When AI Starts Cutting Fat: The Brutal Truth About Vertical AI's Headless Future

I've seen too many AI startup stories end the same tragic way —

They spend two years building a beautiful product: accounting AI, legal AI, medical AI. Workflows, interfaces, domain logic, accountability — all bundled together. Users find it convenient, the company collects revenue, everything looks great.

Then foundation model capabilities leap forward, general AI agents emerge, and yesterday's moat becomes a wall overnight — not a wall keeping competitors out, but one trapping the company inside.

At this point, people start pushing a Headless transformation: hand workflows and interfaces over to agents, expose domain expertise as callable services.

Sounds fashionable, right?

But the paper discussed here says: "going headless" may be a trap, and most companies are jumping right in.

---

What Is "Going Headless"?

Vertical AI companies offer AI solutions in a specific domain (legal, medical, accounting). Their traditional model is bundling — workflows, interfaces, and domain logic packed into one application.

Two benefits:

• Simple user experience: one piece of software solves all problems in a domain
• Clear accountability: when something breaks, users know who to blame

In the era of large models, general-purpose agents can book flights, answer emails, write code. Why not do your books, review contracts, read scans?

So the thinking goes: expose "professional capability" as an API and let agents call it. That's the headless idea. Beautiful in theory. Brutal in reality.

---

A Dangerous Confusion: Interface Boundary vs. Accountability Boundary

The paper's core insight: interface boundaries and accountability boundaries are not the same thing.

• The interface boundary asks: how does the user interact — web, app, or API?
• The accountability boundary asks: who is responsible when things go wrong?

Headless advocates treat moving from app to API as a mere technical decision, but they may be moving the accountability boundary too — and they shouldn't.

Example: accounting software errors cause a botched tax filing and the client's company collapses. Who is responsible? If the software vendor "goes headless" and a general agent calls its API, who answers for the failure — the software company, the agent developer, or the user?

When the accountability boundary blurs, no one is truly responsible.

---

An Analytical Framework: Coase, Teece, and Platform Envelopment

The paper draws on three classic theories:

• Coase's theory of the firm: firms exist because market transactions carry costs; firms form when internal coordination costs less.
• Eisenmann et al.'s platform envelopment framework: platforms gain power by enveloping adjacent products — even as technical interoperability rises.
• Teece's complementary assets theory: durable value often lies in assets highly complementary to core capabilities.

Key conclusion: orchestrators operating on open protocols gain envelopment power even as interoperability increases, and durable value concentrates in "co-specialized accountability assets" — professional sign-off rights, regulated workflows, evidence chains, trusted record systems. These, not algorithms or models, are vertical AI's real moat: the commitment and track record of someone taking responsibility when things fail.

---

A Three-Type Taxonomy

The paper classifies firms not by industry but by task-accountability regime:

1. Component: provides professional capability without accountability. E.g., a tax-calculation API that doesn't guarantee audit survival. 2. Integrated Software Platform: bundles workflow, interface, and domain logic, and carries full accountability. 3. Dual-track: runs both businesses, strictly separated by brand, legal entity, and accountability structure.

The key: your category is determined not by what you do, but by how much responsibility you're willing to bear. An accounting AI that goes headless slides from category two to category one — trading accountability for faster expansion, and losing its moat.

---

Rule Debt: The Hidden Cost

The paper also introduces Rule Debt: when business rules and professional standards migrate from "governed systems" into "prompts and agent instructions," future governance, maintenance, and accountability burdens accumulate like debt on the customer organization.

When accounting software computes your taxes, the logic lives in a regulated system — audited, compliance-checked, logged. Turn it into a prompt for a general agent: will the prompt be modified? Will the agent "misunderstand"? Who reviews its correctness?

These problems don't disappear — they're transferred to customers who lack the expertise to handle them. You appear to "empower" customers; you're actually offloading burdens.

---

Four Principles

1. Decompose by accountability, not by interface. Before deciding what to keep or expose, ask: who owns this piece's accountability? If it's yours, don't go headless with it. 2. Invert the edge, retain the core. Let agents take the peripheral, non-core workflows; grip the accountable core tightly. 3. Position rule debt as a cost the integrated platform saves customers. If your platform shields clients from rule debt, it's worth the price. 4. Avoid dependence on a single orchestrator. Don't expose all capabilities to one agent — you'll become its vassal.

---

A Question for Founders

Headless sounds modern, open, future-facing. But many founders chasing it don't realize they're abandoning their most valuable asset — not technology, not data, but the promise that someone takes responsibility when things go wrong.

An industry where no one is accountable becomes an industry no one dares to use.

Why is medical AI slow? Not the technology — no one will be the defendant when AI errs. Why does accounting AI struggle? Not dumb AI — no firm will sign a tax filing and eat the penalty.

Accountability isn't a burden. It's the moat.

Before going headless, ask: when this fails, who is responsible? If you can't answer, don't do it.

---

References

1. Hydari, M. Z., & Muzaffar, F. (2026). *Going Headless? On the Boundaries of Vertical AI Firms*. arXiv:2605.17812. 2. Coase, R. H. (1937). The nature of the firm. *Economica*, 4(16), 386-405. 3. Eisenmann, T., Parker, G., & Van Alstyne, M. (2006). Strategies for two-sided markets. *Harvard Business Review*, 84(10), 92-101. 4. Teece, D. J. (2010). Business models, innovation strategy, and the catch-up problem. *Strategic Management Journal*, 31(2), 151-174. 5. Zuboff, S. (2019). *The Age of Surveillance Capitalism*. PublicAffairs.