Key Takeaways

  • More than 80% of the world's data remains locked behind corporate firewalls, out of reach for public foundation models such as Anthropic and OpenAI.
  • Approximately 95% of enterprise AI pilots collapse because generic systems cannot access proprietary operational records without trusted software workflows.
  • Workflow applications act as extraction tools, securing internal data sets that public web scrapers will never touch.
  • Companies that secure internal data establish a procyclical feedback loop, mirroring the data flywheel Tesla built for autonomous driving.

Public Models Stop at the Enterprise Firewall

Public foundation models have ingested almost everything on the open internet. The issue is that public internet text does not run a business. Joseph Coyne, co-founder and managing partner of Studio VC, points to a clear data imbalance: “One of our favorite stats, and it actually came through an enterprise that we work very closely with, which is that about 80% or more of the world's data is behind firewalls.” (play from 6:45)

When large companies buy AI tools, they test them on public benchmarks and watch them fail. Coyne notes that 95% of enterprise AI pilots currently fall flat. The failure does not stem from weak algorithms. It happens because foundation models trained on Reddit, Wikipedia, and open web crawls lack the context buried in internal tickets, ERP ledgers, and secure customer histories. Corporate legal teams will not permit foundation model providers to scrape those repositories to train open weights. The data stays sealed inside the perimeter.

Software Workflows Extract Trapped Data

Enterprises will not open their databases to a raw API prompt, but they will yield data to software that solves an immediate operational problem. As David Weisburd observed, “the reason 80% of the data is behind the enterprise firewalls is because it comes from the customer and in order to get that customer data you must provide a good product.” (play from 8:06)

Coyne frames the winners in this market not as algorithm shops, but as specialized infrastructure: “The best of those companies come with an enterprise software workflow that systematically helps companies to unlock their data and then run AI best practices on it in a highly secure way.” (play from 7:53) He adds: “Enterprise software that is creating workflows that unlock new data feels like the oil rig company pulling new oil out of the ground for enterprise.” (play from 8:37)

If an application embeds itself in how workers review contracts, invoice clients, or coordinate logistics, it captures proprietary exhaust as a byproduct of normal operations. That operational exhaust is the only fuel that makes models accurate inside the company.

The Tesla Flywheel Applies to B2B Data

Once software captures that proprietary operational stream, it creates defensive separation. Coyne calls this dynamic a “procyclical advantage, meaning once you're Tesla and you get a lot of data then your model becomes better and you get more users and then you get even more data.” (play from 21:37)

A generic model cannot catch up through compute alone. Compute can re-train weights, but it cannot synthesize private enterprise actions. The software provider that secures the workflow earns the right to fine-tune on private transactions. Each interaction tunes the system to the customer's specific operational rules, widening the gap against horizontal competitors attempting to sell top-down seats.

Why It Matters

This dynamic shifts enterprise software underwriting away from raw model capability and back toward workflow stickiness. Buyers are rejecting horizontal AI overlays, leaving foundation models vulnerable to commoditization unless paired with secure distribution. Investors who underwrite vertical software with defensible data capture are backing the actual gatekeepers of enterprise value.