AI System Capability Taxonomy

AI System Capability Taxonomy

Purpose

This taxonomy is an operational description, not a universal standard. It exists to stop a familiar chat response from hiding meaningful differences in authority and risk.

Saved prompt

A saved prompt stores reusable text. The user chooses when to paste or invoke it. It does not add knowledge, memory, tools, or authority by itself.

Custom assistant

A custom assistant combines persistent instructions with a model. It may also include selected knowledge, examples, conversation starters, or a visual identity. It still responds inside a user-initiated interaction unless other capabilities are added.

Source-grounded assistant

A source-grounded assistant retrieves from an approved corpus and should expose an evidence path. The important distinction is not the product name. It is whether answers are constrained to selected material, how retrieval works, and whether the reader can inspect support.

Workflow

A workflow connects defined steps. Some steps may be deterministic and some may call a model. The sequence, transition rules, error handling, and approval points are designed in advance.

Tool-using workflow

A tool-using workflow can query or change another system. Permissions, credentials, write scope, idempotency, logging, and approval become part of the system boundary.

Agentic loop

An agentic loop can select or revise steps toward a goal, observe results, and continue within a defined budget and authority boundary. It may still require human approval for consequential actions. Autonomy is a degree, not a magic category.

Public persona

A public persona adds identity and audience. It may be only a custom assistant technically, yet its disclosure, consent, impersonation, moderation, bystander, and support risks are larger because strangers can interact with it.

Minimum description

Any system description should name the model, instructions, knowledge, memory, tools, initiation, permissions, execution loop, audience, human approvals, logs, failure handling, owner, and shutdown path. If one of those fields is unknown, the uncertainty belongs in the description.