A system that processes information and produces an output. Not magic. Not a person. A tool that needs the right level of trust for the right job. It can be wrong while sounding completely sure, which is why its freedom is widened step by step rather than granted all at once.
Autonomy is earned through evidence.
Why structure matters
A new employee is not given every key, every account and unlimited authority on their first day. They learn the context, demonstrate judgement and earn responsibility over time.
AI should develop in the same way: begin with visibility, prove reliability, then expand its freedom within boundaries people can understand and govern.
A five-stage maturity ladder
Each step changes what the system can see, decide and do. Progress depends on evidence, not enthusiasm, and different capabilities may sit at different stages.
Summary of the five-stage maturity ladder, from most human control to most system autonomy: stage 01 Observe, Human-led, Reads only what a person gives it in the moment. It produces output for review; nothing is stored or connected. stage 02 Assist, Source-aware, Reads from approved sources and summarises. A human reviews everything before any action is taken. stage 03 Propose, Approval-led, Suggests a specific email, update or change. A human must approve each action before it happens. stage 04 Delegate, Boundary-led, Acts on its own within a defined boundary, escalating anything that falls outside it. stage 05 Trust, Outcome-led, Operates independently across its domain, with human override authority and periodic oversight. Use the arrow keys to move between stages.
Reads only what a person gives it in the moment. It produces output for review; nothing is stored or connected.
ExampleAn assistant that rewrites a passage a colleague has pasted into the chat box.
To advanceA checked run of its answers is accurate, and an approved list of the sources it may read exists in writing.
Reads from approved sources and summarises. A human reviews everything before any action is taken.
ExampleThe same assistant reads an approved folder of documents and returns a summary of what changed.
To advanceChecked summaries match the source documents, and the same action follows from them each time.
Suggests a specific email, update or change. A human must approve each action before it happens.
ExampleIt drafts the reply to an incoming request and leaves it waiting; a person presses send.
To advanceA high share of drafts is approved unchanged across real volume, not a handful of tests.
Acts on its own within a defined boundary, escalating anything that falls outside it.
ExampleIt sends routine replies below an agreed threshold on its own, and escalates everything else.
To advanceA review of its escalations shows the right cases were passed on, and the logged mistakes are rare and were all reversed.
Operates independently across its domain, with human override authority and periodic oversight.
ExampleIt handles the whole category of requests end to end and adapts its approach as cases change.
Ongoing oversightPeople keep override authority and review its outcomes on a regular, scheduled basis.
Place your own use case
Describe something you would like AI to do. A model reads the description and suggests the stage it should begin at, the boundary to hold and the evidence needed before it moves up. A suggestion for discussion, not a decision.
An organisation in practice
Most organisations will have many low-risk capabilities learning at the early stages and fewer operating with broad autonomy.
Illustrative bar chart. Share of a sample portfolio of AI capabilities at each maturity stage: Observe 38 per cent, Assist 29 per cent, Propose 19 per cent, Delegate 10 per cent, Trust 4 per cent. Use the arrow keys to move between bars.
Illustrative figures, not measured data.
A live connection
MCP gives AI systems a structured way to discover and use available capabilities. Here, the connection is visible and bounded rather than hidden or assumed.
What it can do is declared. An AI client sees a finite, named list of tools rather than open-ended access.
The tools are read only. It can ask for the definition, the five stages and the illustrative figures. Nothing else, and it can change nothing.
The boundary is visible. That is stage two of the ladder, in working order.
This is a technical endpoint for AI clients, not a page to read. Opened in an ordinary browser it shows nothing useful: paste it into an AI client instead.
Read only: three tools, nothing else
The core idea
Autonomy is earned through evidence. It is never assumed on day one.
Hrihoreanu.Written by Leonard Hrihoreanu, who designs and governs AI capabilities inside a global enterprise.