The five rungs of AI use

Five rungs of tooling, and almost everyone is on the first without knowing there are five. Chat: you type, it answers, you close the tab and nothing carries over. Project: a container that remembers. Skill: a saved instruction set you call by name. Automation: it happens without you asking. Agent: you hand over a whole messy task and come back later. Each step is free and each takes about ten minutes to try.

The five rungs#

1. Chat. You type, it answers, you close the tab. Nothing carries over. This is where everybody starts and where most people stay.

2. Project. A container that remembers. One per subject, with your notes and your rules in it. You stop introducing yourself every time.

3. Skill. A saved instruction set you call by name. "Run my essay check." You built it once and you improve it when it disappoints you.

4. Automation. It happens without you asking. "Every Friday at five, turn this week's notes into ten questions, and do not give me the answers until I try."

5. Agent. You hand over a whole messy task and come back later. It uses tools, reads files, and does things rather than describing them.

What changes between rung three and rung four#

The rungs are not evenly spaced. One to three are conveniences: the same work with less repetition, and the person is present throughout. Four and five remove the person from the moment of execution, and that is a different kind of step.

Anthropic's own engineering guidance draws the line in the same place, distinguishing workflows, where models and tools are orchestrated through predefined paths, from agents, where the model directs its own process and tool use. Their recommendation is to find the simplest solution that works and to add agentic complexity only when it demonstrably improves outcomes, which is a caution against climbing for its own sake from the people selling the ladder.

The closest cultural analogue is SAE J3016, the six levels of driving automation, instructive for the reason people misuse it: levels three and four are where the human is nominally responsible and not actually attending. That is the same problem, one rung earlier.

The rung where oversight quietly stops working#

Rung four is where this research would put a warning. An automation that runs on a schedule produces output nobody asked for at the moment it appears, and unrequested output is the hardest kind to review attentively.

That is the omission half of automation bias: errors of commission happen when someone acts on a wrong recommendation, and errors of omission happen when nobody notices what the system did not flag. Almost every oversight process ever written catches the first. Parasuraman and Manzey found that the more reliable an automated aid is, the less attention its supervisor pays, which means rung four gets safer and less supervised at the same time.

The practical rule that follows: climbing rungs one to three needs no governance. Rung four needs somebody to answer what happens when this runs and is wrong and nobody looks. Rung five needs an answer to who can override it before it is switched on, not after.

Higher is not the goal, and the deck says so#

The line this framework is taught with is that the future is not typing prompts into a box faster, and the advice attached to it is deliberately modest: get to rung two this week, try rung three in the holidays, you will not need four or five at university but you will at work.

Most of the value is in the step from one to two, which is free, takes ten minutes and simply stops you re-explaining yourself. Anyone at rung five on a task that needed rung two has built a machine to answer a question they could have asked.

What this has not been shown to do#

Nothing has tested it. There is no evidence that people who climb these rungs produce better work, learn more, or supervise better, and no measurement of where the marginal return actually sits. The rung labels also track product features, which means they will date: three of the five did not exist as named things three years ago, and a ladder pinned to a vendor's menu is a ladder that will need rewriting.

The oversight caution attached to rung four is inherited from the automation literature rather than measured on these rungs, and that literature is about aviation and process control, not about a scheduled summary of somebody's notes.

Key sources

• Parasuraman, R. and Manzey, D. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
• Parasuraman, R. and Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors, 39(2).
• Anthropic. Building effective agents. Read at source 6 September 2026.

Related SuperSkills research#

On the request ladder rather than the tooling ladder, the four levels. On rungs four and five, letting an agent act on your behalf, who manages AI agents and designing a stop button people will use. On what stops working as you climb, automation bias and the invisible work of oversight.

About this framework#

Chat, Project, Skill, Automation, Agent is used by Rahim Hirji in teaching and in the Mastering AI deck, most recently in September 2026. No claim of first use is made. No dated first publication exists for it and the Box of Amazing archive carries none, so it anchors to SuperSkills (Kogan Page, 2026) and to the deck. The workflow and agent distinction is Anthropic's, the levels of driving automation are SAE International's, and the automation complacency findings belong to Parasuraman and Manzey. The reading offered here, that rung four is where oversight quietly stops working, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.