At a Glance
Labelling a strategy deck “AI” doesn’t make it a strategy. If your AI features are just a thin wrapper around a rented foundation model, any competitor can copy them in a weekend. Real defensibility comes from three things that have nothing to do with the model itself: the proprietary data you feed it, what it remembers about the customer between sessions, and how deeply it’s woven into your actual operational workflows. None of these moats are permanent either. They have to be continuously re-earned as the underlying models keep improving. AI is an operating imperative, but it can’t substitute for a real strategy built on solving a customer problem only you can solve.
I keep seeing the same slide. “AI Product Strategy” at the top, and underneath it, a weird little chatbot. Or a summarise button. Or a co-pilot bolted onto the side of an existing product. People say ‘phew’ because it feels like progress.
This isn’t strategy. There isn’t even a shallow moat. At best, it might be called an unwanted ‘feature’, and it’s not even hard to copy.
What we tend to forget is that our competitors also have access to the models doing the work in our products. OpenAI, Anthropic, an open-weights model. It really doesn’t matter which. If the AI powering our features is rented from a foundation model, so is theirs. We haven’t built anything they can’t build over a weekend with a decent prompt and a bit of orchestration glue.
That’s the whole problem with what’s been termed “AI wrappers.” Strip one back, and you find a thin interface, a prompt template, and a call out to someone else’s model. There’s nothing underneath that a competitor can’t stand up quickly, and unlike the software margins we got used to in SaaS, every one of those calls costs you money per use. That’s not a strategic differentiator. It can never be a scaling advantage, but tahdah, it can very much be a scaling cost. And because the intelligence isn’t yours, your product’s value is tied to another company’s roadmap and pricing decisions. When either of those changes, so does your feature.
Adel Smee, Head of Engineering at Up, put this sharper than I have on a panel we ran at LTP Events in Melbourne this year. She talked about her engineers running out of token budget every month because the pricing kept shifting under them, and about quoting a four-day company hackathon where the estimated cost from Google ranged anywhere from $200 to $10,000 depending which week she asked. Her point was that betting the farm on things staying the way they are is a terrifying place to build a business. Rent the intelligence, and you’ve rented the volatility that comes with it.
So merely labelling your strategy deck AI doesn’t make it a strategy. The question then is: how do we create that moat that our products need?
I think it comes down to three things, and none of them is about the AI model at all.
Start with what you feed it. A foundation model is only as valuable as the data you give it. If your product pipes years of clean, permissioned, genuinely proprietary data into the prompt, through RAG or fine-tuning or whatever mechanism, the output stops being generic. A competitor can copy your prompt in an afternoon. They can’t copy a decade of clean transactional history, or the telemetry you’ve been collecting since 2019.
Then there’s what it remembers about the customer. Foundation models are stateless. Ask one a question, it answers, it forgets. The moat is in everything you hold onto between sessions: the org structure, the history, the preferences nobody typed into a chat box because they never had to. When your product understands an enterprise customer’s own operation better than that customer could articulate it themselves, that’s not something a wrapper replicates.
But the one I’d lean on hardest is what it’s woven into. The defensibility isn’t in staying away from third-party compute; it’s in what you build around it. An agentic loop that touches your databases, your permission structure, your multi-step operational process, is painful to rip out no matter which model sits behind it. Pulling out a chatbot is a five-minute decision. Pulling out an automated workflow that a customer team relies on every day isn’t, and that’s exactly what makes a product sticky.
It’s worth noting that none of these moats is permanent either. Context windows keep growing, models keep getting better at working with thinner signal, and a data advantage that feels uncrossable today gets narrower every year. The moat isn’t a wall you build once and forget. It’s ground you have to keep re-earning as the models underneath you get better. That’s a harder story to put on a slide than “here’s our AI strategy,” but it’s the much-needed one.
Which is really the point. “We use AI to do X” isn’t a strategy; it’s a statement cloaked as strategy. The strategy question is which customer problem you can solve better than anyone else, because of data and context nobody else has.
No doubt AI is an operating imperative. Use it to make your product practice faster and sharper. Just don’t let it stand in for the strategy itself.
