Separate three use-case classes
Assistive: humans decide; models accelerate drafting, retrieval, and summarization. Semi-automated: models propose actions; humans approve by threshold. Automated: execute inside strict bounds (classification, routing, fixed-template replies). Mixing classes misaligns permissions, audit, and SLAs.
Rule of thumb: higher business risk and lower reversibility stay assistive; automate only when process boundaries are clear and error cost is tolerable.
Risk control is a product requirement, not a post-launch patch
Enterprise GenAI must specify at design time: hallucination and fact-check mechanisms, sensitive-data residency, audit logs for prompts and tool calls, and human takeover paths. Treating these as late “compliance attachments” often costs more than redesigning the use case.
Authority of knowledge sources matters: are internal policies and product facts versioned and citable? Generative answers without recoverable evidence lose trust quickly in marketing and service.
Delivery shapes: factory, embed, co-create
Three shapes can coexist: capability factories (shared prompts, evals, safety, platforms); business embeds (GenAI inside CRM, tickets, supply-chain systems); joint co-creation (end-to-end for high-value processes, including change management).
MICROFACTOR tends to open critical processes with co-creation, accumulate reusable capability in factories, and scale via embeds — not replace operating design with a one-shot model purchase.
Joints with data and process
GenAI quality is often capped by retrieval corpora and process state machines. Without structured process context, chat UIs manufacture more exception tickets. Schedule GenAI with process intelligence and master-data work — not “model first, process later.”