AI agent, AI assistant, and LLM are three structurally different categories in 2026, distinguished by whether the system can reason about a goal (LLM yes), invoke tools to achieve it (assistant adds), and operate autonomously across multi-step workflows (agent adds again). Procurement that conflates the three optimizes the wrong axis: model-quality bake-offs decide the LLM tier, governance scaffolding decides the assistant tier, and operational preconditions (registry, baseline, change-management, threat model) decide whether an agent can scale at all.

Site-wide pillar piece. Captures the 'ai agent vs ai assistant' query family (currently 16+4 imp/month at GSC positions 57-85). Cadence 60-day. Trigger conditions: foundation-model providers reframing categories (e.g. OpenAI redefining 'agent' more narrowly or broadly); new benchmark families that change the multi-step-reliability ceiling cited inline (CRMArena-Pro 35%, TheAgentCompany 30%); regulatory categorisations under EU AI Act guidance that distinguish or merge these classes. Sister claims: AM-141 (What is Agent Mode), AM-029 (88/12 bimodal ROI), AM-122 (agent eval frameworks).