The case in full, rung by rung.
DataFab — the enterprise that understands itself
Every enterprise holds intelligence no competitor can buy — customers, counterparties, decisions, outcomes, and the reasons given at the time. It is scattered across systems that do not agree, and almost none of it can be put a question to. DataFab is the route from owning it to using it.
The asymmetry
General intelligence is converging.
Enormous capital, extraordinary people, real progress — and no argument from us. But general means available to everyone, on the same terms, in the same week. Whatever is built will reach your competitor when it reaches you. The other race is the one nobody in the room is running.
General intelligence
Learned from the world’s text. Identical for everyone who licenses it. Improves on a vendor’s schedule, not yours. A line of cost that buys parity, never advantage.
Artificial Proprietary Intelligence
It already exists, scattered across your systems. Not learned — discovered, resolved, governed and reasoned over in place. Customers, counterparties, decisions, outcomes, and the reasons given at the time.
The asymmetry
It cannot be bought at any price, only built. It compounds for a single organisation. And it is worth more the better the general models get.
General intelligence is a subscription. Artificial Proprietary Intelligence is a balance-sheet item.
The first is an operating cost every competitor also pays. The second compounds for exactly one organisation, and no amount of spending elsewhere reproduces it.The gap
Owning it and being able to use it are different things.
For a decade the answer was to move the data — into a warehouse, a lakehouse, a vendor platform. That worked for plenty of workloads. It breaks precisely where the data matters most: when it is sensitive, regulated, jurisdictional, operational, or too deeply embedded in core systems to be moved safely.
Eleven systems, no agreement
They do not agree on what a customer is. The same organisation appears four times under four spellings, with four identifiers, and no system is wrong.
No path across them
There is no route that carries a question over the estate and returns an answer with evidence you could defend to a regulator, a court or a client.
Governance on a different clock
Controls sit inside each application. An agent crossing five of them is governed by none, and the record of what it did is fragmented across all five.
A directorship sits in a filing. A shareholding sits in a register. A matter sits in the document system. Each fact is written down somewhere. What is written nowhere is the implication — that together they put you on both sides. No search over documents can return a conclusion that exists only between them.
A dashboard can tolerate a representation of the business. An agent cannot.
The bottleneck is no longer model capability, and no longer storage — it is governed enterprise context at the point of actionThe route
Seven rungs — each one the reason the next is possible.
You cannot reason over what you have not resolved. You cannot resolve what you have not discovered. And you cannot govern any of it once the data has moved.
Discover
Reach the systems where they are. Types, keys, distributions, anomalies and undocumented join paths, read rather than described. No extract, no copy, no migration.
Catalogue
The estate proposes its own model — entity types, relationship types, attributes — and a person approves, corrects and constrains it.
Resolve in place
Attribute matching, active learning and structural similarity decide that six records are one real entity. Thresholded, reversible, logged with its basis.
Persist as a graph
Entities and typed relationships with confidence, source and date on every edge — including the derived relationships that appear in no source record.
Govern the whole graph
Inherited access-control lists are the floor — stricter, never looser — enforced at the traversal layer, so reasoning runs over a view that already excludes what it may not see.
Watch, do not wait
A change in the world is matched against what the enterprise knows, and the work is raised. The system stops being a place you go and becomes something that comes to you.
Act under a gate
A model proposes; a deterministic gate disposes. Write-back through the same connectors as read. Versioned, replayable, reversible, autonomy set per task and per risk.
Every rung happens over live sources. Raw records stay behind the boundary; what persists is the resolved layer, with field-level lineage and a snapshot of each cited record so an answer survives the source moving. Nothing is lifted, nothing is copied, and no system is asked to change.
How it is built
Don’t rebuild the enterprise. Draw it.
The model is not authored up front by a specialist team. It emerges from the estate — and stays current, because the fabric maintains it. That is the whole difference between value in weeks and value at the end of a transformation programme.
The usual order
Define the model. Then wait.
- Define an ontology — authored up front by a specialist team
- Map every system into it, one project at a time
- Transform and move the data into a new store
- Populate, reconcile, validate
- Months before the first unit of value
- Every new source is a fresh modelling project
- Goes stale the moment a system changes
The fabric’s order
Connect. Then refine while already using it.
- Connect — point the fabric at a source you already run
- Discover — it catalogues the assets and relationships itself
- Resolve — records matched and merged into golden records
- Derive — a schema-bounded ontology emerges from the sources
- Refine — a person reviews and approves, while value is already flowing
- Maintain — change detection keeps the knowledge state current
- Weeks, not a transformation cycle
The machine that builds the asset
Three layers. One machine.
A foundation that draws the enterprise, a builder that turns it into governed execution, and utilities that ship ready to run. Each layer feeds the one below it as much as the one above — which is why the asset compounds rather than depreciates.
Governed utilities
Ready-made agentic applications — financial crime, due diligence, claim handling, client lifecycle management, remediation, government intelligence. Each grounded in the graph, governed by construction, and enriching the fabric as it runs.
↑ built in the Studio ↻ every execution enriches the Fabric
Layer 2 · the builderKnowledge & Agentic Studio
Where the enterprise composes its own governed agencies — from its own documents, in its own language, on its own schemas. The encoded workflow becomes an asset the firm owns rather than an engagement it renews.
↑ grounds ↓ enriches · governed
Layer 1 · the foundationKnowledge Fabric
Discovers and resolves the systems you already run into one governed, cumulative knowledge state — read in place, never copied. The resolved entity is the unit of the asset, and no competitor can obtain it.
Why it compounds
The Fabric grounds.
The Studio enriches.
Every governed schema, accepted relationship, published agency, workflow, human decision and execution outcome cascades back into the knowledge state — with provenance, version and authority attached. The enterprise does not merely automate work; its operational knowledge compounds as the work is performed. That is what makes it an asset rather than a tool.
This is self-enriching under governance — not self-learning without supervision. The Fabric is never polluted by raw model output, because nothing becomes enterprise truth until it has passed validation and approval.
Observed
An agency discovered something in the estate.
Proposed
The system inferred a relationship or a rule from what it observed.
Validated
Human or deterministic evidence confirmed it.
Approved
Authorised enterprise knowledge — and only now does it count as truth.
Superseded
Later evidence changed it. The record of why is kept.
The difference
Everyone will sell you an agent.
Almost no one will hand you the asset beneath it.
The market has settled into two shapes. One asks you to move into its world — its ontology, its engineers, its deployment measured in quarters and its dependency measured in years, with the intelligence accruing to them. The other hands you a framework and leaves the governed data foundation, the hard half, for you to build. DataFab is neither.
The ontology is yours
On one model the intelligence becomes the vendor’s asset, built on your data, and every asset you create deepens their moat. Here every schema, agency and resolved entity is your intellectual property.
Your people, not a standing bench
One model sends a dozen engineers into your building for a year. The other sends nothing at all — your own people compose an agency in the Studio in an afternoon, from documents they already have.
The working, not just the answer
A black box asks for your trust and offers a headline accuracy number on the box. Every DataFab output carries its evidence — source per statement, confidence per step, and a decision you can replay.
Nothing crosses the boundary
Copied into someone else’s environment, or resolved in place with nothing leaving yours. For regulated, sovereign and classified estates that difference is not a preference — it is the whole conversation.
Weeks, not quarters
No discovery phase before the discovery phase. No ontology programme. Connect a source, and build the first governed agency the same week.
Everyone can retrieve. Almost no one can act.
Retrieval is the easy part and it is commoditising fast. Everything after it — policy, gates, write-back, replay, defensibility — is where the work actually lives.
Where it is proven
Chosen where being wrong is not an option.
The beachhead is deliberately the hardest ground: estates where the data cannot move, the work is regulated and repetitive, and every decision must survive an inspector, a court or an oversight body.
Trust
Governed and certified by construction.
An asset you cannot defend is not an asset. Governance here is not a policy document laid over the top — it is the architecture, enforced at the graph, on every read and every action, for every person and every agency.
Resolve, never copy
Records stay in their system of record in every edition. Only mappings, derived insight and resolved state live in the fabric, and the control plane never holds raw data.
Judgement stays with people
Consequential actions wait on an accountable person, and the gate itself is logged with the basis they had.
Schema-bounded by design
Extraction is constrained by seed schemas. The system cannot invent entities outside them — hallucination prevention by construction.
Tamper-evident audit
Hash-chained and encrypted, auditor role only, with no delete capability. The record is produced as the work happens.
Next step
Owning it is not the same as being able to use it.
The destination is not in dispute. The route is the whole question — and it runs through discovery, resolution, a governed graph and a gate, over systems that never move.