A Capability-by-Capability Test for Going Headless

The decision facing a vertical AI firm is no longer whether its entire product is a platform or a component. It is which capabilities can cross into an agent without taking an ungoverned commitment with them. This comparison turns the accountability model into a practical test using verification cost, error stakes, accumulated context, authority, and transferability of responsibility. It also shows when dual-track operation, staged migration, or no action is more robust than a wholesale architectural shift.

A sector label cannot answer this question. A legal technology firm may expose research and document comparison while retaining a governed filing process. A healthcare company may make summarization callable while keeping diagnosis and care planning inside clinical governance. A software company may modularize code generation while retaining release approval and incident responsibility. The unit of analysis is the capability and the output it helps create.

flowchart TD X["Capability under review"] --> V{"Low-cost verification and limited error stakes?"} V -->|Yes| R{"Can responsibility be transferred?"} R -->|Yes| C["Component path: expose callable capability"] R -->|No| D["Dual-track: expose task, retain commitment"] V -->|No| A{"Does context, evidence, or authority accumulate?"} A -->|Yes| P["Integrated-platform path: retain governed core"] A -->|No| S{"Have migration conditions changed?"} S -->|Yes| D S -->|No| N["Staged review or no-action"] C --> M["Monitor substitution, rule debt, and control drift"] D --> M P --> M M --> S

First classify the output, not the sector

The first gate is verification. Can the result be checked independently, at a cost proportionate to the task, before it creates a consequential commitment? “Cheap” does not mean effortless or unimportant. It means that the customer or another system can determine whether the output is acceptable without reconstructing a large body of hidden context, prior judgment, and authority.

The second gate is responsibility. If the result is wrong, can the organization transfer responsibility cleanly to the party using or accepting it? A modular output may be passed from an agent to a customer workflow because its meaning remains stable and its errors can be identified. If the result changes a person’s status, authorizes an exception, commits a regulated action, or must be defended by a named professional, responsibility is less transferable even when the underlying task looks small.

A component position is appropriate when both gates are favorable. The output is modular, the stakes are limited enough for independent checking, and the customer does not need the vertical firm to preserve the entire surrounding context. Semi-structured data extraction, document classification, standards retrieval, and first-pass summarization can fit this pattern when their results remain separately verifiable.

For a component business, headlessness is distribution. The customer may not want another application; they may want the capability inside an agent, enterprise workflow, or internal system. The vertical firm should make the component reliable to invoke: accuracy, latency, price, schemas, permissions, and broad integration matter. Its product must work as a bounded service across more than one surrounding workflow, because its value is now judged in the places where others assemble it.

That path has a clear failure condition. If a general-purpose model can approximate the output with ordinary prompting, technical availability becomes a commodity rather than a moat.