Earlier this year, Wall Street coined a new word: the SaaSpocalypse.
AI agents and coding tools had gotten good enough that investors started asking an awkward question. What if companies just build what they used to buy? A major software ETF had its worst stretch since 2008.
You can debate whether the panic was overdone. The trend behind it is harder to argue with. In McKinsey’s latest State of AI survey, 32% of organizations said they had already skipped buying at least one software product or feature because they could build it themselves with AI.
So when anyone can build the software, what’s left that’s actually yours?
Your data. Just not in the way most people mean it.
A data moat is proprietary data that competitors can’t copy and general-purpose AI models don’t already know, and that gets more valuable every time you use it. Owning a lot of data isn’t the same thing.
Why more data alone isn’t a data moat
In 2023, Bloomberg trained its own AI model, BloombergGPT, on forty years of financial archives. Within weeks, researchers found that general-purpose models beat it on most of the financial benchmarks they tested, without ever seeing Bloomberg’s data.
Bloomberg’s data is still valuable. It’s why people pay for a Terminal. But pouring a pile of data into a model didn’t create a lasting lead. The general models simply caught up.
Most companies have a quieter version of the same problem: data that exists but isn’t usable. Gartner found 63% of organizations lack, or aren’t sure they have, the right AI-ready data practices.
Now look at Stripe. It trained a model on tens of billions of transactions from its own network and pointed it at fraudsters testing stolen cards. Detection on its largest customers jumped from 59% to 97%.
Same strategy on paper. Very different result. Why?
What a real data moat looks like
Stripe’s data has three things a pile of archives doesn’t:
- A view no one else has. Millions of businesses on one network reveal patterns no single merchant can see.
- Outcomes, not just activity. It’s tied to what actually happened: fraud or not.
- A data flywheel. Every transaction makes the next decision a little sharper.
You’ll find the same traits outside tech. John Deere’s See & Spray cameras spot weeds and spray only them. Across more than five million acres in 2025, farmers cut non-residual herbicide use by nearly half. Deere now trusts its data enough to charge farmers only when the system delivers measurable savings.
That’s the real prize. The right data doesn’t just improve your product. It can change how you get paid.
Frost & Sullivan Best Practices 2026
Global Company of the Year for Digital CX
11 pages of independent analysis. Free PDF, one short form.
Download the report →Three moves to start building your data moat
Run the copy test. For each major data set, ask two questions. Could a competitor get this? Does a general AI model already know it? Whatever survives both is your moat. It’s usually a shorter list than you’d expect.
Record what happened next. Most systems log activity: a claim filed, a ticket opened, a quote sent. Few log the outcome: was it fraud, did the fix hold, did the customer renew? Closing that loop is what turns a log into something AI can learn from.
Build where your data is unique. Buy the rest. Cheap software doesn’t mean rebuilding payroll. Put your build budget where proprietary data drives the decision, and make sure you can get your data back out of everything you buy.
Here’s what that can look like in practice. I would like to share an example from an existing client engagement with a mid-sized industrial equipment distributor. For years, its service team has logged every repair: the machine, the symptom, the part replaced. Useful, but every competitor keeps the same kind of log. Then the team adds one field: did the fix hold for 90 days? A year later, the company knows something no rival and no general AI model can tell you: which repairs actually work, on which machines, in which conditions. That’s a data set that passes the copy test, and it started with one extra column.
The new build vs. buy question in the AI era
The old question was: what should we build, and what should we buy?
The new one is: what do we know that nobody else does, and are we learning more of it every day?
Which data set in your organization would pass the copy test? Tell me in the comments.
Not sure where AI fits in your business?
Compunnel’s Applied AI practice helps mid-market teams pick the one workflow worth automating first – and gets it running inside the tools you already use, with governance built in.
Explore Applied AI →Frequently Asked Questions About Data Moats
What is a data moat?
A data moat is proprietary data that gives a business a lasting edge because competitors can’t copy it and general-purpose AI models don’t already know it. The strongest data moats are tied to real outcomes and become more valuable with every transaction, decision or customer interaction.
Is more data enough to create a competitive advantage with AI?
No. Bloomberg trained BloombergGPT on forty years of financial archives, yet general-purpose models beat it on most of the financial benchmarks researchers tested. Volume alone rarely creates a lasting lead. What matters is a view no one else has, data tied to outcomes, and a flywheel that keeps improving it.
How do you build a data moat?
Start with three moves. Run the copy test on each major data set: could a competitor get it, and does a general AI model already know it? Record outcomes, not just activity. Then put your build budget where proprietary data drives the decision, and buy the rest.
Should companies build or buy software now that AI makes building cheaper?
Build where your data is unique and buy everything else. Cheap software doesn’t mean rebuilding commodity systems like payroll. Focus custom builds on decisions driven by proprietary data, and make sure you can get your data back out of every tool you buy.
