Enterprise AI Data Gap

The asymmetry

Asked why consumers and developers had already had a magical AI moment while enterprises had not, Brendan Foody locates the blocker in data availability rather than model capability.¹ Search worked because models could learn the entire open internet, an already available and instantly accessible context. Coding worked because of GitHub as a foundation plus clear verifiers and unit tests as a success signal, which together supplied abundant material to train against. Enterprise verticals such as investment banking, law, and consulting have no equivalent giant online corpus that answers what a model should actually do in a given case, and no native way to grade an attempt. Closing that gap, in his framing, requires an enormous amount of human work to create the right context and build the evaluation environments that let a model be trained against each vertical, which is also his own company Mercor's business.

Why simply feeding documents into a model fails

The popular idea of scraping a company's files, chat logs, and customer records and dropping them into a long context window fails for two separate reasons. Models are not used to reading that kind of data and understanding how to use it, since it sits outside the distribution they were trained on, so connecting an inbox and a few files and asking for something as simple as a drafted email can cause a model to break down entirely. And a great deal of the most valuable knowledge inside an enterprise was never written down at all; it lives in people's heads, so no amount of document scraping retrieves it, and extracting it requires humans in the loop.

How the gap gets closed

Foody's own company treats this as a future-of-work claim as much as a business model. Roughly eighty percent of Mercor's customer-support staff now spend their time building evaluations and updating context for the support agent rather than answering tickets directly, following the same fixed-cost logic that let software eat routine work in the first place: a person does a redundant task once, trains an agent to do it, and moves on while the agent handles the recurring version. Foody's own summary of the irony is that the future of AI is, in practice, very human, since people remain the bottleneck on getting the right context into a system in the first place.

The same gap read as a moat

A separate account of the same underlying fact reframes it from the opposite commercial position. Robert F. Smith, describing the software companies his firm owns, cites a claim that less than one percent of enterprise data has ever been part of the environment used to train large foundation models.² Where Foody, whose company sells the fix, treats this as a gap to be closed, Smith, who owns companies sitting on the far side of it, treats the same fact as a moat: if a general model cannot see inside a given enterprise, then whoever already holds the workflow can deliver agentic solutions to that industry that no one else can, and capture the resulting value. Both readings are correct at once and the difference is purely positional, a cost to whoever must close the gap and a source of rent to whoever already owns the far side of it.