After Chat: The Three Transitions Between Non-Programmers and Agentic AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Version of record: SSRN 7520019, https://ssrn.com/abstract=7520019 (DOI 10.2139/ssrn.7520019). Cite that version.The same text is also at https://github.com/vadimchernets/papers/tree/main/after-chat; teaching code at https://github.com/vadimchernets/after-chat.This deposit is an archival copy made so that the full text stays retrievable; it makes no separate claim. Most professionals meet AI as a chatbot, while the same models, as AI agents, now act on files and programs. I ask why so few non-programmers have crossed from chat to agentic AI, and how to measure the crossing. The answer rests on a four-level model defined by reach and role and independent of products: chat, a vendor-bounded safe agentic environment, an agent on one's own computer, and repeatable work orchestrated across several models. A transition counts only when an agent has produced an artefact the person asked for. Installing software counts as an attempt. Evidence from 2025 to September 2026 indicates that the barrier to installation and to a first bounded attempt has fallen, while reliability on whole pieces of work, product stability and folder safety have not. Installation is one step, the person's own chatbot can guide it ("reflexive onboarding"), and two vendors merged agent and chat surfaces in 2026. One vendor's press-reported, directional figure puts knowledge workers at a fifth of its coding agent's five million weekly users; it measures use of a terminal-heritage agent, not level-3 adoption as defined here. Use remains broad and shallow: about half of US workers use AI, about one in seven daily, and few for automation. The barrier has moved from installing to directing. Drawing on end-user programming research, I name the missing capability workflow vocabulary and set out the manager's work as a delegation contract: goal, boundary, artefact and check. Web orchestration (one question to several chatbots) is proposed as the bridge that teaches comparison, and CLI orchestration (one agent calling others) as the durable layer, bounded by the fact that correlated models' agreement is weak evidence. The paper closes with fifteen falsifiable propositions, the design of a preregistered cohort programme without its results, and open teaching code.