Seven stones. Each one answers a question the one before it raises.
Read any stone alone. The value of holding all seven is the reduction of lost intent between them — the economic question stays connected to the product decision, the promise to what the system can deliver, governance to architecture. The arch assembles as you read.
What is AI changing?
The starting point is not a vendor, a model or a use case. It is the changing relationship between people and intelligence: does AI help someone understand more and decide better, or produce polished output while they disengage from the problem?
This layer decides what is worth improving — not merely what can be automated.
- →Who receives the gains — better service, learning and time, or simply more pressure?
- →Who carries the consequences when an AI-supported decision goes wrong?
- →Which roles keep their judgement, and which quietly lose the practice that built it?
What becomes cheaper — and what becomes more valuable?
A firm that produces analysis faster may find its customers expect lower fees, commission less first-pass work, or do it themselves. Internal productivity rises while the commercial model comes under pressure — the question is what happens to the price of the output, the unit of sale and the sources of differentiation.
This layer asks where value will live — not merely where effort can be removed.
- →Which assets are becoming stranded, which can be converted, which compound as AI improves?
- →What is the unit of sale — and does the customer still want to buy that unit?
- →The repricing of a task is not the worth of a person: the economics and the workforce analysis stay connected.
What should be pursued?
A well-executed project can still be the wrong investment. Strategy asks what the company should defend, harvest, migrate, construct or stop — and favours the option that remains defensible under informed counterplay over the one with the most exciting unsupported upside.
This layer turns an interpretation of the future into choices that can be tested.
- →What happens when a competitor copies it, a customer negotiates differently, or delivery proves harder?
- →Which assumptions, evidence and rejected alternatives sit behind the recommendation?
- →What named evidence would reopen the decision? The Review is this layer, sold.
Under what conditions may the system act?
A system must be appropriately informed and appropriately constrained — and one does not supply the other. Better context does not grant permission; permission does not create understanding. The authority boundary lives in mechanisms outside the worker’s discretion — permissions, approved state transitions, records — not in the model agreeing to obey an instruction.
Not a final approval stage. It bears on every other stone — which is why it is the keystone.
- →What may the agent see, use and decide — and which actions require approval?
- →What must be recorded, and how is work stopped or recovered when conditions change?
- →Who is accountable — and has it been silently pushed onto a frontline approver?
What must the system understand about this organisation?
A general model does not know this company’s decisions, history, exceptions or standards of judgement. A transcript preserves a conversation; a maintained wiki explains which decision resulted, why, what it superseded and where the evidence sits — and survives a model change or a departure.
This layer makes generic intelligence specific, reusable and connected to the company’s own history.
- →Which decisions exist only in someone’s head?
- →Can an agent — or a new hire — trace a rule back to why it was made?
- →Does a generated answer get mistaken for evidence once it has been written down?
What continues after the conversation ends?
A request, an email, a generated answer and a completed model run are not a completed matter. The organising object is the persistent responsibility — its outcome, state, history, owner, authority and the conditions under which it counts as finished. The agent is the worker; it may stop, restart or be replaced. The obligation does not.
The standard is managerial: does the system absorb work, or manufacture work for the person supervising it?
- →What is open right now, who owns it, and what would close it?
- →Does the agent investigate, act within its authority, wait, follow up — or hand the problem straight back?
- →Where does continuity live when the model forgets?
Where should intelligence sit?
Not “how much of this can be made agentic?” but which parts benefit from judgement and which should not depend on it at all. Interpretation, research, planning and drafting are candidates for model-led work; exact calculation, identity, record state and consequential actions usually call for ordinary software.
This layer makes the ideas operational — and produces the evidence capable of overturning them.
- →What starts the work, what is it aiming at, what survives the run, and what can reject a result?
- →Can the model build tools to investigate without being able to rewrite what grants it authority?
- →Is the harness checking the work — or just looping around an API call?
Research environments, not a product portfolio
Roles, a company context, tools, document workspaces, review and persistent responsibilities — a live environment for testing the operating model.
Ingest, extract transferable patterns, build a library, combine, generate and review — discovery and selection feeding what might be constructed.
The conceptual work read beside project decisions and build history — so the relationship between an idea and its implementation is inspectable.
A marketing agent escalated a request because its tools could not report an image’s dimensions. The fix was not a stricter instruction to be independent — it was giving image generation and reading the measurement. The next build ran without the escalation. Implemented behaviour, not an audited reliability rate — but exactly the shape of the good-staff standard.
Not the best specialist in every discipline.
Less intent lost between them.
Your market, your numbers, your constraints — not a category average.
The economic question stays attached to the product decision and to what the system can actually do.
A bounded build that shows whether the idea survives contact with reality.
Doctrine, wiki and code that the next inquiry starts from — not a deck that expires.
Four claims, only four — from Sell the Compression, Not the Components.
Where the practice sits in the wider research
Context, not certification — none of these validates a particular doctrine or implementation.
AI in the tradition of tools that participate in human thought; the form of the relationship matters.
Greater confidence in AI associated with less reported critical thinking — self-report, so a design question, not a proof of decline.
Context-dependent productivity gains in customer support, larger for less-experienced workers — not a blanket effect.
Governance as cross-cutting and continuous, not a one-time step.
Workflows versus agents; add complexity only where it improves outcomes.
The aim is not simply more AI activity. It is stronger human capability, more viable businesses, and systems that reduce burden while leaving useful knowledge behind.