Sovereignty and Dominion Over Workflows

The test

"Those companies that have sovereignty and dominion over the workflows and data sets have an opportunity to become part one or part two," is how private-equity investor Robert F. Smith sorts enterprise software companies into the ones AI enriches and the ones it eats.¹ The phrase names two things rather than one, and the second is easy to underweight.

Two axes, multiplied together

Dominion over data is the familiar half: irreproducible data an outside model cannot obtain. Smith supplies a mechanism for why this holds at large-enterprise scale rather than only in boutique cases: "less than 1% of enterprise data actually is in the environment that has been used to train these large foundational models."¹ If that figure is roughly right, frontier models are extraordinarily capable in general and nearly blind inside any specific enterprise, and the gap does not close through further training, it closes only when someone grants access, which makes the vendor already sitting on the workflow the gatekeeper.

Dominion over workflow is the underweighted half: the right to act, not merely to know. A model that reasons perfectly about a scheduling decision still cannot write to the relevant calendar on its own; a model that understands a substitution decision perfectly still cannot place the order. The software that owns the workflow holds the write access, the system state, the audit trail, and the permission to actually execute a change, and intelligence is now abundant in a way that authority inside a system of record is not.

The two axes multiply rather than add. Data without workflow produces an analytics product an incumbent can copy once it decides to; workflow without proprietary data produces a thin execution layer any competitor with the same integrations can replicate. Together they produce the ability to deliver agentic solutions to an industry that no other company can.

Why it matters

The test supplies an operational instruction that "have a data moat" on its own does not: whether a management team controls both the workflow and the data, and whether it will still control both next year, implying specific defensive moves such as retaining write access and resisting becoming a passive data source for someone else's assistant. It also reframes a widely discussed problem, the gap between what enterprise data exists and what ever reaches a foundation model's training corpus, as an asset rather than only a cost. Brendan Foody reads the same underlying fact as a bottleneck that requires humans to manually supply missing context; Smith reads it as the moat itself. Both readings can be correct at once, since the same data gap functions as a cost to whoever has to close it and as a rent to whoever already owns the far side of it.

Open questions

The underlying statistic, that under 1 percent of enterprise data made it into training corpora, is vaguely sourced yet load-bearing for the whole argument. Even if true today, it describes a stock rather than a rate: enterprise data increasingly enters model context continuously through connectors, copilots, and agent integrations, so dominion may be eroding in practice even while the training-corpus figure stays flat. Dominion is also granted rather than owned outright, since the enterprise customer still owns its own data and can in principle redirect it elsewhere; a customer that decides its own data should flow to its own agent platform can make that call. And the argument sits above the hyperscalers that provide the underlying AI infrastructure, which raises the possibility that a workflow-owning vendor ends up supplying the backend for someone else's interface, holding dominion over the plumbing while the actual customer relationship migrates upward.