Why inbound data marketplaces have coverage gaps

A data marketplace can only offer what sellers bring to it. That sounds obvious, but it has a consequence AI labs feel quickly: the supply on any inbound platform reflects who decided to list, not what a training or evaluation program actually needs. The gaps that result are not random, and they rarely close on their own.

On an inbound marketplace, a company appears only after someone inside it decides their data is an asset, finds the platform and does the work to list. Each of those steps filters the pool.

The companies that make it through tend to share traits. They are often close to the AI world already, run by people who follow training data closely, or advised by someone who has seen a licence before. They usually have a person with time to manage a listing. And they are rarely under time pressure.

None of that makes their data bad. It means the catalog is a sample of companies that were already thinking about data, which is a small and unusual slice of all the companies sitting on useful operational records.

When a lab writes down what it needs and compares that with what is available inbound, the misses tend to cluster in the same places:

• Sector skew. Software and tech-adjacent companies tend to be overrepresented, because their founders are the most likely to know data licensing exists. Logistics, construction services, healthcare administration, manufacturing and professional services tend to be thinner, even though agents are being built for exactly that kind of work.
• Size skew. Very small companies can list easily but carry limited history. Mid-sized companies, with enough headcount to show real handoffs, often have nobody whose job includes thinking about this.
• Tool-stack gaps. A lab may need companies that ran on a specific combination, such as Jira, Slack and a particular CRM, all at once. Listings rarely let you search for that, and listed datasets are often single-system exports.
• Timing gaps. Companies that are winding down have the most reason to license and the least time. Subscriptions get cancelled and workspaces deleted within weeks, and many close before anyone involved hears that their data had value.
• Regional gaps. Listings come mostly from markets where the idea has already spread. Companies elsewhere stay invisible.

Each of these matters on its own. Together they shape the training mix in ways nobody chose.

It is tempting to assume these gaps will fill as awareness grows. Some will, slowly. But the structural reason they exist does not change: the people who control the most useful data are not looking for a buyer.

Think about who can actually sign a licence. At an operating company, it is a founder, CEO or CFO with a full calendar. At a company that is closing, it is a director, officer, liquidator or assignee focused on creditors, payroll and paperwork. None of them are browsing data marketplaces, and many have never considered that their old Jira projects and Slack channels are worth anything.

A listing cannot reach someone who does not know to look. For winding-down companies in particular, the window may close long before awareness catches up. The archive is gone, and the catalog never knew it existed.

The alternative is to start from the specification rather than the supply. That is how we work with labs. A lab sends a sourcing brief describing what it needs: sector, headcount, tool stack, years of history, regions, and whether the companies should be operating or winding down.

From that brief, we build a target list using company and technographic data, then our reps call the people who can sign. Martal Group has spent 17 years running B2B outbound, with more than 200 onshore sales reps across the US, Canada, the EU and LATAM and experience prospecting in 24 industries. Cold calling, email and LinkedIn are what reach the CFO of a regional freight broker who would never have listed anything.

Companies that agree go through a read-only evaluation, in-house anonymization and a review of an anonymized sample, and data is delivered to the lab on a rolling basis. The pool is shaped by the brief, not by who happened to show up.

Inbound marketplaces are useful for what they hold. If a listed dataset fits your needs, there is no reason to ignore it. The mistake is treating the inbound catalog as a picture of the whole market and letting it quietly define what your agents are trained and evaluated on.

A practical approach is to write the brief first, check what is available inbound, and then source the gaps directly. That keeps the training mix tied to the work you want agents to do, rather than to the sectors that happened to hear about data licensing first. It also makes the gaps visible, which is half the battle. A lab that never writes down what it needs cannot tell what it is missing.

The same applies to evaluation. A private evaluation set drawn only from the companies that chose to list will test agents on a narrow slice of real work, and the scores will look more general than they are.

If you have a specification that inbound supply is not meeting, send us a sourcing brief.