Grafton.
The Theory of the Intelligent Business
The intelligent business is its knowledge base.
Not the model — the model is a commodity: rented, identical under every rival, and improving for free on the labs' dime. The defensible firm is the knowledge base: owned, private, compounding, measurable.
The model=rented engine
·
The knowledge base = the firm
Grafton.
The shape of the firm
A firm is a transducer: four inputs → KB → four outputs
Knowledge work has only four input types — and, symmetrically, the same four output types. The KB is the state in the middle. Sharpen it, and all four output channels sharpen at once.
Four inputs
1Observations / actionsevents & things done
2Communicationsemail, messages
3Meetingscalls, conversations
4Documents / datafiles, records
ingest →
Knowledge Base
the state
owned · private · compounding · measurable
→ produce
Four outputs — the same four
1Actionsthings the firm does
2Communicationswhat it sends
3Meetingswhat it convenes
4Documents / datawhat it produces
The point
There is no fifth kind of knowledge work. And because one firm's output is another's input, these four types are the universal interface of knowledge work — every firm plugs into every other through the same four sockets. Improve the KB and every output improves for free.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
The closed loop
A coefficient governs the spiral
Outputs become future inputs — an action taken is an observation to re-ingest. So it's a loop, and every trip round it either improves the KB or degrades it. A knowledge coefficient — like a viral coefficient — decides which.
KB→
output→
world→
observation→
KB↻
> 1
Compounds — lift-off
Every interaction improves the base on average → the KB compounds exponentially. The firm gets smarter the more it operates.
A self-improving asset
< 1
Decays — a wasting asset
Interactions degrade it on average → noise accumulates, quality rots, the base decays the more it's used. Naive AI ingestion sits here (dump everything in a vector store).
Terminal
The point
Almost nobody clears 1 — which is exactly why it's defensible. The whole game is making every interaction, on average, improve the base rather than degrade it.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
What drives the coefficient · and how you measure it
Quality × throughput — and the KB is the balance sheet
Rate of improvement ≈ (knowledge gained per interaction) × (interactions cleared per unit time). Two coupled inputs feed it — and KB quality is measurable, which is what makes the coefficient credible.
×
The two coupled inputs
Process quality — how smart the queries / triage / context are, so each item is fast and correct to approve.
UX quality — how fast the human clears the queue. A growing backlog is the throughput axis failing — it drags the coefficient back toward 1 however good the processing.
Better UX → more knowledge in → smarter system → fewer, better items → faster still. Triage UX is core to the moat, not cosmetics.
$
KB quality is measurable — the firm's balance sheet
Structure — findable + homes for write-back
Accuracy — is it true?
No conflicts — coherence
Proposal quality — what you approve
Info-gap rate — unanswered asks for missing info
Traditional accounts measure the cash that fell out of past knowledge work. The KB measures the knowledge asset itself — marked to its current quality.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
Clearing the coefficient
A human guardian + answer write-back
You get above 1 with a human guarding ingestion, sharing what they know piece by piece — and the system banking every answer permanently. That second half is the load-bearing piece.
!
Answer write-back
Every answer becomes durable knowledge, so it's never asked again. Without it you re-answer the same things forever — effort spent, nothing banked (coefficient ≈ 1: the "who is speaker 2?" leak).
With it: every answer you ever give, you give once.
▽
Passive ingestion
Something arrives; the guardian gates it; the base improves. The default flow — content comes to you, you judge it in.
Capture → guardian-gated proposal → write-back.
△
Active ingestion
The base knows its own gaps and commissions the input: "discuss & decide these; I'll read the transcript and bank them." The smart pre-meeting agenda is the base's gap-list.
The highest form — the KB pulls knowledge in, aimed exactly where it's thin.
Third lever
Structure — findability for retrieval + homes for write-back. The wasting-asset failure is a coherence failure (contradictions pile up until trust collapses), held off by continuous fact-check, cross-reference, and dedup — the never-finished hygiene that keeps the base true as it grows.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
The workforce · two model tiers
Frontier as architect, free local models as the immune system
Today's frontier models (Opus 4.8 standard) are already good enough — the constraint is the harness, not capability. So own what compounds (the KB + harness); ride what commoditises (the model) for free.
FrontierArchitect + adjudicator
Design the system, make the hard judgements, drive improvement, adjudicate close calls. Scarce and expensive — used where stakes or difficulty are high.
designhard judgementimprovementadjudicate
Open-source · localThe KB's immune system
A limitless, free, private workforce for volume — classify, extract, fact-check, cross-reference, dedup. Free → unbinds throughput; local → keeps the KB private (which is what lets it be productised for regulated firms).
classifyextractfact-checkcross-referencededup
The rule
Both tiers flag for the guardian; they never silently mutate the base. The coefficient metrics are the routing governor — escalate local → frontier on low quality or high stakes. The human's job narrows to guard ingestion · judgement & taste · decisions — and headcount inverts: scale output without scaling people.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
How you build it · and why it changes valuation
Build it as modules — one input type at a time
Make ingestion first-class, input by input — each with the full guardian loop + write-back. Same shape every time, so each module is cheaper than the last and the system compounds as they land.
Module 1
Meetings
In build now
Module 2
Communications
Next
Module 3
Documents / data
Then
Module 4
Observations / actions
Then
Traditional firm
Value is tacit knowledge in heads — undocumented, untransferable. It walks out every evening, and permanently when a key person leaves. You can't mark it, sell it cleanly, or compound it.
Intelligent firm
Value is an explicit, owned, compounding, measurable KB. This deepens the Barbara roll-up — acquire firms to capitalise their tacit knowledge into one compounding base — and is the core of the productise-Labs-for-regulated pitch.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
From theory to numbers · the instrument panel
Measuring the intelligent business — one number per attribute
The theory becomes operational the moment its claims become numbers. Five measures, one per attribute of the firm — and they are a dependency chain, not a list: each is gated by the one before it, and the knowledge coefficient sits upstream of all of them.
A
Knowledge base
Size × Quality.
Coefficient = Δ(Size×Quality) per work item. >1 compounds.
B
Automation
Count of autonomous outcomes × their accuracy %.
C
Throughput
Cycle time through the stages + backlog at each stage.
D
Growth
New business % (logos or revenue) + CAC & LTV.
E
Economics
RPE + wallet ratio = (core + non-core) ÷ core.
A · KB coefficient→
B · accuracy→
B · automation→
E · RPE
C = clock speed · D = the same engine, demand-side
The chain
A compounding KB (A) is what makes (B)'s accuracy possible; proven accuracy is what earns automation; automation is what produces (E)'s RPE. The knowledge coefficient is upstream of them all — which is why it is the first measure, and the others its consequences.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
Measure A · the asset and its coefficient
The KB asset = Size × Quality — and a computable coefficient
Measure two things: the quality of ingestion (the process taking knowledge in — query volume + triage approval rate) and the quality of the base itself (the asset). A tiny perfect base is not the prize; how much true, coherent, findable knowledge you hold is.
Q
KB Quality — an intensive index (~0–1)
A consistent audit over the whole base — the wiki plus the records & files it points to — returns three counts, as densities (per unit of knowledge):
Conflicts — contradictions (coherence)
Accuracy issues — claims that are wrong (accuracy)
Info-gaps — missing pieces worth filling (completeness)
Structure — a cold-session retrieval probe: hit rate + # lookups to find sampled facts
Hold the audit protocol constant and these four are a faithful proxy for KB quality.
×
KB asset = Size × Quality
Size = distinct verified claims (pages as a rough proxy), never characters. Self-defending: padding the base with noise raises Size but lowers Quality — the two cancel, so you only score by adding knowledge that survives the audit.
Coefficient
Δ(Size × Quality) per work item. > 1 → compounds · ≤ 1 → a wasting asset. The viral-coefficient claim, turned into a reading.
The audit's gap-list is the commission list for the next round of active ingestion — the instrument that measures the base also tells it what to go and learn.
The Theory of the Intelligent Business · Grafton Labs · ideas wiki
Grafton.
Where this lives
Theory · OS · embodiment — across three workspaces
The single source of truth is here, in ideas. The others cite it, never duplicate it.
ideas canonical homeThe theory — the IP / the thinking
This theory (single source of truth) + the Verifiable Frontier — the child sub-theory of which outputs can run unattended, via τ (time-to-ground-truth) × stakes.
cockpit_os Labs — building the OS for the intelligent business + testing concepts
The Knowledge Coefficient (the mechanism) and the Labs Platform Overview (the team book). Labs is the OS for the intelligent business.
project_barbara The first embodiment
The AI-native lettings roll-up — the minimum-efficient-scale candidate for an intelligent business. The first real one, built on the engine.
In one line
ideas defines the intelligent business · Labs is the operating system that builds one · Barbara is the first real one. Theory → OS → embodiment.
The Theory of the Intelligent Business · Ed Barroll Brown · 17 June 2026 · ideas wiki