The Interface Is the Control Point: What Workday's $1.1B Bet Tells You About Governing AI

According to a recent essay by Edward Hsu, Workday reportedly paid $1.1 billion to acquire an AI interface layer and, for the first time, opened its data model to outside standards. Read those two moves together and they stop looking like a product update. A company that built its entire valuation on being the place where HR and finance work happens just paid a billion dollars to defend that position, then cracked open the wall it spent two decades sealing.

The instinct is to read this as a model story. It is not. The contested asset here is not the AI underneath. It is the surface where the work lands.

The Threat Is Intermediation, Not Competition

Workday’s problem is not a better HR product. It is that an employee can now ask Microsoft Copilot or Google Agentspace to synthesize a workforce plan, and the answer arrives without anyone ever opening Workday’s interface. Hsu reports that Forrester called this the “existential threat of AI intermediation” (Forrester, via Hsu). The phrase is exact. Intermediation, not competition.

Competition is another vendor selling a rival product. Intermediation is an assistant sitting between the user and the product, doing the coordination the product used to own. The data may still live in Workday. The decision, the synthesis, the moment a human acts on information, moves to wherever the assistant lives. That moment is the asset. Lose it and you keep the database while someone else keeps the customer.

This is why opening the data model matters. An incumbent that goes “headless” stays technically relevant, but relevance is not control. Hsu makes the sharp distinction: an API layer “that nobody is structurally dependent on is still a wedge, and replaceable.” Becoming a data source for other people’s interfaces is survival. It is not power. Power is being the place the work happens.

Five Control Points, Reshuffled by AI

Hsu’s framework names five kinds of control points software can hold: Governor, Language, Aggregator, Exclusive, and Arbitrator. AI does not erase them uniformly. It reshuffles which ones hold.

Complexity-based moats dissolve first. When a product’s defensibility rested on being hard to use or hard to integrate, AI removes the friction that protected it, and the moat drains. Genuine exclusives strengthen instead: proprietary data nobody else has, workflow context that took years to encode, physical infrastructure that cannot be prompted into existence. Those get more valuable, not less, because the scarce thing becomes scarcer relative to the now-abundant thing.

The speed of the reshuffle is the part to sit with. Per Hsu, OpenAI captured three control points in two years. Anthropic’s MCP established a Language control point in roughly one. Google pushed AI Overviews to over a billion monthly users. These figures come secondhand through the essay and should be read as illustrative rather than independently confirmed, but the direction is unmistakable. Control points that took incumbents a decade to build are now being claimed in single-digit quarters.

Hsu offers one more data point that should unsettle any business with a website. He cites BrightEdge figures showing search impressions up roughly 50% year over year while paid click-through rates fell about a third over a parallel period (BrightEdge, 2025, via Hsu). More people see your content. Fewer reach you. The interface, in this case the AI answer, captured the relationship and kept the click. Same pattern as Workday, different surface.

Why This Is a Governance Decision, Not a Vendor Story

Here is the move most readers will miss. The interface is not just where commercial value concentrates. It is where governance attaches.

Audit trails, policy enforcement, access controls, the record of who did what and under whose authority: all of it lives at the layer where work is performed. When an employee builds a workforce plan inside Workday, Workday’s controls govern that action. When the same employee builds the same plan by asking Copilot, the governance moves to Copilot. Your audit trail relocates to a surface you do not own and may not even be able to see.

This is the falsifiable claim worth testing against your own stack: whoever owns the interface where AI work happens owns the control, audit, and policy surface over that work. If that holds, then interface ownership is the primary governance decision an enterprise makes, and model choice is downstream of it. You can run the safest, most compliant model in the world, and if the work happens in an assistant whose logs you cannot inspect, your governance is theater.

Prior thinking in our corpus established that the workflow, not the app, is the defensible asset as AI commoditizes features, and that permissions and the system of record anchor agent governance. This is the next layer down. The system of record holds the data; the interface holds the action. Govern the data and lose the interface, and you have governed the noun while someone else governs the verb.

The Question Procurement Has Not Asked Yet

Most enterprise AI evaluations ask which model is most capable, most compliant, cheapest per token. Reasonable questions. They miss the one that determines whether you can govern anything at all: where does the work actually happen, and who logs it?

When you let an assistant intermediate a system of record, you have made a governance decision whether you meant to or not. You have moved the audit trail. You have changed who can enforce policy. You may have done it through a checkbox in a Copilot rollout that no risk committee ever reviewed. The framework you adopt is itself a governance bet, and so is the interface you let your people work inside.

Do This Now

Run an interface audit, not a model audit. For each system of record that matters (HR, finance, CRM, the data behind your decisions), answer three questions:

First, where does work on this system actually happen today? In the native interface, or increasingly through an assistant that reads and writes on the user’s behalf?

Second, where does the audit trail for that work live? If a regulator asked you to reconstruct who changed what and why, would the record be in a system you control, or in a third-party assistant’s logs you cannot fully access?

Third, if your incumbent vendor goes “headless” and your people start working through Copilot or Agentspace, have you decided that on purpose, with controls in place, or is it happening by drift?

The vendors are spending billions to answer the first question in their favor. Workday’s reported $1.1 billion is the price of one company refusing to let its interface be intermediated. You do not need a billion dollars to make the same decision deliberately for your own stack. You need to make it before someone makes it for you.

---

This analysis synthesizes How AI Changes the Power Dynamics of Software (Edward Hsu (Substack), 2025).

Victorino Group helps enterprises decide which interfaces govern their AI work and keep the audit trail on surfaces they control. Let’s talk.