Source 1
LLM-embedded
What’s baked into the model’s weights from pre-training. Mostly L1, increasingly L2. Commoditising fast with every model release. Vendor-dependent. Gets better whether you do anything or not.
Activate Your Knowledge
Audit made the four knowledge layers of your people’s work visible. Activate shows where those layers actually sit when decisions are made. So you can deploy AI where outcome value is highest.
Activation is extending your architecture to capture the knowledge it doesn’t yet hold.
The Activate pillar in 2:46. Prefer to read? It’s all below. ↓
A note on scope. This corridor maps knowledge work — the judgement and interpretation your decisions depend on — not admin work. Every system below supports both, and every person does both. Our focus is where AI augments judgement; AI automating admin is a separate discipline. See Foundations for the academic lineage.
The Knowledge Corridor
The knowledge your architecture doesn’t yet capture
The Knowledge Corridor shows a typical pattern across enterprises. Where each knowledge layer usually sits in your stack. How much AI usually works with in each cell. Your organisation will vary in the specifics, but the shape tends to hold. Layer 1 (transferable) is reasonably well-covered everywhere, because the general craft is already in the AI instance. Layer 2 (domain) shows up partly across most systems. Layer 3a — your documented organisational knowledge — is the variance layer: heavy in some system-domains, thin in others, and fragmented across silos that don’t talk to each other. Layer 3b is sparse in almost every system except real-time communication, where the running team commentary carries traces of judgement no other system captures.
L3 · Organisational Knowledge
L1
Transferable
Any qualified professional
L2
Domain
Requires field expertise
L3a
Documented
Written down
L3b
Tacit
In people's heads
CRM & customer engagement
Front-office · Sales, service, account records
heavy
Sales craft — qualification, pipeline principles, negotiation patterns
moderate
Sector sales cycles (enterprise B2B vs consumer)
heavy
Account notes, opportunity history, discount matrix, pipeline stages
thin
VP's read on which customer expands, trust signals — mostly in heads
ERP & back-office
Back-office · Finance, supply chain, HR ops
heavy
Finance and ops principles — how AP, AR, and procurement typically work
moderate
Industry ERP flows (discrete vs process mfg; regulated finance)
heavy
Your chart of accounts, custom workflows, approval matrices
sparse
The CFO's read on why exceptions exist — at month-end, not in ERP
Document & knowledge repos
Knowledge capture · SharePoint, Confluence, Drive
thin
General reference material
moderate
Industry handbooks, external standards
heavy
Your policies, playbooks, SOPs, project archives — main L3a concentration
thin
Annotations, working-doc commentary — traces of reasoning
Collaboration suite
Daily work · Email, calendar, meetings
moderate
Writing craft, template patterns, meeting practices
thin
Industry templates (legal contracts, pharma SOPs)
moderate
Actual strategy decks, board papers, memos, meeting notes
thin
Back-channel DMs, unstated agreements — fragments of judgement
Real-time communication
Team coordination · Slack, Teams
sparse
—
sparse
—
moderate
Channels with documented decisions, announcements
heavy
DMs, threads, running commentary — highest L3b density in any system
Data & analytics platform
Decision support · Snowflake, Databricks, BigQuery
moderate
Analytics methods, statistical principles, dashboard patterns
moderate
Industry data models (insurance actuarial, retail forecasting)
heavy
Your dashboards, metric definitions, pipelines, data products
thin
Which analysts trust which data — tribal, unwritten
Product & delivery systems
Build & ship · GitHub, Jira, ServiceNow, MES
heavy
Coding patterns, project management principles, service-delivery craft
moderate
Industry-specific standards (regulated software, mfg, service delivery)
heavy
Your codebase/product specs, ADRs, ticket history, deployment logs
moderate
Why we built it this way — partly in PR/ticket comments, mostly in heads
Identity & access
Infrastructure · Entra, Okta, Google Identity
moderate
Standard auth flows, SSO patterns
moderate
Regulatory identity controls (SOX, HIPAA)
heavy
Your role structure, group memberships, access permissions
thin
Who should have access to what — implicit judgement
Compute & cloud foundation
Infrastructure · Where workloads run
moderate
Reference architectures, standard patterns — model knows these
thin
Industry compute patterns (e.g., regulated workloads)
moderate
Your naming standards, provisioning rules, reference stack
thin
The architect's judgement on what runs where
Cells are shaded by density. Heavier shade = more of that layer typically lives in that system-domain. Layer identity comes from the column; cell shade is pure intensity.
Heavy· most of this layer lives here
Moderate· some of this layer lives here
Thin· small traces of this layer
Sparse· little to none
Typically available to the deployment
Typically not available to the deployment
What’s available to your AI today. L1 (the transferable craft) is in the model. Ready for generic work. On regulated work it’s an integration question. L3a (your documented organisational knowledge) is only available where you’ve built the integration. L3b (the unwritten judgement) is almost nowhere.
But availability isn’t usage. Most of that capacity sits unused. The reasons stack into four rungs.
- 1
Access
There’s no AI tool yet that can do this work.
- 2
Embedding
A tool exists, but it’s only used when someone remembers.
- 3
Workflow
The AI isn’t part of how decisions actually get made.
- 4
Trust
The AI is there. It works. But people don’t yet act on what it says.
Activation is climbing this ladder. Two forces are helping: the models themselves keep improving, and a verification architecture is emerging. Deliberate human-in-the-loop checks, evaluation harnesses, traceable output. That’s what makes AI outputs defensible, and what closes rung four. Activation is the discipline of turning that unused capacity into active value, with the scaffolding that lets the organisation act on it.
Three knowledge sources every AI instance combines
At inference time, every AI deployment draws from some mix of three sources. The mix determines how much of your organisation’s actual knowledge is available to the deployment.
Source 2
Augmented / retrieved
What the deployment can draw on at inference time via RAG, context window, connectors, prompt. Mostly L3a — if the integration is built well. Most aren’t: ASIC found half of Australian financial firms with AI deployments had no fairness or bias policy. Enterprise-controlled, expensive to do well, and the layer where most deployments stall.
Source 3
Codified L3 (as tools / data)
Your own L3a turned into a structured, queryable asset — agent tools, APIs, knowledge graphs, sector role libraries. Highest effort. The layer AI can’t copy. This is where real return on AI is built.
How these two pieces fit together
The Knowledge Corridor above is the organisational diagnostic. It shows where your knowledge sits across your stack. The landscape the Activation Map will drill into.
The Activation Map below is the per-initiative decision. For each AI initiative, you drill down from a specific row of the Corridor (where this initiative’s knowledge actually sits) and plot it on two axes. How close it is to that knowledge today (X), and how much value it would produce (Y).
Most initiatives don’t start in the top-right. They start in Context Debt. The value is obvious, but the knowledge isn’t yet available. The activation work is about moving the initiative to the right by go-live: closing the proximity gap so the value can actually be captured. Like a risk matrix with a current position, a target position, and the work between them.
One Corridor per organisation. One Activation Map per initiative. — with an arrow showing where you need to move it.
The Activation Map
Proximity to knowledge × outcome value
Every AI deployment sits somewhere on two axes. Proximity to knowledge (X) is how close the AI is to where real organisational L3a/L3b actually sits. Outcome value (Y) is how much the activity produces. A low-stakes copilot. Or a strategic recommendation. The four quadrants behave very differently.
↑
Outcome value
High value · Far from knowledge
Context Debt
Ambition without grounding. AI makes high-stakes calls on general-purpose knowledge alone. Outputs can look credible and miss the specifics that drive the decision.
L1L2L3aL3b
High value · Close to knowledge
L3 Advantage
Codified knowledge plus tacit judgement, made available to a reasoning agent. AI synthesises and hands back a view a human can interrogate.
L1L2L3aL3b
Low value · Far from knowledge
L1 Commodity
Model-native productivity. Summarisation, drafting, generic templates. The AI draws only on its pre-training. Easy time savings. Not where differentiation compounds.
L1L2L3aL3b
Low value · Close to knowledge
L3a Foundation
Retrieval-grounded work on documented knowledge. RAG, connectors, knowledge graphs. Reliable foundation for the quadrant above.
L1L2L3aL3b
←Far
Proximity to knowledge
Close→ReachedPartialMissed
How to use the 2×2. Start with L3a Foundation (bottom-right): the entry point where AI adds reliable value because it has real context. Then move up-and-right to L3 Advantage — where durable organisational advantage compounds. The quadrant to keep on the radar is Context Debt (top-left): ambitious goals without the context to support them. Placing each initiative on this grid is how you balance the portfolio.
The Activation Map on one initiative
A contract-review copilot for a legal practice
The pitch. AI that reads contracts the way the senior partner does. Flagging unusual clauses, applying precedent, calling out risk. Big outcome value. Confidence is high.
What the Map shows. The partner’s reasoning isn’t in any system. The contract precedents aren’t structured. There’s no verification layer that would let a reviewer trust the AI’s citation. The initiative sits in Context Debt. High ambition, low proximity to the knowledge the work actually depends on.
The activation work. Extract the precedents into a structured library. Codify the partner’s reasoning patterns. Build the verification layer that lets a reviewer trust the AI’s citation. Each step closes the gap between what was pitched and what the AI can actually do. By go-live, the expected benefits can be realised.
The alternative. A similar investment in similar work delivers a generic copilot. Credible-looking output, missing the partner-level judgement that was expected. Failing silently as use fades over time.
Current position identified. Target state identified. Work between them specified. The Activation Map identifies the risks and activities required to achieve the expected business benefit. See the full project walkthrough →
Extract has three approaches
Building Source 3 — codified L3 — means moving L3b judgement out of specific people’s heads and into something the organisation can use. Three approaches, grounded in Harry Collins’ typology of tacit knowledge. Three distinct types (Relational, Somatic, Collective), each with a different reason for being tacit and a different path to surface it.
Ask, Observe, and Develop turn that typology into practical action verbs. Collins built on Nonaka & Takeuchi’s earlier SECI model; what we add is the operational layer on top. See Foundations for the full lineage.
Ask one diagnostic question of the person holding the knowledge. The answer tells you which approach fits.
Ask · Relational
“Can they explain it when asked the right question?”
The knowledge can be put into words. It’s tacit for social reasons — nobody asked, nobody thought it mattered, or the person holding it didn’t realise others needed it — not because the expert can’t find the words.
How:A structured conversation. Often in a single session.
Cost:Low. The bottleneck is that no one has asked the right way.
AI role:Once extracted, it becomes documented knowledge — L3a — and any AI system you have can use it.
Observe · Somatic
“Can they do it but not explain how?”
The knowledge has been compiled through years of practice below conscious access. The expert can do it but can’t describe it.
How:Watch them work. Codify iteratively. Test against real cases.
Cost:Medium to high. Sustained observation and multiple codification passes.
AI role:Partial. Once you’ve watched the expert, AI can help spot patterns across your notes. The watching itself stays human.
Develop · Collective
“Can others grow it through working alongside them?”
The knowledge belongs to the group, not the individual. Relationship trust, cultural intuition, political judgement.
How:Transfer through immersion — apprenticeship, mentoring, shared practice.
Cost:High and continuous. Can’t be compressed.
AI role:Structurally limited. This knowledge can’t go into a system. It can grow in other people.
Why this matters for your people
Extract isn’t only about making tacit knowledge useful to AI. It’s the start of a loop.
Captured knowledge becomes explicit. Explicit knowledge distributes widely. To people and AI both. Wider distribution sparks new thinking in the people who work with it. New thinking turns into new tacit intuition. And that becomes the next generation of Layer 3b you’ll capture.
That’s the human and organisational growth you can drive through a good Extract programme. It doesn’t just protect what you already know. It can accelerate the rate at which your organisation grows.
See the tools walked through
Two walkthroughs. One applies the Corridor to a decision. The other applies the Activation Map to an AI investment.
Decision walkthrough
A customer at risk. Save, match, or let go?
Classify the knowledge the decision draws on. Locate it across systems on the Corridor. Evaluate four AI approaches to supporting the call.
See the decision walkthrough →Project walkthrough
A contract-review copilot. Will it deliver?
Plot a specific AI investment on the Activation Map. Assess where it sits today, define the target, plan the move, and set the benchmark for go-live.
See the project walkthrough →See how vendors claim the layers
The corridor shows your landscape. The vendor landscape shows the market’s: which AI and enterprise vendors are claiming which Knowledge Strata layers. Useful when a CIO is evaluating overlapping pitches and trying to compose a coherent stack.
The corridor is your map. The 2×2 checks fit. Pillar 3 asks the next question. Once you know where your knowledge sits and which AI initiatives make sense, what do you do commercially with the defensible layer you’ve just identified?
Start a conversation
A question about the framework, or thinking about an engagement for your organisation? Send a note and you’ll hear back within a day or two.