Does proprietary data still create a moat when AI is a commodity?

Is data still a moat when everyone can use the same models?#

Proprietary data is still a moat when AI models are a commodity, but only a narrow kind of data qualifies. When every competitor can rent a capable model, advantage moves to what the model is given and where the product sits in the customer's daily work.

A data moat is not the archive itself. It is the combination of records competitors cannot rebuild, a loop that keeps producing new records through your product or operations, and decisions customers pay you to get right. Remove any one and the archive becomes an asset with value, but not a barrier.

That distinction matters for licensing. A static copy of older records can be valuable to an AI developer without being the thing that keeps competitors out, and many software companies find that most of their archive falls into that category.

Five tests for a real data moat#

Five tests separate a real data moat from a large archive: reproducibility, renewal, decision linkage, control and workflow position. A record family that fails two or more is unlikely to protect the business, however much of it there is.

Run the tests record family by record family. A vertical software company might find that live usage data inside its product passes all five, while its engineering history and old support threads pass only one or two.

What erodes a data moat faster than licensing#

Several forces erode a data moat faster than any license, and owners who guard every record often overlook them. Each one is a reason to know exactly which records protect the business and which do not.

The people force is already visible. As TechCrunch reported in October 2025, the CEO of the data vendor Mercor described AI labs drawing on former senior employees of banks, consulting firms and law firms because the firms themselves would not hand over data that could automate their work; Mercor says it tells contractors not to upload former employers' documents. Refusing to license does not keep expertise inside the company.

The practical response is to map each record family to the product feature or operating decision it feeds. A record family that feeds nothing customers see already offers little protection, and its value as licensed training or evaluation material may be the larger of the two.

• 1Staleness: workflows, regulations and products change, and a moat built on old records shrinks without renewal.
• 2General models catching up: tasks that once needed proprietary examples become routine for models trained on broad data.
• 3Customer control: customers can export their own content and take it to a competitor, and contracts often let them.
• 4Unused records: data that never reaches a product feature protects nothing.
• 5People leaving: much of a company's edge sits in the judgment of experienced staff, who take it with them to rivals or to AI training projects.

When does licensing weaken the moat?#

Licensing weakens a data moat when it hands a developer the records that power your differentiating feature, on terms that let the developer serve your customers. Three patterns cause most of the damage.

The first is licensing a live, continuously refreshed feed of the records your product depends on. The second is granting rights with no field-of-use limit, so a vertical AI startup in your market can build on your records. The third is licensing recent records that capture current pricing, playbooks or customer specifics.

By contrast, a non-exclusive, time-limited license of older engineering or support records for general model training blends your contribution with many other sources. The developer gains examples of how work gets done; it does not gain your customers, your feedback loop or your place in their workflow.

Which license terms protect the moat?#

License terms protect the moat by limiting which records leave, how old they are and what the developer may build with them. The table lists the terms that do most of the work.

Field-of-use limits and restrictions on competitive use need careful drafting, because overly broad restrictions can raise competition-law questions. Have counsel write them for your situation rather than copying another company's clause.

Illustrative: a freight audit software company finds its real moat#

Illustrative: a fictional freight audit software company assumes its invoice dispute history is its moat. Applying the five tests, the CEO finds that newer competitors can collect similar disputes from their own shipper customers, so the history fails on reproducibility. The real moat is a set of carrier-specific resolution rules that analysts update every day inside the product.

The company keeps the live rules and all current customer data out of scope. It licenses older dispute resolution histories and its Jira and GitHub engineering records, with personal details and shipper names removed, non-exclusively and with a field-of-use limit that excludes freight audit and payment products. The board receives a one-page memo listing what stays out, the field-of-use wording and the deletion terms.

Outcome: the company can earn licensing income from records that were not protecting anything, and the feature that wins deals stays untouched.

How SourceX approaches moat questions#

SourceX treats moat questions as part of scoping, before any record is prepared. In the SourceX Enterprise Data Value Framework, uniqueness raises value, reproducibility reduces it and exclusivity raises price, so the tests that define a moat also shape what a package is worth.

Within the SourceX five-step transaction, the Supply step is where moat records are kept out of scope, the Rights step confirms what may be licensed and on what limits, and the supplier signs off field of use, cutoff dates and naming terms at Approval. The SourceX Evidence Packet records permitted use, so the limits travel with the records to the buyer.