Every AI task a business is handed starts the same argument about which tool to use, and skips two questions that decide far more. Who does the work, people or a machine? And who owns it once it is done, the business or a company it rents from? Put the answers on a grid and you get four quadrants. The business keeps one of them. The other three can change hands, and often should, without the business losing anything it knows.
People, owned: the seed
The seed holds what the business knows and believes: its positions, the facts it has checked, how it sounds, and what makes a customer choose it over the alternatives. Nobody rents it. It also holds the raw material underneath those, which is why a marketer who takes AI seriously starts asking the executive team for things that look nothing like marketing. The recordings of customer calls, the bill of materials, the return rate. Those three tell a machine more about a business than any brand guideline, and they are exactly the things a business should never leave a vendor to hold for it.
The System Seed is this quadrant in full, and who leads what is who in a business holds it.
AI, owned: the machinery on your core
Machines can work on things you own. A model drafting from your seed, a script that finds every page and proposal still carrying last year's price, an automation that writes a form response into your own records instead of a vendor's: the engine is rented, but the work lands in a place the business controls, under rules the business wrote. Own the core is where that place is, and why a workflow app that overwrites the past makes a poor one. Cascading updates are what the quadrant is for: change a fact once in the core and it changes everywhere the business says it.
It is also the quadrant where mistakes multiply. A machine working at speed on your own data carries one wrong number into every asset built from it, and the scaling of errors is that failure in full. The guard is Q3, a person checking the output against the seed before it ships.
People, rented: specialists
A business rents people as well as software. The agency running the Google Ads, the freelancer on Meta, the contractor building the email flows: each takes full responsibility for one platform or tactic, however yours is chopped up, and that is the reason to hire them. The check that matters comes before the contract, and it is whether what they learn flows back. If the specialist on Meta finds the headline that works, is there a channel for it to be stored in the seed and tried on the website and in the next email, or does it leave with the agency when the contract ends?
Money is the second check, whether the spend can carry the person, and the incentive is the third, whose interest the advice is in. Hiring specialists sets out the questions to ask before anyone signs.
AI, rented: the models and the tools
ChatGPT, Claude, Gemini, the CRM, the scheduler, whatever ships next quarter: this is the edge, and you rent it. I have watched a meeting argue one CRM against another when, at board level 3 years out, the choice barely registers, and a good team performs with almost any of them. Pick the one that fits, and stick with it. Because it is interchangeable, replacing it later costs the business nothing it knows.
The rule for this quadrant came from platforms: the cycle is going faster and faster, so play the platforms and don't live in them. Models are platforms now. The rented edge is how the window on any one of them closes, and the collapse of differentiation is what everyone renting the same model ends up sounding like.
Where the lines move
Three of the four quadrants are meant to change hands. A task moves from a specialist to a machine when the machine gets good enough, from one model to another when the price or the quality shifts, and from an agency back in-house when the spend grows. None of that should cost the business anything it knows, as long as the work in the rented column keeps writing back to the owned one. A vendor that keeps your history in its own tables, an agency that keeps the winning headline in its own deck, a model fed nothing but its own training data: each of those is a rented quadrant quietly taking over an owned one.
The top left never moves. You can pay someone to help write the seed down, and a machine can draft it, but the positions and the facts belong to the business, and so does the decision about what is true. When that quadrant is rented, the business still has a website, a CRM and an ad account, and no longer has anything to say that a competitor could not say too.
The loop inside the owned half
Work moves through the owned column as a loop of 4 steps, labelled Q1 to Q4 on this site. Q1 is strategy, what the business believes and why it matters, written into the seed. Q2 is execution, a machine turning one conviction into many expressions across channels and formats, fed from the seed and the core, never from nothing. Q3 is validation, a person checking the output against the seed before it ships. Q4 is deployment, the message out across the channels, with what it teaches fed back into Q1 and so into the seed.
Skip Q1 and the machine produces what every other machine produces. Leave out Q3 and one mistake turns up everywhere at once. At a finance advisory firm, a single ActiveCampaign automation sent 16,000 emails. Coca-Cola's AI-generated Christmas ads in 2024 had high production values and a brutal public reaction, because they read as a machine imitating warmth rather than a company talking to its customers, which is a failed Q3 in public.
What the quadrants don't claim
The grid does not claim people will never be replaced in marketing, or that a machine always needs a minder. It claims something narrower. Whatever a business rents, it should be able to stop renting without losing what it knows, and the only way to be sure of that is to keep the seed in its own hands. The four quadrants are a way of writing down, task by task, where that line sits, and the three clocks are the same split between rented and owned, applied to trust.
The diagnostic starts with the seed and then runs the loop's 4 steps against your own operation.