About the framework
What you do, where you work, and what you can run it on
Three things decide most of what AI is actually worth to a person: the work they hand over, the permission they have where they work, and the machine it has to run on. Between them they shape an AI skillset, the next step worth taking, and how someone sees the whole subject.
i
The three are never added up, never ranked and never shown against anyone else. They are a filter, not a grade: the only thing a placement changes is which advice reaches you first.Nothing here describes a tool. A capability — “can get a reliable answer out of a model and knows when not to trust it” — will still be true in three years. A tool name is stale by spring, so tool names live in the content, which is cheap to change, and never in the framework, which is not.
One · Activity, or what you do with AI
One question only: how much of the work do you hand over? Choose your own rung — no test, no score, no ranking against anyone else.
Three of these share one mountain — the trail, the ridge above it, and the launch you run off the top. The fourth is a different aircraft, not a higher ridge. Choose where you actually are: climbing is not automatically the right move, and on the wrong ground it is the wrong one.
-
Asking The trail
You ask. AI answers.
An assistant that has read everything and remembers nothing. You bring the question, the context and the judgement; it brings speed. Where almost all of us start, and a great deal of genuinely valuable work never needs to leave it.
Next rung Treating the conversation as a working session rather than a lookup — hand over the document, the constraint and the audience before you ask.
-
Supervising Ridges
AI builds for you.
You hand over the making, not only the asking — the draft, the model, the deck, the code — and you read every step before it goes anywhere. Still a minority, and the point at which AI stops being a better search box and becomes capacity.
Next rung Letting it work on your own material — your files, your data, your tools — rather than only on what fits inside a chat window.
-
Orchestrating Thermals
Teams of AI work for you.
You set the objective and let it run — several agents, many steps, real tools and real data, without reading each move. The leverage is genuine and so is the exposure. The skill at this altitude is reading the conditions, and knowing when not to launch.
Next rung Evidence. Being able to show the thing worked, and show when it did not, before anyone else is asked to trust it.
Above, and a different aircraft entirely
-
Building AI Above the weather
You work on what is not settled yet.
Research, ethics, evaluation, agents defending against agents — the machinery itself and the arguments about it. Not further up the same mountain but a different aircraft entirely: a few thousand people rather than a few hundred million. Most of us will never fly one and lose nothing by it.
What it gives the rest of us Early warning. What is contested up here today is a risk-committee question in eighteen months. The society's return on its Jets is hearing it first, and hearing it in terms a board can act on.
Two · Environment, or what you are allowed to do
Free to ship
Technical startup
Nobody to ask. You can ship to production this afternoon, build rather than buy, and change your mind next week.
Binding constraint Judgement, not permission. The risk is moving fast in the wrong direction and only noticing at scale.
By consensus
Boutique or executive office
A handful of people, no IT function, no one to delegate to. Almost everything is bought rather than built.
Binding constraint Time and vendor choice. The risk is paying for tools nobody embeds into actual work.
Controlled
Corporate with IT and risk oversight
Clearance before movement. Procurement, data classification, model approval, audit trail.
Binding constraint Permission and evidence. The risk is a pilot that works beautifully and never clears review.
Three · Equipment, or what you can run it on
Whose machine the work happens on, and whether that machine will let you install anything. Never which brand, never which vendor, never your employer’s policy.
Own Kit
A machine you control. You can install what you like.
Own and Issued
Your own machine and a managed one. Two sets of rules, and you know which is which.
Issued Only
Managed hardware and an approved list. Nothing outside it.
Phone Only
No desk machine of your own. It happens in your hand.
This is the layer that decides whether a piece of advice is actionable at all. A member on a managed device with an approved list cannot act on an article that begins “install this” — but a model running inside the estate may be the one route that is open to them, which is why it is worth knowing.
Why both
Take one capability: getting a model to work over your own documents. A Mountaineer in Open Country does it in an afternoon. The same Mountaineer on Marked Trails should almost certainly buy it. In Controlled Airspace the technical work is the easy part — what matters is which data classifications permit it and what evidence the risk function will ask for. Same altitude, same capability, three completely different articles.
Your own placement
Three questions, four taps
No test and no score. Everything published here is tagged against these three layers, so a placement is simply the filter that stops you reading advice written for somebody else's constraints — and stops you being shown worked examples that assume equipment you are not allowed to install. Every option is a real answer, and you can change any of the three whenever you like.
Set up my framework → Members sign in · everyone else applies