their work, using AI tools to make judgement calls.
But if you're big enough, or if this is important enough, there is something else you can do. A lot of these issues arise from using third parties' proprietary models and systems. Having custom AI models can change the game because you own everything: the output, the improvement, and the moat.
At the moment, we are all using frontier LLMs. These have knowledge of everything that humans have ever published on the internet and because of this are able to solve for things across any area of knowledge you can think of. But apart from a small bit of context they learn about you as you use them, your proprietary data, IP, and knowhow are not embedded in them. Building a custom small language model that is an expert in something important to you would allow you to embed that proprietary knowledge and have it influence the responses you get.
We are also at an interesting moment in time with the release of open source models like Kimi K3 that have near frontier reasoning capabilities. This means you can have your own models with near state-of-the-art performance, hosted on your own servers and fine tuned for your data, without exposing your usage to anyone else.
There is of course a tradeoff with custom models: upfront costs of hosting rather than on-demand charges, and not necessarily being able to be at the frontier with your custom models as new frontier models get continuously released.
The Choice
Most companies building with AI today are not building competitive advantage. They are renting commoditised intelligence and calling it strategy. They are leaking their proprietary patterns to model providers with infinite resources to exploit them. And they are betting their long-term economics on the benevolence of a handful of American companies.
Here is your choice: You can continue renting intelligence and hoping your thin prompt layer differentiates you. Or you can start thinking about AI differently—as a capability layer on top of what you uniquely know, not as a replacement for it.
The companies that will win in the next decade won't be those with the best prompts. They'll be those who treat their data, their governance, and their deployment of AI as the real moat. They'll be the ones who own their intelligence instead of renting it.