The platform
What I've built, how it works, and why it's the engine for every Grafton venture — Project Barbara first.
Grafton Labs · the OS for the intelligent business · 2026
Embodies the theory: ideas · intelligent-business-theory
The big idea
A CRM, a notes app, a shared drive — they're filing cabinets. They wait for a person to put everything in. Labs is an engine: it generates the work, routes it, and brings you in to judge — not to type. The point of the engine is a business that is its knowledge base.
What Labs is for
The defensible firm isn't the one with the best model — the model is rented, identical for every rival, and improving for free. It's the one with the best knowledge base: owned, private, compounding, measurable. Labs is the operating system for building one.
A firm is a transducer — four inputs (observations/actions · communications · meetings · documents/data) → the KB → the same four outputs. Labs is that transducer made real. Keep the knowledge coefficient above 1 — every interaction improving the base on average — and it compounds.
The theory: ideas · intelligent-business-theory · the mechanism: knowledge-coefficient
Why this is different
Software is a record of the past. Nothing happens until a human notices, decides, and types it in. The tool is only ever as good as the discipline around it.
An engine for the future. It emits its own triggers, routes each onto a process, drafts the response, and asks you to approve. The work comes to you, already done — you judge it.
The first principle
One system holds the state, the knowledge and the work for every venture, project and personal area — not scattered across inboxes, drives and heads, but in one place: queryable, shared, and scoped to who you are.
The architecture
The database — wiki, registries, tasks, the trust machinery. The substrate is the product.
Standing AI agents + working sessions. They act on the substrate — and obey one rule: propose, never silently change.
The app, the tasks PWA, Slack. They render the substrate; they hold no state of their own.
Read the substrate · act as the runtime · report to the surfaces.
The shape
Grafton Capital, Project Barbara, LegalHelp, WealthVector — and personal areas too — each is a workspace: its own knowledge, work and apps. One window onto all of them: the Labs app. What you can see is scoped to you.
A living wiki per workspace — the durable memory.
Tasks and processes that humans and agents move.
The products and tools each venture runs.
The substrate, up close
Why you can trust it
The hands
Standing AI agents run on a workstation around the clock. Daisy keeps every wiki coherent; briefing agents prepare your day; classifiers sort inbound email, meetings and files. Working sessions drop into any workspace and build. They propose — you approve. The system gets more capable without getting less trustworthy.
The heart of it
AI proposes → you judge it with the context in front of you → approve, edit, or guard.
Approvals, reviews, triage — they're all the same move. The future of knowledge work isn't doing the work; it's approving the work and guarding the context it touches.
What starts the work
An email, a meeting note, a message, a manual update, a file, a system event. The world acting on you — which a filing cabinet can also receive.
A date arriving, a goal coming due, a threshold crossed. The system generating its own work — which a filing cabinet never can. This is the engine.
The model
External or internal pulse arrives.
Routed to the process it belongs to.
Drafts the steps; you judge at the gate.
Approved / edited / declined — the ledger.
The verdict ledger rolls up per process — which is how a process earns its way to autonomy.
The unit of the system
How it starts — scheduled, an event, a response to another process, or by hand.
The typed things it needs — a supplier, a property, an amount, a document.
The ordered actions it performs — each with its own pathway and approval gate.
Name the outcome, give it a recipe, and the system can run it, track it, and improve it.
How far AI can go · the child sub-theory
How far a process can run on its own = how quickly we can know it was right (time to ground truth, τ) × the stakes (cost × how hard it is to undo). The parent theory says a human guards ingestion and judgement; this says which outputs can earn their way out from under that human. That sets a ceiling — a pathway:
AI can complete it. Earns its way to unattended.
AI is the copilot; the human owns the outcome.
A person does it; AI coordinates and tracks.
The safety model
The real product
Author processes top-down where you know them — and discover them bottom-up: the system watches the work, and when the same thing recurs in volume, it proposes turning that pattern into a tracked process. Capture → spot the pattern → define → instrument → earn autonomy.
Why it's the platform for work
Strip any knowledge job to its bones and it's the same shape: a trigger arrives, a process responds, a human judges the outcome. Compliance, lettings, deal-making, your diary — all of it runs on the one spine. Build the spine once; every domain plugs in.
That's why this isn't a tool for one job. It's the platform underneath all of them.
The first embodiment
Barbara — the AI-native lettings roll-up — is the first embodiment of the intelligent business: the minimum-efficient-scale candidate, built on the engine, not beside it. The strategy writes itself from the model:
Own the workflow, the data and the judgement; rent the commodity plumbing (CRM, client-money) until we outgrow it.
Start beside today's tools, mirror them into Labs, augment with AI, then own the process outright.
First lettings processes already modelled: maintenance intake, rent collection, client-money reconciliation — each on its own pathway, earning autonomy as the ledger fills.
And beyond
A platform that holds a regulated firm's knowledge as a governed spine — and runs its processes with earned, evidenced autonomy — is something those firms can't build for themselves. Barbara proves the engine on our own business; the same engine becomes a product others run on.
Prove it on us first. Then sell the factory, not the furniture.
Your part in it
Welcome to Labs — the platform for knowledge work.
First build-along: the accounts / invoice-matching process. See the How to build a process guide in the wiki.
Grafton.